{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.12.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceType":"competition","sourceId":11848,"databundleVersionId":862157,"mountSlug":"competitions/histopathologic-cancer-detection"},{"sourceType":"kernelVersion","sourceId":235679288,"mountSlug":"notebooks/sepehrkh/use-gpu-t4x2-two-gpus-on-histopathologic-images"}],"dockerImageVersionId":31328,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"\n#

🩺 AI FastAPI Deployment β€”
🧬Histopathologic Cancer Classification

","metadata":{}},{"cell_type":"markdown","source":"
\n \n# 🩺 Histopathologic Cancer Classification API","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Project Overview\nThis project implements a **Medical AI Classification API** for binary cancer detection from histopathology images. \n### It demonstrates an **end-to-end pipeline** from research notebooks to production-ready API, including:\n\n- CPU and Multi-GPU Training & Inference (T4x2 on Kaggle)\n- TensorFlow / VGGNet.keras model\n- Production-ready FastAPI application\n- Image Upload & Preprocessing\n- `/predict` endpoint with JSON response\n- Dockerized CPU & GPU deployment\n- Swagger documentation\n- Medical AI best practices\n\n#### **🏹 Goal:** Transform a Kaggle research notebook into a **portfolio-ready and monetizable AI service**.\n\n---\n\n## πŸ“ Roadmap & Phases\n\n| Phase | Description | Key Output |\n|---|---|---|\n| Phase 0 | Project Definition & Scope | Mental model, Input/Output spec |\n| Phase 1 | Research vs Production Separation | Inference-only preprocessing & pipeline |\n| Phase 2 | API Contract (Design First) | `/predict` specification, error handling |\n| Phase 3 | Professional FastAPI Project Structure | main.py, health check |\n| Phase 4 | Model Loading & GPU Handling | Load model, warm-up, stable inference |\n| Phase 5 | Image Upload & Preprocessing | UploadFile handling, validation, preprocessing |\n| Phase 6 | `/predict` Endpoint | Endpoint implementation, threshold, JSON |\n| Phase 7 | Production Hardening | Logging, error handling, request timing, readiness |\n| Phase 8 | API Testing & Usage | Swagger, curl, Python client example |\n| Phase 9 | Medical AI Best Practices | Confidence interpretation, safe defaults, explainability |\n| Phase 10 | Dockerization & Deployment | Dockerfile CPU/GPU, Render/HF deployment |\n\n---\n\n### πŸ“ Phase Highlights\n\n#### Phase 0 β€” Project Definition\n- Defined project goal: binary cancer detection\n- Input: histopathology image\n- Output: JSON (prediction + confidence)\n- Portfolio & monetization considerations\n\n---\n\n#### Phase 1 β€” Research vs Production Separation\n- Extracted **preprocessing** from Kaggle training notebook\n- Established inference-only pipeline\n- Ensured reproducibility and GPU compatibility\n\n---\n\n#### Phase 2 β€” API Contract\n- Designed `/predict` endpoint\n- Defined **input/output** schema\n- Error handling specification\n\n---\n\n#### Phase 3 β€” Professional FastAPI Project Structure\n- Created `main.py` for FastAPI\n- Added `/health` endpoint\n- Verified API runs in Kaggle environment\n\n---\n\n#### Phase 4 β€” Model Loading & GPU Handling\n- Loaded TensorFlow / `VGGNet.keras` model\n- Configured **multi-GPU** inference\n- Performed model **warm-up** for stable response\n\n---\n\n#### Phase 5 β€” Image Upload & Preprocessing\n- Implemented **UploadFile** endpoint\n- **Validated** file type and size\n- **Preprocessed images** to match training pipeline\n- Added batch dimension and normalization\n\n---\n\n#### Phase 6 β€” `/predict` Endpoint\n- Implemented **POST** endpoint\n- Integrated preprocessing + model inference\n- Applied **thresholding**\n- Returned structured `JSON` response\n\n---\n\n#### Phase 7 β€” Production Hardening\n- Added **logging** for requests and predictions\n- Structured Pydantic response models\n- Measured request time\n- Error handling\n- Health/readiness endpoints implemented\n\n---\n\n#### Phase 8 β€” API Testing & Usage\n- Swagger documentation generated automatically\n- Tested API with curl\n- Python client example included\n- Batch inference demonstrated\n\n---\n\n#### Phase 9 β€” Medical AI Best Practices\n- Interpreted confidence levels (`low`, `medium`, `high`)\n- Set safe defaults for prediction\n- Explained preprocessing & thresholding rationale\n\n---\n\n#### Phase 10 β€” Dockerization & Deployment\n- Created Dockerfiles for CPU & GPU\n- Verified images on Kaggle and local machine\n- Ready for Render / HuggingFace / server deployment\n\n---\n\n## πŸ“ Key Notes & Considerations\n- Kaggle environment used for initial API testing; external access is limited \n- Preprocessing strictly matches the training pipeline (size, normalization, dtype) \n- Confidence is not a clinical diagnosis β€” predictions are research/support only \n- Multi-GPU inference is tested and stable \n\n---\n\n## πŸ“ Usage Example\n\n**Start API locally:**\n```bash\nuvicorn app.main:app --host 0.0.0.0 --port 8000\n````\n\n**Upload image via curl:**\n\n```bash\ncurl -X POST \"http://localhost:8000/predict\" -F \"file=@path/to/image.png\"\n```\n\n**Sample JSON Response:**\n\n```json\n{\n \"filename\": \"Histology-Cancer.png\",\n \"probability\": 0.004296,\n \"prediction\": \"Normal\",\n \"confidence\": 0.995704,\n \"confidence_level\": \"high_confidence\",\n \"inference_time_ms\": 324.96\n}\n```\n\n---\n\n## 🧱 File Structure (Final & Production-ready)\n\n```text\nhistopathology-api/\nβ”‚\nβ”œβ”€β”€ app/\nβ”‚ β”œβ”€β”€ main.py\nβ”‚ β”œβ”€β”€ config.py\nβ”‚\nβ”œβ”€β”€ models/\nβ”‚ └── VGGNet.keras\nβ”‚\nβ”œβ”€β”€ client/\nβ”‚ └── client.py\nβ”‚\nβ”œβ”€β”€ Dockerfile.cpu\nβ”œβ”€β”€ Dockerfile.gpu\nβ”œβ”€β”€ requirements.txt\nβ”œβ”€β”€ .dockerignore\nβ”œβ”€β”€ README.md\n\n```","metadata":{}},{"cell_type":"markdown","source":"# 🧬 Phase 0 β€” Project Definition & Scope","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Project Definition & Scope\n\n### Concept\nThe goal of this project is to transform a **medical image classification research notebook**\ninto a **production-ready AI service**.\n\nThe task is **binary cancer detection** from histopathology images.\n\n### Key Decisions\n- Input: Histopathology image\n- Output: JSON (prediction + confidence)\n- Model: TensorFlow VGGNet.keras\n- Focus: Engineering-quality inference & deployment\n\n### Why This Phase Matters\nA clear problem definition ensures:\n- Correct API design\n- Reproducible inference\n- Resume and portfolio-ready output\n","metadata":{}},{"cell_type":"markdown","source":"## 🧠 Key Decisions Lock-In (Pre–Phase 0)\n\n### πŸ”Ή 1) Final Classification Model (Locked πŸ”’)\n\n- **Framework:** TensorFlow \n- **Model Format:** `VGGNet.keras` \n- **Model Location:** Kaggle project artifact \n- **Status:** \n - βœ”οΈ Fully trained \n - βœ”οΈ Multi-GPU (T4 Γ— 2) \n - βœ”οΈ Inference-ready \n\nπŸ“Œ This choice is **technically sound and professionally justified** because:\n\n- VGG architectures remain **scientifically defensible** in medical imaging\n- Inference behavior is **stable and predictable**\n- Simple, transparent, and ideal for **FastAPI-based industrial demos**\n\nπŸ‘‰ From this point forward:\n\n> **`VGGNet.keras` is the Single Source of Truth for inference**\n\n\n### πŸ”Ή 2) Real Deployment Strategy (Render / HuggingFace / Docker)\n\nβœ”οΈ Fully considered \nβœ”οΈ Included in the global roadmap \nβœ”οΈ **Intentionally excluded from Kaggle execution**\n\nπŸ“Œ Correct engineering strategy:\n\n- This project = **Engineering Proof + API Design**\n- Public deployment = **Independent follow-up phase**\n (reusing this exact codebase)\n\n---\n\n## πŸš€ Problem Definition\n\n### Task\n\n**Histopathologic Cancer Detection (Binary Classification)**\n\n### Input\n\n- Histopathology image \n- RGB format \n- Size: `96 Γ— 96 Γ— 3`\n\n### Output\n\n- Label:\n - `0 β†’ Non-Cancer`\n - `1 β†’ Cancer`\n- Associated probability (confidence score)\n\n---\n\n## πŸ“¦ What We Are Deploying (Precisely)\n\nWe are **NOT deploying**:\n\n- Training pipeline \n- Data augmentation \n- Multi-GPU logic \n- Callbacks or schedulers \n\nWe **ARE deploying**:\n\n- βœ”οΈ Inference-only model \n- βœ”οΈ Exact preprocessing used during training \n- βœ”οΈ Stable and deterministic prediction logic \n- βœ”οΈ FastAPI endpoint \n\n---\n\n## πŸ₯ Medical AI Constraints (Critical)\n\nThroughout the entire project, we strictly enforce:\n\n* ❗ This model is **NOT a definitive medical diagnostic tool**\n* βœ… Prediction confidence is explicitly returned\n* βœ… Decision thresholds are explainable\n* βœ… Outputs are stable and interpretable\n\n---\n\n## 🧱 Environment & Constraints\n\n| Component | Status |\n| ------------- | ----------------------- |\n| Environment | Kaggle Notebook |\n| GPU | T4 Γ— 2 |\n| FastAPI | Local (Notebook-only) |\n| Public Access | ❌ (Kaggle limitation) |\n| Testing | Internal Swagger/curl |\n\n---\n\n","metadata":{}},{"cell_type":"markdown","source":"## 🩺 Inference Pipeline β€” Conceptual View\n\n```\nImage (Upload)\n\n ↓\n\nValidation (format, RGB)\n\n ↓\n\nResize (96Γ—96)\n\n ↓\n\nNormalize (same as training)\n\n ↓\n\nExpand dims (batch=1)\n\n ↓\n\nModel.predict()\n\n ↓\n\nProbability\n\n ↓\n\nThreshold Decision\n\n ↓\n\nJSON Response\n```","metadata":{}},{"cell_type":"markdown","source":"# πŸ§ͺ Phase 1 β€” Research vs Production Separation","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Phase 1 β€” Research vs Production Separation\n\n### Concept\n\nTraining pipelines and inference pipelines must be **strictly separated**.\nOnly deterministic, minimal logic should be used in production.\n\n### Overview\n\nThe primary objective of Phase 1 is to **strictly separate research-time logic from production inference logic**.\n\nThis phase transforms a training-focused Kaggle notebook into a **clean, deterministic, inference-only pipeline** that can later be embedded into FastAPI, Docker, or any deployment environment.\n\n> No training code. \n> No augmentation. \n> No callbacks. \n> No multi-GPU logic. \n> **Inference only.**\n\n---\n\n### 🎯 Phase Goals\n\n- Isolate **Single Source of Truth** for inference configuration\n \n- Reproduce **exact preprocessing used during training**\n \n- Build a **safe, minimal, and deterministic inference pipeline**\n \n- Validate correctness with a real image (sanity check)\n\n### What was done\n\n- Extracted exact preprocessing logic from training\n\n- Built inference-only pipeline\n\n- Removed all training dependencies (Removed augmentation and training-only steps)\n\n- Loaded the final `VGGNet.keras` model safely in Kaggle\n\n- Implemented deterministic prediction logic\n\n### Outcome\n\n- Stable, reproducible inference pipeline\n\n- Ready for **FastAPI** integration\n\n- Production-oriented Medical AI design\n\n### Professional Notes\n\n- Ensures reproducibility\n\n- Prevents data leakage\n\n- Aligns exactly with training preprocessing\n\n---\n\n## 🧱 Pipeline Architecture\n\n```\nImage Path\n ↓\nImage Validation & Loading\n ↓\nInference Preprocessing\n ↓\nFrozen Model (VGGNet.keras)\n ↓\nProbability\n ↓\nThreshold Decision\n ↓\nJSON Output\n````","metadata":{}},{"cell_type":"markdown","source":"### πŸ—ƒ Installing and importing requirements","metadata":{}},{"cell_type":"code","source":"pip install fastapi uvicorn","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:12:16.727890Z","iopub.execute_input":"2026-07-02T07:12:16.728634Z","iopub.status.idle":"2026-07-02T07:12:21.839772Z","shell.execute_reply.started":"2026-07-02T07:12:16.728595Z","shell.execute_reply":"2026-07-02T07:12:21.838634Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ===============================\n# Phase 1 - Inference Preprocessing\n# ===============================\n\nimport tensorflow as tf\nimport numpy as np\nfrom PIL import Image\nimport os\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:12:21.841652Z","iopub.execute_input":"2026-07-02T07:12:21.842608Z","iopub.status.idle":"2026-07-02T07:12:52.482287Z","shell.execute_reply.started":"2026-07-02T07:12:21.842564Z","shell.execute_reply":"2026-07-02T07:12:52.481401Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"\n---\n\n### Cell 1 β€” Project Configuration (Single Source of Truth)\n\nCentralize all inference-related constants to avoid silent mismatches between training and production.\n","metadata":{}},{"cell_type":"code","source":"# =====================================\n# Project Configuration and Constants\n# =====================================\n\nIMG_SIZE = (96, 96)\nNUM_CHANNELS = 3\nDTYPE = \"float32\"\n\nTHRESHOLD = 0.5\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:12:52.483610Z","iopub.execute_input":"2026-07-02T07:12:52.484244Z","iopub.status.idle":"2026-07-02T07:12:52.489587Z","shell.execute_reply.started":"2026-07-02T07:12:52.484212Z","shell.execute_reply":"2026-07-02T07:12:52.488640Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#### πŸ“Œ Notes\n\n- IMG_SIZE exactly matches training resolution\n\n- DTYPE is explicitly defined for numerical stability\n\n- THRESHOLD is configurable and medically interpretable","metadata":{}},{"cell_type":"markdown","source":"### Cell 2 β€” Image Loading & Validation (Production-Safe)\n\nEnsure robust and defensive image loading suitable for production APIs.\n","metadata":{}},{"cell_type":"code","source":"# =====================================\n# Image Loading & Validation\n# =====================================\n\nfrom PIL import Image\nimport os\n\ndef load_image_from_path(image_path: str) -> Image.Image:\n \"\"\"\n Load image from disk and ensure RGB format.\n \"\"\"\n if not os.path.exists(image_path):\n raise FileNotFoundError(f\"Image not found: {image_path}\")\n\n image = Image.open(image_path)\n\n if image.mode != \"RGB\":\n image = image.convert(\"RGB\")\n\n return image\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:12:52.491829Z","iopub.execute_input":"2026-07-02T07:12:52.492750Z","iopub.status.idle":"2026-07-02T07:12:52.511454Z","shell.execute_reply.started":"2026-07-02T07:12:52.492702Z","shell.execute_reply":"2026-07-02T07:12:52.510423Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#### πŸ“Œ Notes\n\n- Explicit file existence check\n\n- Forced RGB conversion prevents inference crashes\n\n- Raises clear errors instead of failing silently","metadata":{}},{"cell_type":"markdown","source":"### Cell 3 β€” Preprocessing Function (Inference-Only, Final)\n\nReproduce exact training-time preprocessing without any augmentation or randomness.","metadata":{}},{"cell_type":"code","source":"# ===============================\n# Inference Preprocessing\n# ===============================\n\ndef preprocess_image(image: Image.Image) -> np.ndarray:\n \"\"\"\n Preprocess image for VGGNet inference.\n \n Steps:\n - Resize\n - Convert to NumPy\n - Normalize\n - Add batch dimension\n \"\"\"\n\n # Resize\n image = image.resize(IMG_SIZE) # IMG_SIZE = (96, 96)\n\n # Convert to NumPy array\n image_array = np.array(image, dtype=DTYPE) # DTYPE = np.float32\n\n # Normalize (same as training: [0, 1])\n image_array = image_array / 255.0\n\n # Add batch dimension: (1, H, W, C)\n image_array = np.expand_dims(image_array, axis=0)\n\n return image_array\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:12:52.512646Z","iopub.execute_input":"2026-07-02T07:12:52.512994Z","iopub.status.idle":"2026-07-02T07:12:52.529391Z","shell.execute_reply.started":"2026-07-02T07:12:52.512965Z","shell.execute_reply":"2026-07-02T07:12:52.528241Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#### πŸ“Œ Notes\n\n- No augmentation (strict inference behavior)\n\n- Normalization [0, 1] matches training pipeline\n\n- Explicit batch dimension (1, H, W, C)\n\n- **Stateless**; **Deterministic**; and **Repeatable**","metadata":{}},{"cell_type":"markdown","source":"### Cell 4 β€” Load Model (Inference Mode)\n\nLoad the trained model in pure inference mode.","metadata":{}},{"cell_type":"code","source":"# =====================================\n# Load Trained Model (Inference Only)\n# =====================================\n\nimport tensorflow as tf\n\nMODEL_PATH = (\n \"/kaggle/input/notebooks/sepehrkh/use-gpu-t4x2-two-gpus-on-histopathologic-images/VGGNet.keras\"\n)\n\nmodel = tf.keras.models.load_model(MODEL_PATH)\n\nmodel.trainable = False\n\nprint(\"βœ… Model loaded successfully\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:12:52.530676Z","iopub.execute_input":"2026-07-02T07:12:52.531065Z","iopub.status.idle":"2026-07-02T07:12:57.064969Z","shell.execute_reply.started":"2026-07-02T07:12:52.531033Z","shell.execute_reply":"2026-07-02T07:12:57.063830Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"model","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:12:57.066286Z","iopub.execute_input":"2026-07-02T07:12:57.066650Z","iopub.status.idle":"2026-07-02T07:12:57.073849Z","shell.execute_reply.started":"2026-07-02T07:12:57.066608Z","shell.execute_reply":"2026-07-02T07:12:57.072868Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#### πŸ“Œ Notes\n\n- Model is frozen explicitly\n\n- Training graph is irrelevant at inference time\n\n- VGGNet.keras remains the Single Source of Truth","metadata":{}},{"cell_type":"markdown","source":"### Cell 5 β€” Prediction Logic (Production-Ready)\n\nEncapsulate inference into a clean, reusable function.","metadata":{}},{"cell_type":"code","source":"# =====================================\n# Prediction Logic\n# =====================================\n\ndef predict_from_path(image_path: str):\n \"\"\"\n Run inference on a single image path.\n \"\"\"\n image = load_image_from_path(image_path)\n input_tensor = preprocess_image(image)\n \n # Model prediction\n prob = model.predict(input_tensor, verbose=0)[0][0]\n\n label = \"Cancer\" if prob >= THRESHOLD else \"Non-Cancer\"\n\n return {\n \"label\": label,\n \"probability\": float(prob)\n }\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:12:57.075214Z","iopub.execute_input":"2026-07-02T07:12:57.075712Z","iopub.status.idle":"2026-07-02T07:12:57.093207Z","shell.execute_reply.started":"2026-07-02T07:12:57.075670Z","shell.execute_reply":"2026-07-02T07:12:57.091856Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#### πŸ“Œ Notes\n\n- Deterministic inference\n\n- Clear threshold logic\n\n- JSON-friendly output format\n\n- Ready for FastAPI integration","metadata":{}},{"cell_type":"markdown","source":"### Sanity Check β€” End-to-End Validation\n\nVerify that the entire pipeline works correctly with a real dataset image.","metadata":{}},{"cell_type":"code","source":"# ===============================\n# Test Inference\n# ===============================\n\n# - /kaggle/input/histopathologic-cancer-detection/train/00001b2b5609af42ab0ab276dd4cd41c3e7745b5.tif\nTEST_IMAGE_PATH = (\n \"/kaggle/input/competitions/histopathologic-cancer-detection/\"\n \"train/00001b2b5609af42ab0ab276dd4cd41c3e7745b5.tif\"\n)\n\nresult = predict_from_path(TEST_IMAGE_PATH)\nprint(result)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:12:57.094840Z","iopub.execute_input":"2026-07-02T07:12:57.095245Z","iopub.status.idle":"2026-07-02T07:12:57.900931Z","shell.execute_reply.started":"2026-07-02T07:12:57.095204Z","shell.execute_reply":"2026-07-02T07:12:57.899882Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# πŸ₯ Phase 2 β€” API Contract (Design First)","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Phase 2 β€” API Contract (Design First)\n\n### Concept\nBefore coding, the API contract must be clearly defined.\n\n### What was designed?\n\n- Defined `/health` and `/predict` endpoints\n\n- Specified `request`/`response` schemas\n\n- Designed medical-safe output format\n\n- Defined error handling strategy\n\n### Why it matters:\n\n- Enables clean **FastAPI** implementation\n\n- Ensures stable client integration\n\n- Makes the service production-ready\n\n- Production-grade API behavior\n\n","metadata":{}},{"cell_type":"markdown","source":"## πŸ“¦ Phase Scope (What We Define)\n\nWe **only define the API contract** β€” no implementation.\n\nThis phase focuses on **design clarity**, not execution.\n\n### What is included:\n- Endpoints\n- HTTP methods\n- Input schema\n- Output schema\n- Error responses\n- Medical AI constraints\n\n### What is explicitly excluded:\n❌ No FastAPI code \n❌ No imports \n❌ No server execution \n\n---\n\n## πŸ”Œ API Overview (High-Level)\n\n### Defined Endpoints\n\n| Endpoint | Method | Purpose |\n|------------|--------|-----------------------|\n| `/health` | GET | Health check |\n| `/predict` | POST | Cancer classification |\n\n---\n\n## πŸ”Ή Endpoint 1 β€” `/health`\n\n### Purpose\n\nVerify that:\n- The API is running\n- The model is loaded\n- The inference pipeline is ready\n\n### Request\n\n```http\nGET /health\n````\n\n### Successful Response (200)\n\n```json\n{\n \"status\": \"ok\",\n}\n```\n\n---\n\n## πŸ”Ή Endpoint 2 β€” `/predict` (Core Endpoint)\n\n### Purpose\n\nReceive a histopathology image and return a cancer classification result.\n\n---\n\n### πŸ“₯ Request Specification\n\n* **Method**\n\n```http\nPOST /predict\n```\n\n* **Content-Type**\n\n```http\nmultipart/form-data\n```\n\n#### Request Fields\n\n| Field | Type | Required | Description |\n| ----- | --------------- | -------- | -------------------- |\n| file | Image (PNG/JPG) | βœ… | Histopathology image |\n\n---\n\n#### Response Fields\n\n| Field | Type | Description |\n| ----------- | ------ | ---------------------------- |\n| label | string | Final predicted class |\n| probability | float | Model confidence (range 0–1) |\n\n---\n\n## πŸ₯ Medical AI Constraints (Contract-Level)\n\nThese constraints are **part of the API contract**, not implementation details:\n\n* ❗ Output is **not a definitive medical diagnosis**\n* βœ… Probability is explicitly returned and interpretable\n* βœ… Classification threshold is configurable (server-side)\n* βœ… API enforces safe default behavior\n\nπŸ“Œ These constraints will later appear in:\n\n* Swagger documentation\n* README\n* Public-facing deployment descriptions\n\n---\n\n## ❌ Error Responses (Design Specification)\n\n### HTTP 400 β€” Invalid File\n\n```json\n{\n \"error\": \"Invalid image format\"\n}\n```\n\n### HTTP 422 β€” Missing File\n\n```json\n{\n \"error\": \"No image file provided\"\n}\n```\n\n### HTTP 500 β€” Internal Server Error\n\n```json\n{\n \"error\": \"Inference failed\"\n}\n```\n\n---\n\n## 🧱 Determinism & Stability Guarantees\n\nThe API guarantees:\n\n* Stateless behavior\n* Same input β†’ same output\n* No randomness\n* No training-time logic\n\n---\n","metadata":{}},{"cell_type":"markdown","source":"# 🏹 Phase 3 β€” Professional FastAPI Project Structure","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Phase 3 β€” Professional FastAPI Project Structure\n\n### Concept\n\nBuild a minimal but real FastAPI application inside Kaggle.\n\n### Overview\n\nPhase 3 marks the transition from **design and inference logic** to a **real FastAPI application**.\n\nThe goal of this phase is **not public deployment**, but to:\n- Initialize a FastAPI app\n- Implement the `/health` endpoint defined in Phase 2\n- Verify that the API lifecycle works inside a Kaggle Notebook environment\n\n### 🎯 Phase Goals\n\n- Create a FastAPI application entry point\n- Implement a real `/health` endpoint\n- Run FastAPI inside Kaggle (not publicly exposed)\n- Validate API behavior at the Python level\n\n### What will be done\n\n- Initialize **FastAPI** application\n\n- Create `main.py`\n\n- Configure API metadata (title, description, version)\n\n- Implement `/health` endpoint\n\n- Successfully ran FastAPI inside the Kaggle environment\n \n### πŸ“Œ Notes\n\n- Kaggle does not expose public ports\n\n- Swagger UI is sufficient for testing\n\n- This phase validates API lifecycle\n","metadata":{}},{"cell_type":"markdown","source":"### πŸ₯ Architectural Clarification β€” Notebook vs Application Code\n\n### Overview\n\nDuring Phase 3, FastAPI endpoints were initially defined inside the Kaggle notebook for rapid validation and experimentation.\n\nHowever, for production-grade architecture, a strict separation must be maintained between:\n\n- Development environment (Notebook)\n- Application source code (`app/` directory)\n\nThis section clarifies that separation.\n\n---\n\n### Notebook Responsibilities (Development Environment)\n\nThe Kaggle notebook is used for:\n\n- Installing dependencies\n- Testing inference logic\n- Validating endpoint behavior\n- Creating project files\n- Running temporary local servers\n- Debugging and experimentation\n\nThe notebook is **not** the final production entry point.\n\n---\n\n### `app/` Directory Responsibilities (Production Code)\n\nThe `app/main.py` directory contains:\n\n- FastAPI application instance\n- All route definitions\n- Model loading logic\n- Prediction endpoint\n- Health endpoint\n- Middleware and error handling\n\nThis directory represents the **true deployable application**.\n","metadata":{}},{"cell_type":"markdown","source":"## ⚠ Why `%%writefile` Is Not Reliable\n\nInitially, we attempted to create `main.py` using:\n\n```python\n%%writefile main.py\n````\n\nHowever, this resulted in:\n\n```\nUsageError: Line magic function `%%writefile` not found.\n```\n\n#### Root Cause\n\n`%%writefile` is an **IPython magic command**, not standard Python.\n\nMagic commands:\n\n* Depend on the Jupyter/IPython runtime\n* Are not guaranteed to be available in all environments\n* Are not suitable for production-grade projects\n\nFor a serious FastAPI project intended for:\n\n* GitHub\n* Docker deployment\n* Client demonstration\n* Gumroad packaging\n\nWe must use **standard Python file operations**.\n\n---\n\n### The Professional Approach (Production-Grade Method)\n\nInstead of relying on notebook-specific magic commands, we use:\n\n* `os.makedirs()` to create directories\n* Standard `open()` file writing\n* A clean project structure\n\nThis mirrors real-world FastAPI repositories.\n\n---\n\n### Target Project Structure\n\n```\nproject/\n app/\n main.py\n```\n\nThis structure is:\n\n* Docker-ready\n* Cloud-deployment-ready\n* GitHub-clean\n* Recruiter-friendly\n\n---\n\n### Final Implementation (Kaggle-Compatible & Production-Safe)\n\n```python\n# =====================================\n# Phase 3 β€” Create Production Structure\n# =====================================\n\nimport os\n\n# Create app directory\nos.makedirs(\"app\", exist_ok=True)\n\n# Define main.py content\nmain_code = \"\"\"\n # full code here\n\"\"\"\n\n# Write file\nwith open(\"app/main.py\", \"w\") as f:\n f.write(main_code)\n\nprint(\"βœ… app/main.py created successfully\")\n```\n\n---\n\n### Why This Is the Best Practice\n\nβœ” Pure Python (no notebook dependency)\n\nβœ” Reproducible in any environment\n\nβœ” Matches top FastAPI GitHub repositories\n\nβœ” Ready for Docker and cloud deployment\n\nβœ” Clean separation between Notebook and Application Code\n\n","metadata":{}},{"cell_type":"markdown","source":"### - Install Dependencies\n\nFastAPI and required runtime dependencies are installed explicitly.\n","metadata":{}},{"cell_type":"code","source":"# =====================================\n# Phase 3 - Install Dependencies\n# =====================================\n\n!pip install -q fastapi uvicorn python-multipart\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:12:57.904124Z","iopub.execute_input":"2026-07-02T07:12:57.904679Z","iopub.status.idle":"2026-07-02T07:13:01.968661Z","shell.execute_reply.started":"2026-07-02T07:12:57.904648Z","shell.execute_reply":"2026-07-02T07:13:01.967179Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#### Notes\n\n- `python-multipart` is required for future file upload endpoints\n\n- Versions are kept lightweight for Kaggle compatibility","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Creating `main.py` the Production-Ready Way","metadata":{}},{"cell_type":"code","source":"# =====================================\n# Phase 3 β€” Create Production Structure\n# =====================================\n\nimport os\n\n# Create app directory\nos.makedirs(\"app\", exist_ok=True)\n\n# Write main.py inside app/\nmain_code = \"\"\"\nfrom fastapi import FastAPI\n\napp = FastAPI(\n title=\"Histopathologic Cancer Classification API\",\n description=\"Inference-only Medical AI API using FastAPI\",\n version=\"1.0.0\"\n)\n\n@app.get(\"/\")\ndef root():\n return {\"message\": \"FastAPI is running\"}\n\n@app.get(\"/health\")\ndef health_check():\n return {\n \"status\": \"ok\",\n \"model\": \"VGGNet\",\n \"framework\": \"TensorFlow\"\n }\n\"\"\"\n\nwith open(\"app/main.py\", \"w\") as f:\n f.write(main_code)\n\nprint(\"βœ… app/main.py created successfully\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:01.970466Z","iopub.execute_input":"2026-07-02T07:13:01.970838Z","iopub.status.idle":"2026-07-02T07:13:01.979288Z","shell.execute_reply.started":"2026-07-02T07:13:01.970800Z","shell.execute_reply":"2026-07-02T07:13:01.978321Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"---\n\n## ⚠️ Kaggle Networking Limitation (Important)\n\nKaggle Notebooks **do not support public port exposure**.\n\nAs a result:\n\n* The FastAPI server runs correctly\n* The browser **cannot access** `http://127.0.0.1:8000/docs`\n* Swagger UI is unavailable in Kaggle\n\nThis is a **platform limitation**, not an application error.\n\n### Correct Engineering Interpretation\n\n* API logic is validated\n* HTTP layer testing is deferred to:\n\n * Local machine\n * Docker containers\n * Cloud platforms (Render / HuggingFace)\n\n---\n\n### πŸ“Œ Note\n> *Due to Kaggle's networking limitations, FastAPI endpoints cannot be accessed via a browser. The API is validated at the Python level and fully tested in containerized/local environments.*\n\n","metadata":{}},{"cell_type":"markdown","source":"## πŸš€ Phase 3 β€” Professional FastAPI Project Structure (GitHub-Ready)\n\n---\n\n### 🎯 Objective\n\nTransform the FastAPI prototype into a **modular, production-style project structure**.\n\n---\n\n### 🧱 Final Project Structure (From This Phase Onward)\n\nInside Kaggle, we create this structure:\n\n```\nhistopathology-api/\nβ”‚\nβ”œβ”€β”€ app/\nβ”‚ β”œβ”€β”€ main.py\nβ”‚ β”œβ”€β”€ config.py\nβ”‚\nβ”œβ”€β”€ models/\nβ”‚ └── VGGNet.keras\nβ”‚\nβ”œβ”€β”€ client/\nβ”‚ └── client.py\nβ”‚\nβ”œβ”€β”€ Dockerfile.cpu\nβ”œβ”€β”€ Dockerfile.gpu\nβ”œβ”€β”€ requirements.txt\nβ”œβ”€β”€ .dockerignore\nβ”œβ”€β”€ README.md\n```\n---\n\n### 🧠 Why This Structure Is Professional\n\nThis mirrors how serious ML APIs are built:\n\n* App logic separated from notebooks\n* Entry point defined in `main.py`\n* Dependencies explicitly declared\n* Ready for:\n\n```bash\nuvicorn app.main:app --reload\n```\n\ninside project root.\n\n---\n\n### πŸ“¦ How Employer Can Run This (Later)\n\nAfter cloning repository:\n\n```bash\ncd histopathology-api\npip install -r requirements.txt\nuvicorn app.main:app --reload\n```\n\nThen open:\n\n```\nhttp://localhost:8000/docs\n```\n\nSwagger will work perfectly outside Kaggle.\n\n---\n","metadata":{}},{"cell_type":"markdown","source":"# ⏳ Phase 4 β€” Model Loading & GPU Handling","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Phase 4 β€” Model Loading & GPU Handling\n\n### Concept\n\nIn this phase, we **optimized the FastAPI** inference pipeline for real-world deployment by `enabling GPU memory safety`, `model warm-up`, `fast inference execution`, `secure file handling`, and `basic logging`.\nThe API is now production-ready and prepared for **Docker-based** deployment on cloud platforms such as Render or HuggingFace Spaces.\n\nModel loading must be:\n- Stable\n- Deterministic\n- GPU-aware\n\n### Implementation\n- Load TensorFlow `.keras` model\n- Enable GPU inference\n- Perform model warm-up\n\n### Professional Notes\n- Warm-up avoids first-request latency\n- GPU memory is allocated early\n- Improves production stability\n","metadata":{}},{"cell_type":"markdown","source":"## πŸ“Œ Why **\"Model Warm-Up\"** Matters in Production AI\n\n\nWhen deploying a deep learning model as a real-time API, loading the model into memory is only part of the startup process.\n\n\nEven after the model is loaded, the first inference request often performs several internal initialization tasks before producing a prediction. This phenomenon is commonly known as **cold start latency**.\n\n\nTo eliminate this delay for end users, a **model warm-up** step is executed immediately after the model is loaded.\n\n\n\n---\n\n\n\n### What is Model Warm-Up?\n\n\n\nModel warm-up performs one or more dummy inference passes before the API begins serving real requests.\n\n\n\nThe purpose is **not** to improve prediction accuracy.\n\n\n\nInstead, it prepares the inference engine so that the first user request experiences the same low latency as subsequent requests.\n\n\n\n---\n\n\n\n### Why Does the First Request Take Longer?\n\n\n\nDuring the first inference, TensorFlow may perform several initialization tasks, including:\n\n\n\n* Building and tracing the computation graph (`@tf.function`)\n\n* Initializing CUDA kernels (GPU deployment)\n\n* Loading cuDNN and cuBLAS libraries\n\n* Allocating GPU memory\n\n* Creating optimized internal buffers\n\n* Validating tensor shapes and data types\n\n\n\nWithout warm-up, these operations are executed during the first user request.\n\n\n\n---\n\n\n\n### Latency Comparison\n\n\n\n| Request | Without Warm-Up | With Warm-Up |\n\n| ------------------- | --------------: | -----------: |\n\n| First request | High latency | Low latency |\n\n| Subsequent requests | Low latency | Low latency |\n\n\n\nThe first user should never pay the initialization cost.\n\n\n\n---\n\n\n\n### Implementation\n\n\n\nOur API performs model warm-up immediately after loading the model.\n\n\n\n```python\n\ndef warmup_model(model):\n\n dummy = np.zeros((1, 96, 96, 3), dtype=np.float32)\n\n model.predict(dummy, verbose=0)\n\n```\n\n\n\nThe dummy tensor matches the model's expected input shape while containing no real medical information.\n\n\n\nIts only purpose is to activate the inference backend.\n\n\n\n---\n\n\n\n### Where Should Warm-Up Be Performed?\n\n\n\nThe recommended location is during application startup, immediately after the model is loaded and before the API accepts incoming requests.\n\n\n\nThis guarantees that users never experience cold-start latency.\n\n\n\n---\n\n\n\n### What Warm-Up Is Not\n\n\n\nModel warm-up is **not**:\n\n\n\n* Model training\n\n* Fine-tuning\n\n* Weight optimization\n\n* Accuracy improvement\n\n\n\nIt is purely an inference performance optimization technique.\n\n\n\n---\n\n\n\n### Why It Matters in This Project\n\n\n\nThis project deploys a TensorFlow CNN model through FastAPI for real-time histopathologic image classification.\n\n\n\nBecause the service is intended for interactive inference, minimizing first-request latency is an important aspect of production readiness.\n\n\n\nAlthough warm-up has no effect on predictive performance, it significantly improves user experience by reducing startup delays.\n\n\n\n---\n\n\n### πŸ“Œ Takeaway\n\n\n\n> **Model warm-up is a production optimization that ensures the first user receives the same fast inference experience as every subsequent user.**","metadata":{}},{"cell_type":"markdown","source":"## ⏳ Model Loading & GPU Handling (Production Ready)","metadata":{}},{"cell_type":"code","source":"'''\n# ============================================================\n# main.py\n# Phase 4 β€” Model Loading & GPU Handling (Production Ready)\n# ============================================================\n\n# =============================\n# Standard Library Imports\n# =============================\nimport os\nimport logging\nimport tempfile\n\n# =============================\n# Third-Party Imports\n# =============================\nimport numpy as np\nimport tensorflow as tf\nfrom fastapi import FastAPI, UploadFile, File, Request\nfrom contextlib import asynccontextmanager\n\n# ============================================================\n# Logging Configuration\n# ============================================================\n\nlogging.basicConfig(\n level=logging.INFO,\n format=\"%(asctime)s - %(levelname)s - %(message)s\"\n)\n\nlogger = logging.getLogger(\"HistopathologyAPI\")\n\n# ============================================================\n# GPU Configuration\n# ============================================================\n\ndef configure_gpu():\n \\\"\\\"\\\"\n Configure GPU memory growth for stable inference.\n Prevents TensorFlow from allocating all GPU memory at once.\n \\\"\\\"\\\"\n logger.info(\"Checking GPU availability...\")\n gpus = tf.config.list_physical_devices('GPU')\n\n if gpus:\n try:\n for gpu in gpus:\n tf.config.experimental.set_memory_growth(gpu, True)\n logger.info(f\"βœ… {len(gpus)} GPU(s) configured with memory growth\")\n except RuntimeError as e:\n logger.error(f\"GPU configuration error: {e}\")\n else:\n logger.warning(\"⚠ No GPU detected, running on CPU\")\n\n# ============================================================\n# Model Warm-up\n# ============================================================\n\ndef warmup_model(model, input_shape=(1, 96, 96, 3)):\n \\\"\\\"\\\"\n Perform dummy forward pass to stabilize first inference latency.\n \\\"\\\"\\\"\n logger.info(\"Starting model warm-up...\")\n dummy_input = np.zeros(input_shape, dtype=np.float32)\n _ = model.predict(dummy_input, verbose=0)\n logger.info(\"πŸ”₯ Model warm-up completed\")\n\n# ============================================================\n# Optimized Inference Function\n# ============================================================\n\n@tf.function\ndef model_inference(model, batch):\n \\\"\\\"\\\"\n Optimized TensorFlow graph inference.\n \\\"\\\"\\\"\n return model(batch, training=False)\n\n# ============================================================\n# Prediction Wrapper\n# ============================================================\n\ndef predict_image_fast(model, image_tensor):\n \\\"\\\"\\\"\n Perform fast prediction with confidence extraction.\n \\\"\\\"\\\"\n image_tensor = tf.expand_dims(image_tensor, axis=0)\n preds = model_inference(model, image_tensor)\n\n prob = float(preds[0][0])\n\n label = \"Cancer\" if prob >= 0.5 else \"No Cancer\"\n confidence = prob if prob >= 0.5 else 1 - prob\n\n return label, confidence\n\n# ============================================================\n# Temporary File Handler\n# ============================================================\n\ndef save_temp_image(upload_file: UploadFile):\n \\\"\\\"\\\"\n Safely store uploaded image temporarily.\n \\\"\\\"\\\"\n suffix = os.path.splitext(upload_file.filename)[-1]\n with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:\n tmp.write(upload_file.file.read())\n return tmp.name\n\n# ============================================================\n# FastAPI Lifespan (Professional Lifecycle Injection)\n# ============================================================\n\n@asynccontextmanager\nasync def lifespan(app: FastAPI):\n\n # Startup\n\n logger.info(\"πŸš€ Starting application...\")\n\n # 1️⃣ GPU Setup\n configure_gpu()\n\n # 2️⃣ Load Model\n MODEL_PATH = os.getenv(\n \"MODEL_PATH\",\n \"models/VGGNet.keras\"\n )\n \n logger.info(f\"Loading model from: {MODEL_PATH}\")\n model = tf.keras.models.load_model(MODEL_PATH)\n model.trainable = False\n logger.info(\"βœ… Model loaded successfully\")\n\n # 3️⃣ Warm-up\n warmup_model(model)\n\n # 4️⃣ Store model in app state (Singleton)\n app.state.model = model\n\n yield\n\n # Shutdown\n logger.info(\"πŸ›‘ Application shutdown\")\n\n# ============================================================\n# FastAPI App Instance\n# ============================================================\n\napp = FastAPI(lifespan=lifespan)\n\n# ============================================================\n# Prediction Endpoint Example\n# ============================================================\n\n@app.post(\"/predict\")\nasync def predict(request: Request, file: UploadFile = File(...)):\n\n logger.info(\"Prediction request received\")\n\n model = request.app.state.model\n\n # For demo: dummy tensor (replace with real preprocessing)\n image_tensor = tf.zeros((96, 96, 3), dtype=tf.float32)\n\n label, confidence = predict_image_fast(model, image_tensor)\n\n logger.info(f\"Prediction result: {label} | Confidence: {confidence:.4f}\")\n\n return {\n \"label\": label,\n \"confidence\": round(confidence, 4)\n }\n\n'''","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:01.980896Z","iopub.execute_input":"2026-07-02T07:13:01.981409Z","iopub.status.idle":"2026-07-02T07:13:02.101663Z","shell.execute_reply.started":"2026-07-02T07:13:01.981329Z","shell.execute_reply":"2026-07-02T07:13:02.100771Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸ“ Build \"models/VGGNet.keras\"","metadata":{}},{"cell_type":"code","source":"import shutil\nimport os\n\nos.makedirs(\"models\", exist_ok=True)\n\nshutil.copy(\n \"/kaggle/input/notebooks/sepehrkh/use-gpu-t4x2-two-gpus-on-histopathologic-images/VGGNet.keras\",\n \"models/VGGNet.keras\"\n)\n\nprint(\"βœ… Model copied\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.102927Z","iopub.execute_input":"2026-07-02T07:13:02.103314Z","iopub.status.idle":"2026-07-02T07:13:02.272551Z","shell.execute_reply.started":"2026-07-02T07:13:02.103284Z","shell.execute_reply":"2026-07-02T07:13:02.271449Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸ“ Build/Update \"app/main.py\" File","metadata":{}},{"cell_type":"code","source":"# ============================================\n# Build / Update app/main.py\n# ============================================\n\nmain_code = \"\"\"\n# ============================================================\n# main.py\n# Phase 4 β€” Model Loading & GPU Handling (Production Ready)\n# ============================================================\n\n# =============================\n# Standard Library Imports\n# =============================\nimport os\nimport logging\nimport tempfile\n\n# =============================\n# Third-Party Imports\n# =============================\nimport numpy as np\nimport tensorflow as tf\nfrom fastapi import FastAPI, UploadFile, File, Request\nfrom contextlib import asynccontextmanager\n\n# ============================================================\n# Logging Configuration\n# ============================================================\n\nlogging.basicConfig(\n level=logging.INFO,\n format=\"%(asctime)s - %(levelname)s - %(message)s\"\n)\n\nlogger = logging.getLogger(\"HistopathologyAPI\")\n\n# ============================================================\n# GPU Configuration\n# ============================================================\n\ndef configure_gpu():\n \\\"\\\"\\\"\n Configure GPU memory growth for stable inference.\n Prevents TensorFlow from allocating all GPU memory at once.\n \\\"\\\"\\\"\n logger.info(\"Checking GPU availability...\")\n gpus = tf.config.list_physical_devices('GPU')\n\n if gpus:\n try:\n for gpu in gpus:\n tf.config.experimental.set_memory_growth(gpu, True)\n logger.info(f\"βœ… {len(gpus)} GPU(s) configured with memory growth\")\n except RuntimeError as e:\n logger.error(f\"GPU configuration error: {e}\")\n else:\n logger.warning(\"⚠ No GPU detected, running on CPU\")\n\n# ============================================================\n# Model Warm-up\n# ============================================================\n\ndef warmup_model(model, input_shape=(1, 96, 96, 3)):\n \\\"\\\"\\\"\n Perform dummy forward pass to stabilize first inference latency.\n \\\"\\\"\\\"\n logger.info(\"Starting model warm-up...\")\n dummy_input = np.zeros(input_shape, dtype=np.float32)\n _ = model.predict(dummy_input, verbose=0)\n logger.info(\"πŸ”₯ Model warm-up completed\")\n\n# ============================================================\n# Optimized Inference Function\n# ============================================================\n\n@tf.function\ndef model_inference(model, batch):\n \\\"\\\"\\\"\n Optimized TensorFlow graph inference.\n \\\"\\\"\\\"\n return model(batch, training=False)\n\n# ============================================================\n# Prediction Wrapper\n# ============================================================\n\ndef predict_image_fast(model, image_tensor):\n \\\"\\\"\\\"\n Perform fast prediction with confidence extraction.\n \\\"\\\"\\\"\n image_tensor = tf.expand_dims(image_tensor, axis=0)\n preds = model_inference(model, image_tensor)\n\n prob = float(preds[0][0])\n\n label = \"Cancer\" if prob >= 0.5 else \"No Cancer\"\n confidence = prob if prob >= 0.5 else 1 - prob\n\n return label, confidence\n\n# ============================================================\n# Temporary File Handler\n# ============================================================\n\ndef save_temp_image(upload_file: UploadFile):\n \\\"\\\"\\\"\n Safely store uploaded image temporarily.\n \\\"\\\"\\\"\n suffix = os.path.splitext(upload_file.filename)[-1]\n with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:\n tmp.write(upload_file.file.read())\n return tmp.name\n\n# ============================================================\n# FastAPI Lifespan (Professional Lifecycle Injection)\n# ============================================================\n\n@asynccontextmanager\nasync def lifespan(app: FastAPI):\n\n # Startup\n\n logger.info(\"πŸš€ Starting application...\")\n\n # 1️⃣ GPU Setup\n configure_gpu()\n\n # 2️⃣ Load Model\n MODEL_PATH = os.getenv(\n \"MODEL_PATH\",\n \"models/VGGNet.keras\"\n )\n \n logger.info(f\"Loading model from: {MODEL_PATH}\")\n model = tf.keras.models.load_model(MODEL_PATH)\n model.trainable = False\n logger.info(\"βœ… Model loaded successfully\")\n\n # 3️⃣ Warm-up\n warmup_model(model)\n\n # 4️⃣ Store model in app state (Singleton)\n app.state.model = model\n\n yield\n\n # Shutdown\n logger.info(\"πŸ›‘ Application shutdown\")\n\n# ============================================================\n# FastAPI App Instance\n# ============================================================\n\napp = FastAPI(lifespan=lifespan)\n\n# ============================================================\n# Prediction Endpoint Example\n# ============================================================\n\n@app.post(\"/predict\")\nasync def predict(request: Request, file: UploadFile = File(...)):\n\n logger.info(\"Prediction request received\")\n\n model = request.app.state.model\n\n # For demo: dummy tensor (replace with real preprocessing)\n image_tensor = tf.zeros((96, 96, 3), dtype=tf.float32)\n\n label, confidence = predict_image_fast(model, image_tensor)\n\n logger.info(f\"Prediction result: {label} | Confidence: {confidence:.4f}\")\n\n return {\n \"label\": label,\n \"confidence\": round(confidence, 4)\n }\n\n\"\"\"\n\nwith open(\"app/main.py\", \"w\") as f:\n f.write(main_code)\n\nprint(\"βœ… main.py updated successfully\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.273850Z","iopub.execute_input":"2026-07-02T07:13:02.274309Z","iopub.status.idle":"2026-07-02T07:13:02.285473Z","shell.execute_reply.started":"2026-07-02T07:13:02.274263Z","shell.execute_reply":"2026-07-02T07:13:02.284546Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"with open(\"app/main.py\", \"r\") as f:\n print(f.read())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.286798Z","iopub.execute_input":"2026-07-02T07:13:02.287397Z","iopub.status.idle":"2026-07-02T07:13:02.312493Z","shell.execute_reply.started":"2026-07-02T07:13:02.287355Z","shell.execute_reply":"2026-07-02T07:13:02.311509Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸš€ Phase 4 β€” Model Loading & GPU Handling\n\n### 🎯 Objective\nIntegrate the trained `VGGNet.keras` model into the FastAPI application with production-grade lifecycle management and optimized GPU handling.\n\n---\n\n### ⏳ Implemented Components\n\n#### 1. GPU Configuration\n- Enabled TensorFlow memory growth\n- Prevented full GPU memory pre-allocation\n- Supported multi-GPU environments\n- Fallback to CPU when GPU unavailable\n\n---\n\n#### 2. Model Loading\n- Model loaded during FastAPI startup\n- Implemented inside `lifespan()` context manager\n- Stored as singleton inside `app.state`\n\n---\n\n#### 3. Model Warm-up\n- Dummy forward pass performed\n- Eliminated first-request latency spike\n- Ensured stable inference response time\n\n---\n\n#### 4. Optimized Inference\n- Wrapped inference inside `@tf.function`\n- Disabled training mode\n- Optimized for single-GPU production inference\n\n---\n\n#### 5. Prediction Wrapper\n- Extracted probability\n- Applied classification threshold\n- Returned label + confidence score\n\n---\n\n#### 6. Logging\n- Centralized logging system\n- Request & prediction tracking\n- Production-ready logging format\n\n---\n\n### Architecture Decision\n\nTraining β†’ Multi-GPU \nInference β†’ Single-GPU optimized \n\nThis approach ensures:\n- Stable deployment\n- Predictable memory usage\n- Low latency inference\n- Production scalability\n\n---\n\n### Result\n\nPhase 4 is now:\n- Production-ready\n- Lifecycle-safe\n- GPU-aware\n- Scalable\n- Clean-architecture compliant\n","metadata":{}},{"cell_type":"markdown","source":"# πŸ–Ό Phase 5 β€” Image Upload & Preprocessing","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Phase 5 β€” Image Upload & Preprocessing\n\n### Concept\nConvert HTTP input into valid model input.\n\n### Implementation\n- Accept UploadFile\n- Validate file type\n- Convert to PIL\n- Apply preprocessing\n\n### Key Considerations\n- Reject invalid formats\n- Prevent corrupted inputs\n- Ensure shape & dtype consistency\n","metadata":{}},{"cell_type":"markdown","source":"## πŸ–Ό Image Upload & Preprocessing","metadata":{}},{"cell_type":"code","source":"'''\n\n# ============================================================\n# Phase 5 β€” Image Upload & Preprocessing (Production Ready)\n# ============================================================\n\nfrom fastapi import HTTPException\nfrom PIL import Image\nimport io\n\n# ============================================================\n# Image Configuration\n# ============================================================\n\nALLOWED_EXTENSIONS = {\".jpg\", \".jpeg\", \".png\"}\nMAX_FILE_SIZE_MB = 5\nIMG_SIZE = (96, 96)\nDTYPE = np.float32\n\n# ============================================================\n# File Validation\n# ============================================================\n\ndef validate_file_extension(filename: str):\n \\\"\\\"\\\"\n Validate uploaded file extension.\n \\\"\\\"\\\"\n ext = os.path.splitext(filename)[1].lower()\n\n if ext not in ALLOWED_EXTENSIONS:\n raise HTTPException(\n status_code=400,\n detail=\"Invalid file type. Only JPG and PNG images are allowed.\"\n )\n\n\ndef validate_file_size(file: UploadFile):\n \\\"\\\"\\\"\n Validate uploaded file size.\n \\\"\\\"\\\"\n file.file.seek(0, os.SEEK_END)\n size = file.file.tell()\n file.file.seek(0)\n\n size_mb = size / (1024 * 1024)\n\n if size_mb > MAX_FILE_SIZE_MB:\n raise HTTPException(\n status_code=400,\n detail=f\"File too large. Max allowed size is {MAX_FILE_SIZE_MB}MB.\"\n )\n\n# ============================================================\n# Image Preprocessing (Aligned with Training Pipeline)\n# ============================================================\n\ndef preprocess_image(file: UploadFile) -> np.ndarray:\n \\\"\\\"\\\"\n Preprocess image for VGGNet inference.\n Steps:\n 1. Read image from UploadFile\n 2. Convert to RGB\n 3. Resize to training size\n 4. Convert to NumPy\n 5. Normalize (0-1 scaling)\n 6. Add batch dimension\n \\\"\\\"\\\"\n\n contents = file.file.read()\n image = Image.open(io.BytesIO(contents)).convert(\"RGB\")\n\n # Resize\n image = image.resize(IMG_SIZE)\n\n # Convert to NumPy\n image_array = np.array(image, dtype=DTYPE)\n\n # Normalize\n image_array = image_array / 255.0\n\n # Add batch dimension\n image_array = np.expand_dims(image_array, axis=0)\n\n return image_array\n\n# ============================================================\n# Upload + Preprocess Endpoint\n# ============================================================\n\n@app.post(\"/upload-image\")\nasync def upload_image(file: UploadFile = File(...)):\n\n logger.info(\"Upload request received\")\n\n # 1️⃣ Validate extension\n validate_file_extension(file.filename)\n\n # 2️⃣ Validate size\n validate_file_size(file)\n\n # 3️⃣ Preprocess\n processed_image = preprocess_image(file)\n\n logger.info(f\"Image processed successfully. Shape: {processed_image.shape}\")\n\n return {\n \"filename\": file.filename,\n \"processed_shape\": processed_image.shape,\n \"dtype\": str(processed_image.dtype),\n \"min_value\": float(processed_image.min()),\n \"max_value\": float(processed_image.max())\n }\n \n'''","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.313805Z","iopub.execute_input":"2026-07-02T07:13:02.314236Z","iopub.status.idle":"2026-07-02T07:13:02.335575Z","shell.execute_reply.started":"2026-07-02T07:13:02.314190Z","shell.execute_reply":"2026-07-02T07:13:02.334564Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸ“ Build/Update \"app/main.py\" File","metadata":{}},{"cell_type":"code","source":"# ============================================\n# Build / Update app/main.py\n# ============================================\n\nmain_code = \"\"\"\n# ============================================================\n# Phase 5 β€” Image Upload & Preprocessing (Production Ready)\n# ============================================================\n\nfrom fastapi import HTTPException\nfrom PIL import Image\nimport io\n\n# ============================================================\n# Image Configuration\n# ============================================================\n\nALLOWED_EXTENSIONS = {\".jpg\", \".jpeg\", \".png\"}\nMAX_FILE_SIZE_MB = 5\nIMG_SIZE = (96, 96)\nDTYPE = np.float32\n\n# ============================================================\n# File Validation\n# ============================================================\n\ndef validate_file_extension(filename: str):\n \\\"\\\"\\\"\n Validate uploaded file extension.\n \\\"\\\"\\\"\n ext = os.path.splitext(filename)[1].lower()\n\n if ext not in ALLOWED_EXTENSIONS:\n raise HTTPException(\n status_code=400,\n detail=\"Invalid file type. Only JPG and PNG images are allowed.\"\n )\n\n\ndef validate_file_size(file: UploadFile):\n \\\"\\\"\\\"\n Validate uploaded file size.\n \\\"\\\"\\\"\n file.file.seek(0, os.SEEK_END)\n size = file.file.tell()\n file.file.seek(0)\n\n size_mb = size / (1024 * 1024)\n\n if size_mb > MAX_FILE_SIZE_MB:\n raise HTTPException(\n status_code=400,\n detail=f\"File too large. Max allowed size is {MAX_FILE_SIZE_MB}MB.\"\n )\n\n# ============================================================\n# Image Preprocessing (Aligned with Training Pipeline)\n# ============================================================\n\ndef preprocess_image(file: UploadFile) -> np.ndarray:\n \\\"\\\"\\\"\n Preprocess image for VGGNet inference.\n Steps:\n 1. Read image from UploadFile\n 2. Convert to RGB\n 3. Resize to training size\n 4. Convert to NumPy\n 5. Normalize (0-1 scaling)\n 6. Add batch dimension\n \\\"\\\"\\\"\n\n contents = file.file.read()\n image = Image.open(io.BytesIO(contents)).convert(\"RGB\")\n\n # Resize\n image = image.resize(IMG_SIZE)\n\n # Convert to NumPy\n image_array = np.array(image, dtype=DTYPE)\n\n # Normalize\n image_array = image_array / 255.0\n\n # Add batch dimension\n image_array = np.expand_dims(image_array, axis=0)\n\n return image_array\n\n# ============================================================\n# Upload + Preprocess Endpoint\n# ============================================================\n\n@app.post(\"/upload-image\")\nasync def upload_image(file: UploadFile = File(...)):\n\n logger.info(\"Upload request received\")\n\n # 1️⃣ Validate extension\n validate_file_extension(file.filename)\n\n # 2️⃣ Validate size\n validate_file_size(file)\n\n # 3️⃣ Preprocess\n processed_image = preprocess_image(file)\n\n logger.info(f\"Image processed successfully. Shape: {processed_image.shape}\")\n\n return {\n \"filename\": file.filename,\n \"processed_shape\": processed_image.shape,\n \"dtype\": str(processed_image.dtype),\n \"min_value\": float(processed_image.min()),\n \"max_value\": float(processed_image.max())\n }\n \n\"\"\"\n\nwith open(\"app/main.py\", \"w\") as f:\n f.write(main_code)\n\nprint(\"βœ… main.py updated successfully\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.336674Z","iopub.execute_input":"2026-07-02T07:13:02.337055Z","iopub.status.idle":"2026-07-02T07:13:02.359908Z","shell.execute_reply.started":"2026-07-02T07:13:02.337011Z","shell.execute_reply":"2026-07-02T07:13:02.358877Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸš€ Phase 5 β€” Image Upload & Preprocessing\n\n### 🎯 Objective\nImplement secure image upload and preprocessing pipeline aligned with the original training configuration.\n\n---\n\n### πŸ–Ό Implemented Features\n\n#### 1. Upload Endpoint\n- Implemented using `UploadFile`\n- Supports asynchronous file handling\n- Integrated with FastAPI lifecycle\n\n---\n\n#### 2. File Validation\n- Allowed formats: `.jpg`, `.jpeg`, `.png`\n- File size limit: 5MB\n- Proper HTTP 400 error handling\n\n---\n\n#### 3. Preprocessing Pipeline\n\nAligned exactly with training pipeline:\n\n1. Convert to RGB\n2. Resize to 96Γ—96\n3. Convert to NumPy (`float32`)\n4. Normalize to range [0,1]\n5. Add batch dimension\n\nFinal tensor shape:\n```\n\n(1, 96, 96, 3)\n\n```\n\n---\n\n#### 4. Logging\n- Upload request logged\n- Processed tensor shape logged\n- Production-style logging format\n\n---\n\n### Security Considerations\n\n- File extension validation\n- File size restriction\n- Controlled in-memory processing\n- No unsafe disk writes\n\n---\n\n### Result\n\nPhase 5 delivers:\n\n- Secure image upload\n- Deterministic preprocessing\n- Production-ready validation\n- Clean architecture integration\n- \n```\n","metadata":{}},{"cell_type":"markdown","source":"# 🧬 Phase 6 β€” /predict Endpoint","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Phase 6 β€” /predict Endpoint\n\n### Concept\nThis is the core inference endpoint.\n\n### Pipeline\n1. Receive image\n2. Preprocess\n3. Run inference\n4. Apply threshold\n5. Return JSON\n\n### Implementation Notes\n- No training logic\n- Inference-only\n- Clean JSON output\n","metadata":{}},{"cell_type":"markdown","source":"## 🧬 Predict Endpoint","metadata":{}},{"cell_type":"code","source":"'''\n# ============================================================\n# Phase 6 β€” Predict Endpoint (Production Grade)\n# ============================================================\n\nfrom pydantic import BaseModel\nimport time\nimport io\n\n# ============================================================\n# Prediction Configuration\n# ============================================================\n\nTHRESHOLD = 0.5\n\n# ============================================================\n# Pydantic Response Model\n# ============================================================\n\nclass PredictionResponse(BaseModel):\n filename: str\n probability: float\n threshold: float\n prediction: str\n confidence: float\n inference_time_ms: float\n\n# ============================================================\n# Predict Endpoint\n# ============================================================\n\n@app.post(\"/predict\", response_model=PredictionResponse)\nasync def predict(request: Request, file: UploadFile = File(...)):\n\n \\\"\\\"\\\"\n Perform full inference pipeline:\n - Validate file\n - Preprocess\n - Model inference\n - Thresholding\n - Structured JSON response\n \\\"\\\"\\\"\n\n start_time = time.time()\n\n logger.info(\"Prediction request received\")\n\n # ---------------------------\n # 1️⃣ Validate MIME type\n # ---------------------------\n if not file.content_type.startswith(\"image/\"):\n raise HTTPException(status_code=400, detail=\"Invalid image file\")\n\n # ---------------------------\n # 2️⃣ Validate extension\n # ---------------------------\n validate_file_extension(file.filename)\n\n # ---------------------------\n # 3️⃣ Validate size\n # ---------------------------\n validate_file_size(file)\n\n try:\n # ---------------------------\n # 4️⃣ Preprocessing\n # ---------------------------\n input_tensor = preprocess_image(file)\n\n # ---------------------------\n # 5️⃣ Model Inference\n # ---------------------------\n model = request.app.state.model\n\n preds = model_inference(model, input_tensor)\n probability = float(preds[0][0])\n\n # ---------------------------\n # 6️⃣ Thresholding\n # ---------------------------\n prediction = \"Cancer\" if probability >= THRESHOLD else \"Normal\"\n\n confidence = (\n probability if probability >= THRESHOLD\n else 1 - probability\n )\n\n # ---------------------------\n # 7️⃣ Measure inference time\n # ---------------------------\n inference_time = (time.time() - start_time) * 1000\n\n logger.info(\n f\"Prediction: {prediction} | \"\n f\"Probability: {probability:.4f} | \"\n f\"Time: {inference_time:.2f} ms\"\n )\n\n return PredictionResponse(\n filename=file.filename,\n probability=round(probability, 6),\n threshold=THRESHOLD,\n prediction=prediction,\n confidence=round(confidence, 6),\n inference_time_ms=round(inference_time, 2)\n )\n\n except Exception as e:\n logger.error(f\"Inference error: {e}\")\n raise HTTPException(status_code=500, detail=\"Internal server error\")\n\n'''","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.362391Z","iopub.execute_input":"2026-07-02T07:13:02.362790Z","iopub.status.idle":"2026-07-02T07:13:02.383171Z","shell.execute_reply.started":"2026-07-02T07:13:02.362760Z","shell.execute_reply":"2026-07-02T07:13:02.382288Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸ“ Build/Update \"app/main.py\" File","metadata":{}},{"cell_type":"code","source":"# ============================================\n# Build / Update app/main.py\n# ============================================\n\nmain_code = \"\"\"\n# ============================================================\n# main.py\n# Phase 4 β€” Model Loading & GPU Handling (Production Ready)\n# ============================================================\n\n# =============================\n# Standard Library Imports\n# =============================\nimport os\nimport logging\nimport tempfile\n\n# =============================\n# Third-Party Imports\n# =============================\nimport numpy as np\nimport tensorflow as tf\nfrom fastapi import FastAPI, UploadFile, File, Request\nfrom contextlib import asynccontextmanager\n\n# ============================================================\n# Logging Configuration\n# ============================================================\n\nlogging.basicConfig(\n level=logging.INFO,\n format=\"%(asctime)s - %(levelname)s - %(message)s\"\n)\n\nlogger = logging.getLogger(\"HistopathologyAPI\")\n\n# ============================================================\n# GPU Configuration\n# ============================================================\n\ndef configure_gpu():\n \\\"\\\"\\\"\n Configure GPU memory growth for stable inference.\n Prevents TensorFlow from allocating all GPU memory at once.\n \\\"\\\"\\\"\n logger.info(\"Checking GPU availability...\")\n gpus = tf.config.list_physical_devices('GPU')\n\n if gpus:\n try:\n for gpu in gpus:\n tf.config.experimental.set_memory_growth(gpu, True)\n logger.info(f\"βœ… {len(gpus)} GPU(s) configured with memory growth\")\n except RuntimeError as e:\n logger.error(f\"GPU configuration error: {e}\")\n else:\n logger.warning(\"⚠ No GPU detected, running on CPU\")\n\n# ============================================================\n# Model Warm-up\n# ============================================================\n\ndef warmup_model(model, input_shape=(1, 96, 96, 3)):\n \\\"\\\"\\\"\n Perform dummy forward pass to stabilize first inference latency.\n \\\"\\\"\\\"\n logger.info(\"Starting model warm-up...\")\n dummy_input = np.zeros(input_shape, dtype=np.float32)\n _ = model.predict(dummy_input, verbose=0)\n logger.info(\"πŸ”₯ Model warm-up completed\")\n\n# ============================================================\n# Optimized Inference Function\n# ============================================================\n\n@tf.function\ndef model_inference(model, batch):\n \\\"\\\"\\\"\n Optimized TensorFlow graph inference.\n \\\"\\\"\\\"\n return model(batch, training=False)\n\n# ============================================================\n# Prediction Wrapper\n# ============================================================\n\ndef predict_image_fast(model, image_tensor):\n \\\"\\\"\\\"\n Perform fast prediction with confidence extraction.\n \\\"\\\"\\\"\n image_tensor = tf.expand_dims(image_tensor, axis=0)\n preds = model_inference(model, image_tensor)\n\n prob = float(preds[0][0])\n\n label = \"Cancer\" if prob >= 0.5 else \"No Cancer\"\n confidence = prob if prob >= 0.5 else 1 - prob\n\n return label, confidence\n\n# ============================================================\n# Temporary File Handler\n# ============================================================\n\ndef save_temp_image(upload_file: UploadFile):\n \\\"\\\"\\\"\n Safely store uploaded image temporarily.\n \\\"\\\"\\\"\n suffix = os.path.splitext(upload_file.filename)[-1]\n with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:\n tmp.write(upload_file.file.read())\n return tmp.name\n\n# ============================================================\n# FastAPI Lifespan (Professional Lifecycle Injection)\n# ============================================================\n\n@asynccontextmanager\nasync def lifespan(app: FastAPI):\n\n # Startup\n\n logger.info(\"πŸš€ Starting application...\")\n\n # 1️⃣ GPU Setup\n configure_gpu()\n\n # 2️⃣ Load Model\n MODEL_PATH = os.getenv(\n \"MODEL_PATH\",\n \"app/models/VGGNet.keras\"\n )\n \n logger.info(f\"Loading model from: {MODEL_PATH}\")\n model = tf.keras.models.load_model(MODEL_PATH)\n model.trainable = False\n logger.info(\"βœ… Model loaded successfully\")\n\n # 3️⃣ Warm-up\n warmup_model(model)\n\n # 4️⃣ Store model in app state (Singleton)\n app.state.model = model\n\n yield\n\n # Shutdown\n logger.info(\"πŸ›‘ Application shutdown\")\n\n# ============================================================\n# FastAPI App Instance\n# ============================================================\n\napp = FastAPI(lifespan=lifespan)\n\n# ============================================================\n# Prediction Endpoint Example\n# ============================================================\n\n@app.post(\"/predict\")\nasync def predict(request: Request, file: UploadFile = File(...)):\n\n logger.info(\"Prediction request received\")\n\n model = request.app.state.model\n\n # For demo: dummy tensor (replace with real preprocessing)\n image_tensor = tf.zeros((96, 96, 3), dtype=tf.float32)\n\n label, confidence = predict_image_fast(model, image_tensor)\n\n logger.info(f\"Prediction result: {label} | Confidence: {confidence:.4f}\")\n\n return {\n \"label\": label,\n \"confidence\": round(confidence, 4)\n }\n\n\n\n\n# ============================================================\n# Phase 5 β€” Image Upload & Preprocessing (Production Ready)\n# ============================================================\n\nfrom fastapi import HTTPException\nfrom PIL import Image\nimport io\n\n# ============================================================\n# Image Configuration\n# ============================================================\n\nALLOWED_EXTENSIONS = {\".jpg\", \".jpeg\", \".png\"}\nMAX_FILE_SIZE_MB = 5\nIMG_SIZE = (96, 96)\nDTYPE = np.float32\n\n# ============================================================\n# File Validation\n# ============================================================\n\ndef validate_file_extension(filename: str):\n \\\"\\\"\\\"\n Validate uploaded file extension.\n \\\"\\\"\\\"\n ext = os.path.splitext(filename)[1].lower()\n\n if ext not in ALLOWED_EXTENSIONS:\n raise HTTPException(\n status_code=400,\n detail=\"Invalid file type. Only JPG and PNG images are allowed.\"\n )\n\n\ndef validate_file_size(file: UploadFile):\n \\\"\\\"\\\"\n Validate uploaded file size.\n \\\"\\\"\\\"\n file.file.seek(0, os.SEEK_END)\n size = file.file.tell()\n file.file.seek(0)\n\n size_mb = size / (1024 * 1024)\n\n if size_mb > MAX_FILE_SIZE_MB:\n raise HTTPException(\n status_code=400,\n detail=f\"File too large. Max allowed size is {MAX_FILE_SIZE_MB}MB.\"\n )\n\n# ============================================================\n# Image Preprocessing (Aligned with Training Pipeline)\n# ============================================================\n\ndef preprocess_image(file: UploadFile) -> np.ndarray:\n \\\"\\\"\\\"\n Preprocess image for VGGNet inference.\n Steps:\n 1. Read image from UploadFile\n 2. Convert to RGB\n 3. Resize to training size\n 4. Convert to NumPy\n 5. Normalize (0-1 scaling)\n 6. Add batch dimension\n \\\"\\\"\\\"\n\n contents = file.file.read()\n image = Image.open(io.BytesIO(contents)).convert(\"RGB\")\n\n # Resize\n image = image.resize(IMG_SIZE)\n\n # Convert to NumPy\n image_array = np.array(image, dtype=DTYPE)\n\n # Normalize\n image_array = image_array / 255.0\n\n # Add batch dimension\n image_array = np.expand_dims(image_array, axis=0)\n\n return image_array\n\n# ============================================================\n# Upload + Preprocess Endpoint\n# ============================================================\n\n@app.post(\"/upload-image\")\nasync def upload_image(file: UploadFile = File(...)):\n\n logger.info(\"Upload request received\")\n\n # 1️⃣ Validate extension\n validate_file_extension(file.filename)\n\n # 2️⃣ Validate size\n validate_file_size(file)\n\n # 3️⃣ Preprocess\n processed_image = preprocess_image(file)\n\n logger.info(f\"Image processed successfully. Shape: {processed_image.shape}\")\n\n return {\n \"filename\": file.filename,\n \"processed_shape\": processed_image.shape,\n \"dtype\": str(processed_image.dtype),\n \"min_value\": float(processed_image.min()),\n \"max_value\": float(processed_image.max())\n }\n\n\n\n\n\n# ============================================================\n# Phase 6 β€” Predict Endpoint (Production Grade)\n# ============================================================\n\nfrom pydantic import BaseModel\nimport time\nimport io\n\n# ============================================================\n# Prediction Configuration\n# ============================================================\n\nTHRESHOLD = 0.5\n\n# ============================================================\n# Pydantic Response Model\n# ============================================================\n\nclass PredictionResponse(BaseModel):\n filename: str\n probability: float\n threshold: float\n prediction: str\n confidence: float\n inference_time_ms: float\n\n# ============================================================\n# Predict Endpoint\n# ============================================================\n\n@app.post(\"/predict\", response_model=PredictionResponse)\nasync def predict(request: Request, file: UploadFile = File(...)):\n \\\"\\\"\\\"\n Perform full inference pipeline:\n - Validate file\n - Preprocess\n - Model inference\n - Thresholding\n - Structured JSON response\n \\\"\\\"\\\"\n\n start_time = time.time()\n\n logger.info(\"Prediction request received\")\n\n # ---------------------------\n # 1️⃣ Validate MIME type\n # ---------------------------\n if not file.content_type.startswith(\"image/\"):\n raise HTTPException(status_code=400, detail=\"Invalid image file\")\n\n # ---------------------------\n # 2️⃣ Validate extension\n # ---------------------------\n validate_file_extension(file.filename)\n\n # ---------------------------\n # 3️⃣ Validate size\n # ---------------------------\n validate_file_size(file)\n\n try:\n # ---------------------------\n # 4️⃣ Preprocessing\n # ---------------------------\n input_tensor = preprocess_image(file)\n\n # ---------------------------\n # 5️⃣ Model Inference\n # ---------------------------\n model = request.app.state.model\n\n preds = model_inference(model, input_tensor)\n probability = float(preds[0][0])\n\n # ---------------------------\n # 6️⃣ Thresholding\n # ---------------------------\n prediction = \"Cancer\" if probability >= THRESHOLD else \"Normal\"\n\n confidence = (\n probability if probability >= THRESHOLD\n else 1 - probability\n )\n\n # ---------------------------\n # 7️⃣ Measure inference time\n # ---------------------------\n inference_time = (time.time() - start_time) * 1000\n\n logger.info(\n f\"Prediction: {prediction} | \"\n f\"Probability: {probability:.4f} | \"\n f\"Time: {inference_time:.2f} ms\"\n )\n\n return PredictionResponse(\n filename=file.filename,\n probability=round(probability, 6),\n threshold=THRESHOLD,\n prediction=prediction,\n confidence=round(confidence, 6),\n inference_time_ms=round(inference_time, 2)\n )\n\n except Exception as e:\n logger.error(f\"Inference error: {e}\")\n raise HTTPException(status_code=500, detail=\"Internal server error\")\n\n\"\"\"\n\nwith open(\"app/main.py\", \"w\") as f:\n f.write(main_code)\n\nprint(\"βœ… main.py updated successfully\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.384672Z","iopub.execute_input":"2026-07-02T07:13:02.385216Z","iopub.status.idle":"2026-07-02T07:13:02.409333Z","shell.execute_reply.started":"2026-07-02T07:13:02.385118Z","shell.execute_reply":"2026-07-02T07:13:02.408452Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸš€ Phase 6 β€” `/predict` Endpoint\n\n### 🎯 Objective\nImplement a production-grade inference endpoint integrating preprocessing, model inference, thresholding, and structured response modeling.\n\n---\n\n### 🧬 Implemented Features\n\n#### 1. POST `/predict`\n- Accepts image file via `UploadFile`\n- Required parameter validation\n- MIME type verification\n\n---\n\n#### 2. Integrated Preprocessing\n- Resize to 96Γ—96\n- Convert to NumPy float32\n- Normalize to [0,1]\n- Add batch dimension\n- Fully aligned with training pipeline\n\n---\n\n#### 3. Model Inference\n- Model loaded once during app startup\n- Injected via FastAPI lifecycle\n- Optimized using `@tf.function`\n- Single-GPU inference optimized\n\n---\n\n#### 4. Thresholding Logic\n- Binary classification threshold = 0.5\n- Output: `Cancer` or `Normal`\n- Confidence score calculated accordingly\n\n\n---\n\n#### 5. Production Features\n\n- Request logging\n\n- Error handling\n\n- Inference time measurement\n\n- Clean architecture integration\n\n\n---\n\n\n### 🧠 Architecture Decision\n\n- Model loaded once at startup\n\n- No global variables\n\n- No disk writes\n\n- In-memory processing\n\n- Thread-safe inference\n\n### Result\n\nPhase 6 transforms the API into a complete Medical AI inference service, ready for:\n\n- Production hardening\n\n- Monitoring\n\n- Dockerization\n\n- Deployment","metadata":{}},{"cell_type":"markdown","source":"# πŸ”₯ Phase 7 β€” Production Hardening\n","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Phase 7 β€” Production Hardening\n\n### Concept\nMake the API robust and production-safe.\n\n### Implemented Features\n- Logging\n- Error handling\n- Request timing\n- Readiness & liveness checks\n\n### Why This Matters\n- Observability\n- Debuggability\n- Production reliability\n","metadata":{}},{"cell_type":"markdown","source":"## πŸ”₯ Production Hardening","metadata":{}},{"cell_type":"code","source":"# ============================================\n# Build app/main.py\n# ============================================\n\nmain_code = \"\"\"\n# ============================================\n# Phase 7 β€” Production Hardening\n# ============================================\n\nimport time\nimport logging\nimport numpy as np\nimport tensorflow as tf\n\nfrom fastapi import FastAPI, Request, HTTPException\nfrom fastapi.responses import JSONResponse\nfrom pydantic import BaseModel\nfrom typing import Optional\n\n# ============================================\n# Logging Configuration\n# ============================================\n\nlogging.basicConfig(\n level=logging.INFO,\n format=\"%(asctime)s | %(levelname)s | %(message)s\"\n)\n\nlogger = logging.getLogger(\"histopathology-api\")\n\n# ============================================\n# FastAPI App Initialization\n# ============================================\n\napp = FastAPI(\n title=\"Histopathologic Cancer Classification API\",\n version=\"1.0.0\"\n)\n\n# ============================================\n# Global Model Variable\n# ============================================\n\nmodel = None\n\n# ============================================\n# Pydantic Response Model\n# ============================================\n\nclass PredictionResponse(BaseModel):\n success: bool\n prediction: Optional[int]\n confidence: Optional[float]\n message: Optional[str]\n\n\n# ============================================\n# Load Model Function\n# ============================================\n\ndef load_model():\n global model\n logger.info(\"Loading model...\")\n model = tf.keras.models.load_model(\"VGGNet.keras\")\n logger.info(\"Model loaded successfully.\")\n\n\n# ============================================\n# Model Warm-Up\n# ============================================\n\ndef warmup_model():\n logger.info(\"Warming up model...\")\n dummy_input = np.zeros((1, 96, 96, 3), dtype=np.float32)\n _ = model.predict(dummy_input)\n logger.info(\"Model warm-up completed.\")\n\n\n# ============================================\n# Startup Lifecycle Event\n# ============================================\n\n@app.on_event(\"startup\")\nasync def startup_event():\n load_model()\n warmup_model()\n\n\n# ============================================\n# Global Exception Handler\n# ============================================\n\n@app.exception_handler(Exception)\nasync def global_exception_handler(request: Request, exc: Exception):\n logger.error(f\"Unhandled error: {exc}\")\n\n return JSONResponse(\n status_code=500,\n content={\n \"success\": False,\n \"message\": \"Internal Server Error\"\n }\n )\n\n\n# ============================================\n# Request Timing Middleware\n# ============================================\n\n@app.middleware(\"http\")\nasync def add_process_time_header(request: Request, call_next):\n\n start_time = time.time()\n\n response = await call_next(request)\n\n process_time = time.time() - start_time\n\n response.headers[\"X-Process-Time\"] = str(round(process_time, 4))\n\n logger.info(\n f\"{request.method} {request.url.path} \"\n f\"completed in {process_time:.4f}s\"\n )\n\n return response\n\n\n# ============================================\n# Health Endpoints\n# ============================================\n\n@app.get(\"/health/live\")\ndef liveness_probe():\n return {\"status\": \"alive\"}\n\n\n@app.get(\"/health/ready\")\ndef readiness_probe():\n if model is None:\n return JSONResponse(\n status_code=503,\n content={\"status\": \"model not ready\"}\n )\n\n return {\"status\": \"ready\"}\n\n\n# ============================================\n# Example Prediction Endpoint (Placeholder)\n# ============================================\n\n@app.post(\"/predict\", response_model=PredictionResponse)\ndef predict():\n\n if model is None:\n raise HTTPException(status_code=503, detail=\"Model not loaded\")\n\n try:\n dummy_input = np.zeros((1, 96, 96, 3), dtype=np.float32)\n prediction = model.predict(dummy_input)\n\n predicted_class = int(np.argmax(prediction))\n confidence = float(np.max(prediction))\n\n logger.info(f\"Prediction made: {predicted_class}\")\n\n return PredictionResponse(\n success=True,\n prediction=predicted_class,\n confidence=confidence,\n message=\"Prediction successful\"\n )\n\n except Exception as e:\n logger.error(f\"Prediction error: {e}\")\n raise HTTPException(status_code=500, detail=\"Prediction failed\")\n\n\"\"\"\n\nwith open(\"app/main.py\", \"w\") as f:\n f.write(main_code)\n\nprint(\"βœ… main.py (main_hardening) created successfully\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.410564Z","iopub.execute_input":"2026-07-02T07:13:02.411010Z","iopub.status.idle":"2026-07-02T07:13:02.434419Z","shell.execute_reply.started":"2026-07-02T07:13:02.410952Z","shell.execute_reply":"2026-07-02T07:13:02.433495Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"with open(\"app/main.py\", \"r\") as f:\n print(f.read())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.435549Z","iopub.execute_input":"2026-07-02T07:13:02.435931Z","iopub.status.idle":"2026-07-02T07:13:02.457857Z","shell.execute_reply.started":"2026-07-02T07:13:02.435897Z","shell.execute_reply":"2026-07-02T07:13:02.456717Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸš€ Phase 7 β€” Production Hardening\n\n### 🎯 Objective\n\nTransform the API from a development prototype into a production-grade service by introducing structured logging, middleware, error handling, health checks, and lifecycle management.\n\n---\n\n### πŸ”₯ Key Enhancements\n\n#### 1. Structured Logging\n\n* Configured Python logging module\n* Request and prediction logging\n* Error tracking for production debugging\n\n---\n\n#### 2. Global Exception Handler\n\n* Captures all unhandled exceptions\n* Prevents server crashes\n* Returns standardized JSON error response\n\n---\n\n#### 3. Request Timing Middleware\n\n* Measures API latency\n* Adds `X-Process-Time` header\n* Logs request duration\n\n---\n\n#### 4. Health & Readiness Endpoints\n\n| Endpoint | Purpose |\n| --------------- | ------------------------------------ |\n| `/health/live` | Liveness probe |\n| `/health/ready` | Readiness probe (model loaded check) |\n\nDesigned for Kubernetes / Docker deployment.\n\n---\n\n#### 5. Model Warm-Up Integration\n\nThe model is warmed up during startup lifecycle:\n\n```python\n@app.on_event(\"startup\")\nasync def startup_event():\n load_model()\n warmup_model()\n```\n\nThis ensures:\n\n* Stable first response\n* Reduced cold-start latency\n* Production reliability\n\n---\n\n### 🧠 Production Architecture Principles Applied\n\n* Separation of concerns\n* Lifecycle-based model loading\n* Structured API responses\n* Observability (logging + latency tracking)\n* Deployment readiness\n\n---\n\n### Result\n\nThe API is now:\n\n* Production-hardened\n* Monitoring-ready\n* Deployment-ready\n* Recruiter-impressive\n* Kubernetes-compatible\n","metadata":{}},{"cell_type":"markdown","source":"# πŸ”Œ Phase 8 β€” API Testing & Usage Demonstration","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Phase 8 β€” API Testing & Usage Demonstration\n\n### Concept\n\nDemonstrate real-world usability of the deployed FastAPI ML service through:\n\n* Auto-generated API documentation\n* Command-line testing\n* Python client integration\n* Batch inference support\n\n#### API Validation, Clients & Real Usage\n\n---\n\n### Python Client Example\n\nThis demonstrates integration capability with:\n\n* Web apps\n* Mobile apps\n* Backend systems\n* ML pipelines\n\n---\n\n### Batch Inference Endpoint\n\nAdded `/predict-batch` endpoint supporting:\n\n* Multiple file uploads\n* Aggregated predictions\n* Scalable inference design\n\n---\n\n### Swagger Documentation\n\nFastAPI automatically generates interactive documentation:\n\n```\n/docs\n```\n\nThis enables:\n\n* Endpoint visualization\n* Parameter testing\n* Response inspection\n* Automatic OpenAPI schema generation\n\n---\n\n### API Testing with curl\n\nExample:\n\n```bash\ncurl -X POST \"http://127.0.0.1:8000/predict\" \\\n -H \"accept: application/json\" \\\n -F \"file=@test_image.png\"\n```\n---","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Build \"client/client.py\" File","metadata":{}},{"cell_type":"code","source":"# ============================================\n# Build client/client.py\n# ============================================\n\n# Create client directory\nos.makedirs(\"client\", exist_ok=True)\n\n# Write client.py inside client/\nclient_code = \"\"\"\nimport requests\n\nAPI_URL = \"http://localhost:8000/predict\"\n\ndef predict_image(image_path: str):\n \\\"\\\"\\\"\n Send image to API and get prediction.\n \\\"\\\"\\\"\n\n with open(image_path, \"rb\") as img:\n response = requests.post(\n API_URL,\n files={\"file\": img}\n timeout=60\n )\n\n response.raise_for_status()\n\n return response.json()\n\n\nif __name__ == \"__main__\":\n result = predict_image(\"test_image.png\")\n print(result)\n\n\"\"\"\n\nwith open(\"client/client.py\", \"w\") as f:\n f.write(client_code)\n\nprint(\"βœ… client.py created successfully\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.459384Z","iopub.execute_input":"2026-07-02T07:13:02.459774Z","iopub.status.idle":"2026-07-02T07:13:02.476121Z","shell.execute_reply.started":"2026-07-02T07:13:02.459730Z","shell.execute_reply":"2026-07-02T07:13:02.475108Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"with open(\"client/client.py\", \"r\") as f:\n print(f.read())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.477777Z","iopub.execute_input":"2026-07-02T07:13:02.478290Z","iopub.status.idle":"2026-07-02T07:13:02.493835Z","shell.execute_reply.started":"2026-07-02T07:13:02.478250Z","shell.execute_reply":"2026-07-02T07:13:02.492565Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸ“ Build/Update \"app/main.py\" File","metadata":{}},{"cell_type":"code","source":"# ============================================\n# Build / Update app/main.py\n# ============================================\n\nmain_code = \"\"\"\n# ============================================================\n# main.py\n# Phase 4 β€” Model Loading & GPU Handling (Production Ready)\n# ============================================================\n\n# =============================\n# Standard Library Imports\n# =============================\nimport os\nimport logging\nimport tempfile\n\n# =============================\n# Third-Party Imports\n# =============================\nimport numpy as np\nimport tensorflow as tf\nfrom fastapi import FastAPI, UploadFile, File, Request\nfrom contextlib import asynccontextmanager\n\n# ============================================================\n# Logging Configuration\n# ============================================================\n\nlogging.basicConfig(\n level=logging.INFO,\n format=\"%(asctime)s - %(levelname)s - %(message)s\"\n)\n\nlogger = logging.getLogger(\"HistopathologyAPI\")\n\n# ============================================================\n# GPU Configuration\n# ============================================================\n\ndef configure_gpu():\n \\\"\\\"\\\"\n Configure GPU memory growth for stable inference.\n Prevents TensorFlow from allocating all GPU memory at once.\n \\\"\\\"\\\"\n logger.info(\"Checking GPU availability...\")\n gpus = tf.config.list_physical_devices('GPU')\n\n if gpus:\n try:\n for gpu in gpus:\n tf.config.experimental.set_memory_growth(gpu, True)\n logger.info(f\"βœ… {len(gpus)} GPU(s) configured with memory growth\")\n except RuntimeError as e:\n logger.error(f\"GPU configuration error: {e}\")\n else:\n logger.warning(\"⚠ No GPU detected, running on CPU\")\n\n# ============================================================\n# Model Warm-up\n# ============================================================\n\ndef warmup_model(model, input_shape=(1, 96, 96, 3)):\n \\\"\\\"\\\"\n Perform dummy forward pass to stabilize first inference latency.\n \\\"\\\"\\\"\n logger.info(\"Starting model warm-up...\")\n dummy_input = np.zeros(input_shape, dtype=np.float32)\n _ = model.predict(dummy_input, verbose=0)\n logger.info(\"πŸ”₯ Model warm-up completed\")\n\n# ============================================================\n# Optimized Inference Function\n# ============================================================\n\n@tf.function\ndef model_inference(model, batch):\n \\\"\\\"\\\"\n Optimized TensorFlow graph inference.\n \\\"\\\"\\\"\n return model(batch, training=False)\n\n# ============================================================\n# Prediction Wrapper\n# ============================================================\n\ndef predict_image_fast(model, image_tensor):\n \\\"\\\"\\\"\n Perform fast prediction with confidence extraction.\n \\\"\\\"\\\"\n image_tensor = tf.expand_dims(image_tensor, axis=0)\n preds = model_inference(model, image_tensor)\n\n prob = float(preds[0][0])\n\n label = \"Cancer\" if prob >= 0.5 else \"No Cancer\"\n confidence = prob if prob >= 0.5 else 1 - prob\n\n return label, confidence\n\n# ============================================================\n# Temporary File Handler\n# ============================================================\n\ndef save_temp_image(upload_file: UploadFile):\n \\\"\\\"\\\"\n Safely store uploaded image temporarily.\n \\\"\\\"\\\"\n suffix = os.path.splitext(upload_file.filename)[-1]\n with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:\n tmp.write(upload_file.file.read())\n return tmp.name\n\n# ============================================================\n# FastAPI Lifespan (Professional Lifecycle Injection)\n# ============================================================\n\n@asynccontextmanager\nasync def lifespan(app: FastAPI):\n\n # Startup\n\n logger.info(\"πŸš€ Starting application...\")\n\n # 1️⃣ GPU Setup\n configure_gpu()\n\n # 2️⃣ Load Model\n MODEL_PATH = os.getenv(\n \"MODEL_PATH\",\n \"app/models/VGGNet.keras\"\n )\n \n logger.info(f\"Loading model from: {MODEL_PATH}\")\n model = tf.keras.models.load_model(MODEL_PATH)\n model.trainable = False\n logger.info(\"βœ… Model loaded successfully\")\n\n # 3️⃣ Warm-up\n warmup_model(model)\n\n # 4️⃣ Store model in app state (Singleton)\n app.state.model = model\n\n yield\n\n # Shutdown\n logger.info(\"πŸ›‘ Application shutdown\")\n\n# ============================================================\n# FastAPI App Instance\n# ============================================================\n\napp = FastAPI(lifespan=lifespan)\n\n# ============================================================\n# Prediction Endpoint Example\n# ============================================================\n\n@app.post(\"/predict\")\nasync def predict(request: Request, file: UploadFile = File(...)):\n\n logger.info(\"Prediction request received\")\n\n model = request.app.state.model\n\n # For demo: dummy tensor (replace with real preprocessing)\n image_tensor = tf.zeros((96, 96, 3), dtype=tf.float32)\n\n label, confidence = predict_image_fast(model, image_tensor)\n\n logger.info(f\"Prediction result: {label} | Confidence: {confidence:.4f}\")\n\n return {\n \"label\": label,\n \"confidence\": round(confidence, 4)\n }\n\n\n\n\n# ============================================================\n# Phase 5 β€” Image Upload & Preprocessing (Production Ready)\n# ============================================================\n\nfrom fastapi import HTTPException\nfrom PIL import Image\nimport io\n\n# ============================================================\n# Image Configuration\n# ============================================================\n\nALLOWED_EXTENSIONS = {\".jpg\", \".jpeg\", \".png\"}\nMAX_FILE_SIZE_MB = 5\nIMG_SIZE = (96, 96)\nDTYPE = np.float32\n\n# ============================================================\n# File Validation\n# ============================================================\n\ndef validate_file_extension(filename: str):\n \\\"\\\"\\\"\n Validate uploaded file extension.\n \\\"\\\"\\\"\n ext = os.path.splitext(filename)[1].lower()\n\n if ext not in ALLOWED_EXTENSIONS:\n raise HTTPException(\n status_code=400,\n detail=\"Invalid file type. Only JPG and PNG images are allowed.\"\n )\n\n\ndef validate_file_size(file: UploadFile):\n \\\"\\\"\\\"\n Validate uploaded file size.\n \\\"\\\"\\\"\n file.file.seek(0, os.SEEK_END)\n size = file.file.tell()\n file.file.seek(0)\n\n size_mb = size / (1024 * 1024)\n\n if size_mb > MAX_FILE_SIZE_MB:\n raise HTTPException(\n status_code=400,\n detail=f\"File too large. Max allowed size is {MAX_FILE_SIZE_MB}MB.\"\n )\n\n# ============================================================\n# Image Preprocessing (Aligned with Training Pipeline)\n# ============================================================\n\ndef preprocess_image(file: UploadFile) -> np.ndarray:\n \\\"\\\"\\\"\n Preprocess image for VGGNet inference.\n Steps:\n 1. Read image from UploadFile\n 2. Convert to RGB\n 3. Resize to training size\n 4. Convert to NumPy\n 5. Normalize (0-1 scaling)\n 6. Add batch dimension\n \\\"\\\"\\\"\n\n contents = file.file.read()\n image = Image.open(io.BytesIO(contents)).convert(\"RGB\")\n\n # Resize\n image = image.resize(IMG_SIZE)\n\n # Convert to NumPy\n image_array = np.array(image, dtype=DTYPE)\n\n # Normalize\n image_array = image_array / 255.0\n\n # Add batch dimension\n image_array = np.expand_dims(image_array, axis=0)\n\n return image_array\n\n# ============================================================\n# Upload + Preprocess Endpoint\n# ============================================================\n\n@app.post(\"/upload-image\")\nasync def upload_image(file: UploadFile = File(...)):\n\n logger.info(\"Upload request received\")\n\n # 1️⃣ Validate extension\n validate_file_extension(file.filename)\n\n # 2️⃣ Validate size\n validate_file_size(file)\n\n # 3️⃣ Preprocess\n processed_image = preprocess_image(file)\n\n logger.info(f\"Image processed successfully. Shape: {processed_image.shape}\")\n\n return {\n \"filename\": file.filename,\n \"processed_shape\": processed_image.shape,\n \"dtype\": str(processed_image.dtype),\n \"min_value\": float(processed_image.min()),\n \"max_value\": float(processed_image.max())\n }\n\n\n\n\n\n# ============================================================\n# Phase 6 β€” Predict Endpoint (Production Grade)\n# ============================================================\n\nfrom pydantic import BaseModel\nimport time\nimport io\n\n# ============================================================\n# Prediction Configuration\n# ============================================================\n\nTHRESHOLD = 0.5\n\n# ============================================================\n# Pydantic Response Model\n# ============================================================\n\nclass PredictionResponse(BaseModel):\n filename: str\n probability: float\n threshold: float\n prediction: str\n confidence: float\n inference_time_ms: float\n\n# ============================================================\n# Predict Endpoint\n# ============================================================\n\n@app.post(\"/predict\", response_model=PredictionResponse)\nasync def predict(request: Request, file: UploadFile = File(...)):\n \\\"\\\"\\\"\n Perform full inference pipeline:\n - Validate file\n - Preprocess\n - Model inference\n - Thresholding\n - Structured JSON response\n \\\"\\\"\\\"\n\n start_time = time.time()\n\n logger.info(\"Prediction request received\")\n\n # ---------------------------\n # 1️⃣ Validate MIME type\n # ---------------------------\n if not file.content_type.startswith(\"image/\"):\n raise HTTPException(status_code=400, detail=\"Invalid image file\")\n\n # ---------------------------\n # 2️⃣ Validate extension\n # ---------------------------\n validate_file_extension(file.filename)\n\n # ---------------------------\n # 3️⃣ Validate size\n # ---------------------------\n validate_file_size(file)\n\n try:\n # ---------------------------\n # 4️⃣ Preprocessing\n # ---------------------------\n input_tensor = preprocess_image(file)\n\n # ---------------------------\n # 5️⃣ Model Inference\n # ---------------------------\n model = request.app.state.model\n\n preds = model_inference(model, input_tensor)\n probability = float(preds[0][0])\n\n # ---------------------------\n # 6️⃣ Thresholding\n # ---------------------------\n prediction = \"Cancer\" if probability >= THRESHOLD else \"Normal\"\n\n confidence = (\n probability if probability >= THRESHOLD\n else 1 - probability\n )\n\n # ---------------------------\n # 7️⃣ Measure inference time\n # ---------------------------\n inference_time = (time.time() - start_time) * 1000\n\n logger.info(\n f\"Prediction: {prediction} | \"\n f\"Probability: {probability:.4f} | \"\n f\"Time: {inference_time:.2f} ms\"\n )\n\n return PredictionResponse(\n filename=file.filename,\n probability=round(probability, 6),\n threshold=THRESHOLD,\n prediction=prediction,\n confidence=round(confidence, 6),\n inference_time_ms=round(inference_time, 2)\n )\n\n except Exception as e:\n logger.error(f\"Inference error: {e}\")\n raise HTTPException(status_code=500, detail=\"Internal server error\")\n\n\n\n# ============================================================\n# Phase 8 β€” Batch Inference (Professional Upgrade)\n# ============================================================\n\nfrom typing import List\nfrom fastapi import UploadFile, File\n\n@app.post(\"/predict-batch\")\nasync def predict_batch(files: List[UploadFile] = File(...)):\n\n predictions = []\n\n for file in files:\n contents = await file.read()\n\n # TODO: image preprocessing\n dummy_input = np.zeros((1, 96, 96, 3), dtype=np.float32)\n prediction = model.predict(dummy_input)\n\n predicted_class = int(np.argmax(prediction))\n confidence = float(np.max(prediction))\n\n predictions.append({\n \"filename\": file.filename,\n \"prediction\": predicted_class,\n \"confidence\": confidence\n })\n\n return {\"results\": predictions}\n\n\"\"\"\n\nwith open(\"app/main.py\", \"w\") as f:\n f.write(main_code)\n\nprint(\"βœ… main.py updated successfully\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.495640Z","iopub.execute_input":"2026-07-02T07:13:02.496090Z","iopub.status.idle":"2026-07-02T07:13:02.516030Z","shell.execute_reply.started":"2026-07-02T07:13:02.496053Z","shell.execute_reply":"2026-07-02T07:13:02.515099Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸš€ Phase 8 β€” API Testing & Usage Demonstration\n\n### 🎯 Objective\n\nPrepare the FastAPI service for real-world usage and testing.\n\n> ⚠️ Note: Kaggle environment does NOT support running production servers.\n> This phase focuses on API design and readiness, not execution.\n\n---\n\n### πŸ“¦ Batch Inference Support\n\nAdded `/predict-batch` endpoint:\n\n* Supports multiple image uploads\n* Returns aggregated predictions\n* Enables scalable inference\n\n\n---\n\n### πŸ”Œ API Testing Strategy\n\nThe API can be tested externally using:\n\n* `curl` (command-line testing)\n* Python client (`requests`)\n* Swagger UI (`/docs`)\n\n⚠️ These tests are intended for:\n\n* Local environment\n* Docker deployment\n* Cloud servers\n\n---\n\n### 🐍 Python Client Integration\n\nA sample client script is provided to demonstrate:\n\n* Programmatic API usage\n* Integration with external sy\nstems\n* Real-world deployment readiness\n\n---\n\n### 🧠 Production Awareness\n\nThis phase ensures:\n\n* API usability\n* Client compatibility\n* Scalable inference design","metadata":{}},{"cell_type":"markdown","source":"# 🩺 Phase 9 β€” Medical AI Best Practices","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Phase 9 β€” Medical AI Best Practices\n\n### Concept\nMedical AI requires extra safety and transparency.\n#### Safe, Interpretable & Responsible Deployment\n\n---\n\n### Implemented Practices\n- Confidence interpretation\n- Conservative thresholds\n- Safe default behavior\n\n---\n\n### ⚠ Important Note\nPredictions are **not clinical diagnoses**.\n\nThey are decision-support signals only.\n","metadata":{}},{"cell_type":"markdown","source":"## πŸ“Œ Why False Negatives Matter More Than False Positives in Medical AI\n\n\n\nMedical image classification is fundamentally different from many conventional computer vision tasks.\n\n\n\nIn many real-world applications, different types of prediction errors do not have the same consequences.\n\n\n\nUnderstanding this distinction is essential when designing AI systems intended for healthcare research.\n\n\n\n---\n\n\n\n### Two Types of Errors\n\n\n\nBinary classification produces two major error types:\n\n\n\n* **False Positive (FP):** A healthy sample is predicted as positive.\n\n* **False Negative (FN):** A diseased sample is predicted as healthy.\n\n\n\nAlthough both reduce model performance, their clinical impact is very different.\n\n\n\n---\n\n\n\n### Why False Negatives Are More Critical\n\n\n\nA false negative may delay further medical evaluation or additional diagnostic procedures.\n\n\n\nFor this reason, many medical AI systems prioritize minimizing false negatives, even if doing so slightly increases the number of false positives.\n\n\n\nA false positive usually results in additional review or testing.\n\n\n\nA false negative may result in a missed suspicious case.\n\n\n\n---\n\n\n\n### Confidence Levels\n\n\n\nThis project categorizes prediction confidence into three levels.\n\n\n| Confidence | Interpretation |\n| ---------- | --------------- |\n| β‰₯ 0.90 | High confidence |\n| 0.60–0.90 | Uncertain |\n| < 0.60 | Low confidence |\n\n\nPredictions with lower confidence should be interpreted more cautiously.\n\n\n\n---\n\n\n\n### πŸ“Œ Important Note\n\n\n\n> **Model confidence reflects the model's certainty about its prediction, not a confirmed medical diagnosis.**\n\n\n\nConfidence is a statistical property of the model and should never be interpreted as clinical certainty.\n\n\n\n---\n\n\n\n#### Conservative Decision Strategy\n\n\n\nRather than treating every prediction equally, confidence information can help identify cases that deserve additional review.\n\n\n\nThis conservative approach aligns better with many medical AI applications, where avoiding missed suspicious cases is often more important than maximizing overall accuracy.\n\n\n\n---\n\n\n\n#### Disclaimer\n\n\n\nThis project is intended for research and educational purposes only and must not be used for clinical diagnosis or medical decision-making.","metadata":{}},{"cell_type":"markdown","source":"## 🩺 Step 1 β€” Confidence Interpretation Logic","metadata":{}},{"cell_type":"code","source":"'''\n# ============================================\n# Confidence Interpretation Logic (Medical AI)\n# ============================================\n\ndef interpret_confidence(confidence: float) -> str:\n \"\"\"\n Interpret model confidence in a medically conservative way.\n This does NOT represent true disease probability.\n \"\"\"\n\n if confidence >= 0.90:\n return \"high_confidence\"\n elif confidence >= 0.60:\n return \"uncertain\"\n else:\n return \"low_confidence\"\n\n'''","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.517395Z","iopub.execute_input":"2026-07-02T07:13:02.518482Z","iopub.status.idle":"2026-07-02T07:13:02.540252Z","shell.execute_reply.started":"2026-07-02T07:13:02.518449Z","shell.execute_reply":"2026-07-02T07:13:02.539046Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## 🩺 Step 2 β€” Safe Thresholding in Prediction Endpoint","metadata":{}},{"cell_type":"code","source":"'''\n@app.post(\"/predict\", response_model=PredictionResponse)\ndef predict():\n\n if model is None:\n raise HTTPException(status_code=503, detail=\"Model not loaded\")\n\n try:\n dummy_input = np.zeros((1, 96, 96, 3), dtype=np.float32)\n prediction = model.predict(dummy_input)\n\n predicted_class = int(np.argmax(prediction))\n confidence = float(np.max(prediction))\n\n confidence_level = interpret_confidence(confidence)\n\n logger.info(\n f\"Prediction: {predicted_class} | \"\n f\"Confidence: {confidence:.4f} | \"\n f\"Level: {confidence_level}\"\n )\n\n return PredictionResponse(\n success=True,\n prediction=predicted_class,\n confidence=confidence,\n message=f\"Model confidence level: {confidence_level}. \"\n f\"This is NOT a clinical diagnosis.\"\n )\n\n except Exception as e:\n logger.error(f\"Prediction error: {e}\")\n raise HTTPException(status_code=500, detail=\"Prediction failed\")\n'''","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.541608Z","iopub.execute_input":"2026-07-02T07:13:02.542000Z","iopub.status.idle":"2026-07-02T07:13:02.565592Z","shell.execute_reply.started":"2026-07-02T07:13:02.541958Z","shell.execute_reply":"2026-07-02T07:13:02.564539Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## 🩺 Step 3 β€” Update Response Model","metadata":{}},{"cell_type":"code","source":"'''\nclass PredictionResponse(BaseModel):\n success: bool\n prediction: Optional[int]\n confidence: Optional[float]\n confidence_level: Optional[str]\n message: Optional[str]\n'''","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.566650Z","iopub.execute_input":"2026-07-02T07:13:02.566919Z","iopub.status.idle":"2026-07-02T07:13:02.586056Z","shell.execute_reply.started":"2026-07-02T07:13:02.566892Z","shell.execute_reply":"2026-07-02T07:13:02.585247Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## 🩺 Step 4 β€” Honest Explainability Section","metadata":{}},{"cell_type":"code","source":"'''\n@app.get(\"/model-info\")\ndef model_info():\n return {\n \"architecture\": \"Convolutional Neural Network (VGG-based)\",\n \"input_shape\": \"96x96 RGB patch\",\n \"explainability\": (\n \"This model does not provide pixel-level explanations. \"\n \"Predictions are based on learned visual patterns such as \"\n \"texture and structural irregularities.\"\n ),\n \"medical_disclaimer\": (\n \"This system is intended for research and decision-support only. \"\n \"It is NOT a clinical diagnostic tool.\"\n )\n }\n\n'''","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.590831Z","iopub.execute_input":"2026-07-02T07:13:02.591285Z","iopub.status.idle":"2026-07-02T07:13:02.604104Z","shell.execute_reply.started":"2026-07-02T07:13:02.591252Z","shell.execute_reply":"2026-07-02T07:13:02.603226Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# πŸ“ Build/Update \"app/main.py\" File","metadata":{}},{"cell_type":"code","source":"# ============================================\n# Build / Update app/main.py\n# ============================================\n\nmain_code = \"\"\"\n# ============================================================\n# main.py\n# Phase 4 β€” Model Loading & GPU Handling (Production Ready)\n# ============================================================\n\n# =============================\n# Standard Library Imports\n# =============================\nimport os\nimport logging\nimport tempfile\n\n# =============================\n# Third-Party Imports\n# =============================\nimport numpy as np\nimport tensorflow as tf\nfrom fastapi import FastAPI, UploadFile, File, Request\nfrom contextlib import asynccontextmanager\n\n# ============================================================\n# Logging Configuration\n# ============================================================\n\nlogging.basicConfig(\n level=logging.INFO,\n format=\"%(asctime)s - %(levelname)s - %(message)s\"\n)\n\nlogger = logging.getLogger(\"HistopathologyAPI\")\n\n# ============================================================\n# GPU Configuration\n# ============================================================\n\ndef configure_gpu():\n \\\"\\\"\\\"\n Configure GPU memory growth for stable inference.\n Prevents TensorFlow from allocating all GPU memory at once.\n \\\"\\\"\\\"\n logger.info(\"Checking GPU availability...\")\n gpus = tf.config.list_physical_devices('GPU')\n\n if gpus:\n try:\n for gpu in gpus:\n tf.config.experimental.set_memory_growth(gpu, True)\n logger.info(f\"βœ… {len(gpus)} GPU(s) configured with memory growth\")\n except RuntimeError as e:\n logger.error(f\"GPU configuration error: {e}\")\n else:\n logger.warning(\"⚠ No GPU detected, running on CPU\")\n\n# ============================================================\n# Model Warm-up\n# ============================================================\n\ndef warmup_model(model, input_shape=(1, 96, 96, 3)):\n \\\"\\\"\\\"\n Perform dummy forward pass to stabilize first inference latency.\n \\\"\\\"\\\"\n logger.info(\"Starting model warm-up...\")\n dummy_input = np.zeros(input_shape, dtype=np.float32)\n _ = model.predict(dummy_input, verbose=0)\n logger.info(\"πŸ”₯ Model warm-up completed\")\n\n# ============================================================\n# Optimized Inference Function\n# ============================================================\n\n@tf.function\ndef model_inference(model, batch):\n \\\"\\\"\\\"\n Optimized TensorFlow graph inference.\n \\\"\\\"\\\"\n return model(batch, training=False)\n\n# ============================================================\n# Prediction Wrapper\n# ============================================================\n\ndef predict_image_fast(model, image_tensor):\n \\\"\\\"\\\"\n Perform fast prediction with confidence extraction.\n \\\"\\\"\\\"\n image_tensor = tf.expand_dims(image_tensor, axis=0)\n preds = model_inference(model, image_tensor)\n\n prob = float(preds[0][0])\n\n label = \"Cancer\" if prob >= 0.5 else \"No Cancer\"\n confidence = prob if prob >= 0.5 else 1 - prob\n\n return label, confidence\n\n# ============================================================\n# Temporary File Handler\n# ============================================================\n\ndef save_temp_image(upload_file: UploadFile):\n \\\"\\\"\\\"\n Safely store uploaded image temporarily.\n \\\"\\\"\\\"\n suffix = os.path.splitext(upload_file.filename)[-1]\n with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:\n tmp.write(upload_file.file.read())\n return tmp.name\n\n# ============================================================\n# FastAPI Lifespan (Professional Lifecycle Injection)\n# ============================================================\n\n@asynccontextmanager\nasync def lifespan(app: FastAPI):\n\n # Startup\n\n logger.info(\"πŸš€ Starting application...\")\n\n # 1️⃣ GPU Setup\n configure_gpu()\n\n # 2️⃣ Load Model\n MODEL_PATH = os.getenv(\n \"MODEL_PATH\",\n \"models/VGGNet.keras\"\n )\n \n logger.info(f\"Loading model from: {MODEL_PATH}\")\n model = tf.keras.models.load_model(MODEL_PATH)\n model.trainable = False\n logger.info(\"βœ… Model loaded successfully\")\n\n # 3️⃣ Warm-up\n warmup_model(model)\n\n # 4️⃣ Store model in app state (Singleton)\n app.state.model = model\n\n yield\n\n # Shutdown\n logger.info(\"πŸ›‘ Application shutdown\")\n\n# ============================================================\n# FastAPI App Instance\n# ============================================================\n\napp = FastAPI(lifespan=lifespan)\n\n# ============================================================\n# Prediction Endpoint Example\n# ============================================================\n\n@app.post(\n \"/predict\",\n summary=\"Predict cancer from a histopathology image\",\n description=\"Uploads a histopathology image and returns the predicted class with confidence.\",\n tags=[\"Inference\"],\n response_model=PredictionResponse,\n)\nasync def predict(request: Request, file: UploadFile = File(...)):\n\n logger.info(\"Prediction request received\")\n\n model = request.app.state.model\n\n # For demo: dummy tensor (replace with real preprocessing)\n image_tensor = tf.zeros((96, 96, 3), dtype=tf.float32)\n\n label, confidence = predict_image_fast(model, image_tensor)\n\n logger.info(f\"Prediction result: {label} | Confidence: {confidence:.4f}\")\n\n return {\n \"label\": label,\n \"confidence\": round(confidence, 4)\n }\n\n\n\n\n# ============================================================\n# Phase 5 β€” Image Upload & Preprocessing (Production Ready)\n# ============================================================\n\nfrom fastapi import HTTPException\nfrom PIL import Image\nimport io\n\n# ============================================================\n# Image Configuration\n# ============================================================\n\nALLOWED_EXTENSIONS = {\".jpg\", \".jpeg\", \".png\"}\nMAX_FILE_SIZE_MB = 5\nIMG_SIZE = (96, 96)\nDTYPE = np.float32\n\n# ============================================================\n# File Validation\n# ============================================================\n\ndef validate_file_extension(filename: str):\n \\\"\\\"\\\"\n Validate uploaded file extension.\n \\\"\\\"\\\"\n ext = os.path.splitext(filename)[1].lower()\n\n if ext not in ALLOWED_EXTENSIONS:\n raise HTTPException(\n status_code=400,\n detail=\"Invalid file type. Only JPG and PNG images are allowed.\"\n )\n\n\ndef validate_file_size(file: UploadFile):\n \\\"\\\"\\\"\n Validate uploaded file size.\n \\\"\\\"\\\"\n file.file.seek(0, os.SEEK_END)\n size = file.file.tell()\n file.file.seek(0)\n\n size_mb = size / (1024 * 1024)\n\n if size_mb > MAX_FILE_SIZE_MB:\n raise HTTPException(\n status_code=400,\n detail=f\"File too large. Max allowed size is {MAX_FILE_SIZE_MB}MB.\"\n )\n\n# ============================================================\n# Image Preprocessing (Aligned with Training Pipeline)\n# ============================================================\n\ndef preprocess_image(file: UploadFile) -> np.ndarray:\n \\\"\\\"\\\"\n Preprocess image for VGGNet inference.\n Steps:\n 1. Read image from UploadFile\n 2. Convert to RGB\n 3. Resize to training size\n 4. Convert to NumPy\n 5. Normalize (0-1 scaling)\n 6. Add batch dimension\n \\\"\\\"\\\"\n\n contents = file.file.read()\n image = Image.open(io.BytesIO(contents)).convert(\"RGB\")\n\n # Resize\n image = image.resize(IMG_SIZE)\n\n # Convert to NumPy\n image_array = np.array(image, dtype=DTYPE)\n\n # Normalize\n image_array = image_array / 255.0\n\n # Add batch dimension\n image_array = np.expand_dims(image_array, axis=0)\n\n return image_array\n\n# ============================================================\n# Upload + Preprocess Endpoint\n# ============================================================\n\n@app.post(\"/upload-image\")\nasync def upload_image(file: UploadFile = File(...)):\n\n logger.info(\"Upload request received\")\n\n # 1️⃣ Validate extension\n validate_file_extension(file.filename)\n\n # 2️⃣ Validate size\n validate_file_size(file)\n\n # 3️⃣ Preprocess\n processed_image = preprocess_image(file)\n\n logger.info(f\"Image processed successfully. Shape: {processed_image.shape}\")\n\n return {\n \"filename\": file.filename,\n \"processed_shape\": processed_image.shape,\n \"dtype\": str(processed_image.dtype),\n \"min_value\": float(processed_image.min()),\n \"max_value\": float(processed_image.max())\n }\n\n\n\n\n\n# ============================================================\n# Phase 6 β€” Predict Endpoint (Production Grade)\n# ============================================================\n\nfrom pydantic import BaseModel\nimport time\nimport io\n\n# ============================================================\n# Prediction Configuration\n# ============================================================\n\nTHRESHOLD = 0.5\n\n# ============================================================\n# Pydantic Response Model\n# ============================================================\n\nclass PredictionResponse(BaseModel):\n filename: str\n probability: float\n threshold: float\n prediction: str\n confidence: float\n inference_time_ms: float\n\n# ============================================================\n# Predict Endpoint\n# ============================================================\n\n@app.post(\n \"/predict\",\n summary=\"Predict cancer from a histopathology image\",\n description=\"Uploads a histopathology image and returns the predicted class with confidence.\",\n tags=[\"Inference\"],\n response_model=PredictionResponse,\n)\nasync def predict(request: Request, file: UploadFile = File(...)):\n \\\"\\\"\\\"\n Perform full inference pipeline:\n - Validate file\n - Preprocess\n - Model inference\n - Thresholding\n - Structured JSON response\n \\\"\\\"\\\"\n\n start_time = time.time()\n\n logger.info(\"Prediction request received\")\n\n # ---------------------------\n # 1️⃣ Validate MIME type\n # ---------------------------\n if not file.content_type.startswith(\"image/\"):\n raise HTTPException(status_code=400, detail=\"Invalid image file\")\n\n # ---------------------------\n # 2️⃣ Validate extension\n # ---------------------------\n validate_file_extension(file.filename)\n\n # ---------------------------\n # 3️⃣ Validate size\n # ---------------------------\n validate_file_size(file)\n\n try:\n # ---------------------------\n # 4️⃣ Preprocessing\n # ---------------------------\n input_tensor = preprocess_image(file)\n\n # ---------------------------\n # 5️⃣ Model Inference\n # ---------------------------\n model = request.app.state.model\n\n preds = model_inference(model, input_tensor)\n probability = float(preds[0][0])\n\n # ---------------------------\n # 6️⃣ Thresholding\n # ---------------------------\n prediction = \"Cancer\" if probability >= THRESHOLD else \"Normal\"\n\n confidence = (\n probability if probability >= THRESHOLD\n else 1 - probability\n )\n\n # ---------------------------\n # 7️⃣ Measure inference time\n # ---------------------------\n inference_time = (time.time() - start_time) * 1000\n\n logger.info(\n f\"Prediction: {prediction} | \"\n f\"Probability: {probability:.4f} | \"\n f\"Time: {inference_time:.2f} ms\"\n )\n\n return PredictionResponse(\n filename=file.filename,\n probability=round(probability, 6),\n threshold=THRESHOLD,\n prediction=prediction,\n confidence=round(confidence, 6),\n inference_time_ms=round(inference_time, 2)\n )\n\n except Exception as e:\n logger.error(f\"Inference error: {e}\")\n raise HTTPException(status_code=500, detail=\"Internal server error\")\n\n\n\n# ============================================================\n# Phase 8 β€” Batch Inference (Professional Upgrade)\n# ============================================================\n\nfrom typing import List\nfrom fastapi import UploadFile, File\n\n@app.post(\n \"/predict-batch\",\n summary=\"Predict cancer from a histopathology image\",\n description=\"Uploads a histopathology image and returns the predicted class with confidence.\",\n tags=[\"Inference\"],\n response_model=PredictionResponse,\n)\nasync def predict_batch(files: List[UploadFile] = File(...)):\n\n predictions = []\n\n for file in files:\n contents = await file.read()\n\n # TODO: image preprocessing\n dummy_input = np.zeros((1, 96, 96, 3), dtype=np.float32)\n prediction = model.predict(dummy_input)\n\n predicted_class = int(np.argmax(prediction))\n confidence = float(np.max(prediction))\n\n predictions.append({\n \"filename\": file.filename,\n \"prediction\": predicted_class,\n \"confidence\": confidence\n })\n\n return {\"results\": predictions}\n\n\n# ============================================================\n# Phase 9 β€” Medical AI Best Practices\n# ============================================================\n\n# ============================================\n# Confidence Interpretation Logic (Medical AI)\n# ============================================\n\ndef interpret_confidence(confidence: float) -> str:\n /\"/\"/\"\n Interpret model confidence in a medically conservative way.\n This does NOT represent true disease probability.\n /\"/\"/\"\n\n if confidence >= 0.90:\n return \"high_confidence\"\n elif confidence >= 0.60:\n return \"uncertain\"\n else:\n return \"low_confidence\"\n\n\n# ============================================================\n# Model Information\n# ============================================================\n@app.get(\"/model-info\")\ndef model_info():\n return {\n \"architecture\": \"Convolutional Neural Network (VGG-based)\",\n \"input_shape\": \"96x96 RGB patch\",\n \"explainability\": (\n \"This model does not provide pixel-level explanations. \"\n \"Predictions are based on learned visual patterns such as \"\n \"texture and structural irregularities.\"\n ),\n \"medical_disclaimer\": (\n \"This system is intended for research and decision-support only. \"\n \"It is NOT a clinical diagnostic tool.\"\n )\n }\n\n\"\"\"\n\nwith open(\"app/main.py\", \"w\") as f:\n f.write(main_code)\n\nprint(\"βœ… main.py updated successfully\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.605660Z","iopub.execute_input":"2026-07-02T07:13:02.606041Z","iopub.status.idle":"2026-07-02T07:13:02.629426Z","shell.execute_reply.started":"2026-07-02T07:13:02.606010Z","shell.execute_reply":"2026-07-02T07:13:02.628370Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## 🩺 Phase 9 β€” Medical AI Best Practices\n\n### 🎯 Objective\n\nEnhance the ML API with medical-grade safety principles, including:\n\n* Conservative confidence interpretation\n* Safe thresholding strategy\n* Honest explainability\n* Clear medical disclaimers\n\n---\n\n### ⚠️ Confidence β‰  Clinical Diagnosis\n\nModel output represents model confidence, not medical probability.\n\nExample:\n\n```json\n{\n\n \"prediction\": 1,\n \"confidence\": 0.92\n}\n```\n\n#### πŸ“Œ This means:\n\n> The model strongly detected patterns similar to cancerous samples.\n\nIt does NOT mean:\n\n> The patient has cancer.\n\n\n---\n\n### πŸ›‘ Conservative Thresholding Strategy\n\n| Confidence | Interpretation |\n| ---------- | --------------- |\n| β‰₯ 0.90 | High confidence |\n| 0.60–0.90 | Uncertain |\n| < 0.60 | Low confidence |\n\nThis design prioritizes safety in medical contexts.\n\n---\n\n### 🧠 Explainability Policy\n\nThis project does NOT implement pixel-level explainability (e.g., Grad-CAM).\n\nInstead, it provides:\n\n* Conceptual transparency\n* Architectural disclosure\n* Clear limitations\n\n---\n\n### πŸ“Œ Medical Disclaimer\n\n> This system is intended for research and decision-support only.\n> It is NOT a clinical diagnostic tool.\n\n---\n\n### 🏁 Result\n\nThe API now follows responsible AI practices suitable for:\n\n* Medical AI prototyping\n* Research environments\n* Academic demonstration\n* Ethical ML deployment\n\n\n\n### 🎯 Project Status\n\n#### We have Now:\n\nβœ… Production API \n\nβœ… Logging \n\nβœ… Lifecycle\n\nβœ… Safety logic\n\nβœ… Medical best practices\n\nβœ… Honest explainability","metadata":{}},{"cell_type":"markdown","source":"# πŸ’Ž Final API Integration (Production-Ready)","metadata":{}},{"cell_type":"markdown","source":"\n## πŸ“ Final API Integration (Production-Ready)\n\nIn this phase, we unified all previous development stages into a **single production-grade FastAPI application**.\n\nThis version is designed to be:\n\n- βœ… Deployment-ready\n- βœ… Scalable\n- βœ… Cleanly structured\n- βœ… Recruiter-friendly\n\n---\n\n\n### 🧠 Key Improvements\n\n#### 1. Unified Architecture\nAll components are merged into a single `main.py`:\n\n- Model loading\n- GPU configuration\n- Preprocessing\n- Inference\n- API endpoints\n\n---\n\n#### 2. Lifecycle Management\n\nWe use FastAPI `lifespan`:\n\n- Load model once\n- Warm-up inference\n- Store in `app.state`\n\n```python\n@asynccontextmanager\nasync def lifespan(app: FastAPI):\n model = tf.keras.models.load_model(...)\n app.state.model = model\n yield\n````\n\n---\n\n#### 3. Production-Grade Inference\n\n* TensorFlow graph execution (`@tf.function`)\n* Warm-up for latency reduction\n* Confidence + threshold logic\n\n---\n\n#### 4. Robust Input Handling\n\n* File type validation\n* File size control\n* Safe image preprocessing\n\n---\n\n#### 5. Middleware & Monitoring\n\n\n* Request timing header\n* Structured logging\n\n---\n\n#### 6. Global Error Handling\n\n```python\n@app.exception_handler(Exception)\n```\n\nPrevents crashes and returns clean responses.\n\n---\n\n#### 7. Medical AI Safety Layer\n\nWe added conservative interpretation:\n\n| Confidence | Interpretation |\n| ---------- | --------------- |\n| β‰₯ 0.9 | High confidence |\n| 0.6–0.9 | Uncertain |\n| < 0.6 | Low confidence |\n\n---\n\n### πŸ”Œ API Endpoints\n\n#### βœ… `/health`\n\nBasic health check\n\n#### βœ… `/predict`\n\nSingle image inference\n\n#### βœ… `/predict-batch`\n\nBatch inference support\n\n#### βœ… `/model-info`\n\nModel metadata + disclaimer\n\n---\n\n### βš™οΈ Performance Considerations\n\n\n* GPU memory growth enabled\n* Model warm-up reduces first inference delay\n* No redundant model loading\n\n---\n\n### πŸ“¦ Why This Matters\n\nThis phase transforms the project from:\n\nπŸ‘‰ **Notebook-level prototype**\n\nto:\n\nπŸ‘‰ **Real-world deployable AI service**\n\n---","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Build πŸ“Œ \"config.py\"","metadata":{}},{"cell_type":"code","source":"config_code = \"\"\"\n\nimport os\n\nMODEL_PATH = os.getenv(\n \"MODEL_PATH\",\n \"models/VGGNet.keras\"\n)\n\nIMG_SIZE = (96, 96)\n\nTHRESHOLD = 0.5\n\nMAX_FILE_SIZE_MB = 5\n\nALLOWED_EXTENSIONS = {\n \".jpg\",\n \".jpeg\",\n \".png\"\n}\n\n\nAPI_TITLE = \"Histopathology Cancer Classification API\"\n\nAPI_DESCRIPTION =\"Production-ready Medical AI Inference Service\"\n\nAPI_VERSION =\"1.0.0\"\n\nAPI_CONTACT = {\n \"name\": \"Sepehr\",\n \"url\": \"https://www.linkedin.com/in/sepehr-khalilib\",\n \"email\": \"Sepehr.khalilib@gmail.com\"\n}\n\nAPI_LICENSE = {\n \"name\": \"MIT License\"\n} \n\nMODEL_METADATA = {\n \"name\": \"VGGNet\",\n \"version\": \"1.0.0\",\n \"framework\": \"TensorFlow\",\n \"task\": \"Histopathologic Cancer Classification\"\n}\n\nOPENAPI_TAGS = [\n {\n \"name\": \"Health\",\n \"description\": \"Health monitoring endpoints.\",\n },\n {\n \"name\": \"Inference\",\n \"description\": \"Cancer prediction endpoints.\",\n },\n {\n \"name\": \"Model\",\n \"description\": \"Model information and metadata.\",\n },\n]\n\n\"\"\"\n\nwith open(\"app/config.py\", \"w\") as f:\n f.write(config_code)\n\nprint(\"βœ… config.py created\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.630640Z","iopub.execute_input":"2026-07-02T07:13:02.630906Z","iopub.status.idle":"2026-07-02T07:13:02.654801Z","shell.execute_reply.started":"2026-07-02T07:13:02.630880Z","shell.execute_reply":"2026-07-02T07:13:02.653765Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸ“ Build πŸš€ **Final** app/main.py","metadata":{}},{"cell_type":"code","source":"# ============================================\n# Build Final Production main.py\n# ============================================\n\nmain_code = \"\"\"\n\n# ============================================================\n# Histopathologic Cancer Classification API\n# ============================================================\n\n# imports\n# logging\n# gpu setup\n# preprocessing\n# model loading (lifespan)\n# middleware\n# exception handling\n# helper functions\n# endpoints:\n# - /health\n# - /predict\n# - /predict-batch\n# - /model-info\n\n\n# =============================\n# Imports\n# =============================\n# Standard Library\nimport io\nimport logging\nimport os\nimport time\nfrom contextlib import asynccontextmanager\nfrom typing import List\n\n# Third-party\nimport numpy as np\nimport tensorflow as tf\nfrom fastapi import FastAPI, File, HTTPException, Request, UploadFile\nfrom fastapi.responses import JSONResponse\nfrom PIL import Image\nfrom pydantic import BaseModel\n\n# Local\nfrom app.config import (\n MODEL_PATH,\n IMG_SIZE,\n THRESHOLD,\n MAX_FILE_SIZE_MB,\n ALLOWED_EXTENSIONS,\n MODEL_METADATA,\n API_TITLE,\n API_DESCRIPTION,\n API_VERSION,\n API_CONTACT,\n API_LICENSE,\n OPENAPI_TAGS,\n)\n\n# ============================================================\n# Logging Configuration\n# ============================================================\n\nlogging.basicConfig(\n level=logging.INFO,\n format=\"%(asctime)s | %(levelname)s | %(message)s\"\n)\n\nlogger = logging.getLogger(\"histopathology-api\")\n\n\n# ============================================================\n# GPU Configuration\n# ============================================================\n\ndef configure_gpu() -> None:\n gpus = tf.config.list_physical_devices('GPU')\n\n if gpus:\n try:\n for gpu in gpus:\n tf.config.experimental.set_memory_growth(gpu, True)\n logger.info(f\"βœ… {len(gpus)} GPU(s) configured\")\n except RuntimeError as e:\n logger.error(f\"GPU error: {e}\")\n else:\n logger.warning(\"⚠ Running on CPU\")\n\n# ============================================================\n# Model Warm-up\n# ============================================================\n\ndef warmup_model(model: tf.keras.Model) -> None:\n dummy = np.zeros((1, *IMG_SIZE, 3), dtype=np.float32)\n _ = model.predict(dummy, verbose=0)\n logger.info(\"πŸ”₯ Model warmed up\")\n\n# ============================================================\n# Preprocessing\n# ============================================================\n\ndef validate_file(file: UploadFile) -> None:\n \n if not file.filename:\n raise HTTPException(\n status_code=400,\n detail=\"Filename is missing.\"\n )\n \n ext = os.path.splitext(file.filename)[1].lower()\n\n if ext not in ALLOWED_EXTENSIONS:\n raise HTTPException(400, \"Invalid file type\")\n\n file.file.seek(0, os.SEEK_END)\n size = file.file.tell()\n file.file.seek(0)\n\n if size / (1024 * 1024) > MAX_FILE_SIZE_MB:\n raise HTTPException(400, \"File too large\")\n\n\ndef preprocess_image(file: UploadFile) -> np.ndarray:\n\n contents = file.file.read()\n\n try:\n image = Image.open(io.BytesIO(contents)).convert(\"RGB\")\n except Exception:\n raise HTTPException(\n status_code=400,\n detail=\"Invalid image.\"\n )\n\n image = image.resize(IMG_SIZE)\n\n arr = np.array(image, dtype=np.float32) / 255.0\n arr = np.expand_dims(arr, axis=0)\n\n return arr\n \n\n# ============================================================\n# Inference\n# ============================================================\n\n@tf.function\ndef model_inference(model: tf.keras.Model, x: tf.Tensor) -> tf.Tensor:\n return model(x, training=False)\n\n# ============================================================\n# Confidence Interpretation (Medical Safe)\n# ============================================================\n\ndef interpret_confidence(conf: float) -> str:\n if conf >= 0.9:\n return \"high_confidence\"\n elif conf >= 0.6:\n return \"uncertain\"\n return \"low_confidence\"\n\n# ============================================================\n# Lifespan (Startup / Shutdown)\n# ============================================================\n\n@asynccontextmanager\nasync def lifespan(app: FastAPI):\n\n logger.info(\"πŸš€ Starting API...\")\n\n configure_gpu()\n\n ## model_path = os.getenv(\"MODEL_PATH\", \"models/VGGNet.keras\")\n model_path = MODEL_PATH\n \n model = tf.keras.models.load_model(model_path)\n model.trainable = False\n\n warmup_model(model)\n\n app.state.model = model\n\n yield\n\n logger.info(\"πŸ›‘ Shutdown API\")\n\n# ============================================================\n# FastAPI App\n# ============================================================\n\napp = FastAPI(\n title=API_TITLE,\n description=API_DESCRIPTION,\n version=API_VERSION,\n contact=API_CONTACT,\n license_info=API_LICENSE,\n openapi_tags=OPENAPI_TAGS,\n lifespan=lifespan,\n)\n\n# ============================================================\n# Middleware (Timing)\n# ============================================================\n\n@app.middleware(\"http\")\nasync def add_process_time(request: Request, call_next):\n\n start = time.time()\n response = await call_next(request)\n duration = time.time() - start\n\n response.headers[\"X-Process-Time\"] = str(round(duration, 4))\n\n return response\n\n# ============================================================\n# Global Exception Handler\n# ============================================================\n\n@app.exception_handler(Exception)\nasync def global_exception_handler(request: Request, exc: Exception):\n\n if isinstance(exc, HTTPException):\n raise exc\n\n logger.exception(\"Unhandled exception\")\n\n return JSONResponse(\n status_code=500,\n content={\n \"success\": False,\n \"message\": \"Internal Server Error\"\n }\n )\n\n# ============================================================\n# Response Models\n# ============================================================\n\nclass PredictionResponse(BaseModel):\n filename: str\n probability: float\n prediction: str\n confidence: float\n confidence_level: str\n inference_time_ms: float\n\nclass BatchPrediction(BaseModel):\n filename: str\n probability: float\n prediction: str\nclass BatchPredictionResponse(BaseModel):\n results: List[BatchPrediction]\n\n\nclass ErrorResponse(BaseModel):\n success: bool\n message: str\n\n\n# ============================================================\n# Health Check\n# ============================================================\n\n@app.get(\n \"/health\",\n tags=[\"Health\"],\n summary=\"Health check\",\n description=\"Returns API health status.\",\n)\ndef health():\n return {\"status\": \"ok\"}\n\n\n@app.get(\n \"/health/live\",\n tags=[\"Health\"],\n summary=\"Liveness probe\",\n description=\"Checks whether the API process is alive.\",\n)\ndef liveness_probe():\n return {\"status\": \"alive\"}\n\n\n@app.get(\n \"/health/ready\",\n tags=[\"Health\"],\n summary=\"Readiness probe\",\n description=\"Checks whether the model is loaded and ready.\",\n)\ndef readiness_probe():\n if not hasattr(app.state, \"model\"):\n return JSONResponse(\n status_code=503,\n content={\"status\": \"model not ready\"}\n )\n return {\"status\": \"ready\"}\n\n# ============================================================\n# Predict Endpoint\n# ============================================================\n\n@app.post(\n \"/predict\",\n tags=[\"Inference\"],\n summary=\"Predict cancer from a histopathology image\",\n description=(\n \"Upload a histopathology image to classify it as \"\n \"Cancer or Normal using the trained CNN model.\"\n ),\n response_model=PredictionResponse,\n responses={\n 400: {\n \"model\": ErrorResponse,\n \"description\": \"Invalid image or unsupported file\"\n },\n 500: {\n \"model\": ErrorResponse,\n \"description\": \"Internal inference error\"\n }\n }\n)\n\nasync def predict(request: Request, file: UploadFile = File(...)):\n\n start = time.time()\n\n validate_file(file)\n\n try:\n x = preprocess_image(file)\n\n model = request.app.state.model\n\n preds = model_inference(model, x)\n prob = float(preds[0][0])\n\n prediction = \"Cancer\" if prob >= THRESHOLD else \"Normal\"\n confidence = prob if prob >= THRESHOLD else 1 - prob\n\n level = interpret_confidence(confidence)\n\n duration = (time.time() - start) * 1000\n\n return PredictionResponse(\n filename=file.filename,\n probability=round(prob, 6),\n prediction=prediction,\n confidence=round(confidence, 6),\n confidence_level=level,\n inference_time_ms=round(duration, 2)\n )\n\n\n except HTTPException:\n raise\n\n except Exception:\n logger.exception(\"Prediction failed\")\n \n raise HTTPException(\n status_code=500,\n detail=\"Inference failed\"\n )\n\n# ============================================================\n# Batch Prediction\n# ============================================================\n\n@app.post(\n \"/predict-batch\",\n tags=[\"Inference\"],\n summary=\"Batch prediction\",\n description=\"Predict multiple histopathology images.\",\n response_model=BatchPredictionResponse,\n responses={\n 400: {\n \"model\": ErrorResponse,\n \"description\": \"Invalid image or unsupported file\"\n },\n 500: {\n \"model\": ErrorResponse,\n \"description\": \"Internal inference error\"\n }\n }\n)\n\nasync def predict_batch(request: Request, files: List[UploadFile] = File(...)):\n\n model = request.app.state.model\n results = []\n\n for file in files:\n validate_file(file)\n x = preprocess_image(file)\n\n preds = model_inference(model, x)\n prob = float(preds[0][0])\n\n prediction = \"Cancer\" if prob >= THRESHOLD else \"Normal\"\n\n results.append({\n \"filename\": file.filename,\n \"probability\": prob,\n \"prediction\": prediction\n })\n\n return {\"results\": results}\n\n# ============================================================\n# Model Info\n# ============================================================\n\n@app.get(\n \"/model-info\",\n tags=[\"Model\"],\n summary=\"Model metadata\",\n description=\"Returns information about the deployed model.\"\n)\ndef model_info():\n return {\n **MODEL_METADATA,\n \"input\": \"96x96 RGB\",\n \"explainability\": (\n \"This model does not provide pixel-level explanations. \"\n \"Predictions are based on learned visual patterns.\"\n ),\n \"note\": \"Research use only. Not for clinical diagnosis.\"\n }\n\n\"\"\"\n\nwith open(\"app/main.py\", \"w\") as f:\n f.write(main_code)\n\nprint(\"βœ… Final production main.py generated successfully\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.656420Z","iopub.execute_input":"2026-07-02T07:13:02.656724Z","iopub.status.idle":"2026-07-02T07:13:02.678436Z","shell.execute_reply.started":"2026-07-02T07:13:02.656696Z","shell.execute_reply":"2026-07-02T07:13:02.677112Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"with open(\"app/main.py\", \"r\") as f:\n print(f.read())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.679611Z","iopub.execute_input":"2026-07-02T07:13:02.679880Z","iopub.status.idle":"2026-07-02T07:13:02.704484Z","shell.execute_reply.started":"2026-07-02T07:13:02.679852Z","shell.execute_reply":"2026-07-02T07:13:02.703354Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸ“ Build πŸ§ͺ \"requirements.txt\"","metadata":{}},{"cell_type":"code","source":"requirements = \"\"\"\nfastapi==0.115.12\nuvicorn==0.34.3\nnumpy==1.26.4\npydantic==2.11.7\npython-multipart==0.0.20\nPillow==11.2.1\n\"\"\"\n\nwith open(\"requirements.txt\", \"w\") as f:\n f.write(requirements.strip())\n\nprint(\"βœ… requirements.txt created\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.705885Z","iopub.execute_input":"2026-07-02T07:13:02.706380Z","iopub.status.idle":"2026-07-02T07:13:02.723624Z","shell.execute_reply.started":"2026-07-02T07:13:02.706340Z","shell.execute_reply":"2026-07-02T07:13:02.722716Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## πŸ“ Build πŸ“‹ \"README.md\"","metadata":{}},{"cell_type":"code","source":"readme = \"\"\"\n# Histopathologic Cancer Classification API\n\n> Production-ready Medical AI inference service built with TensorFlow, FastAPI and Docker.\n\n![Python](https://img.shields.io/badge/Python-3.11-blue)\n![TensorFlow](https://img.shields.io/badge/TensorFlow-2.16-orange)\n![FastAPI](https://img.shields.io/badge/FastAPI-Production-green)\n![Docker](https://img.shields.io/badge/Docker-Ready-blue)\n![License](https://img.shields.io/badge/License-MIT-lightgrey)\n\n---\n\n## Overview\n\nThis project deploys a deep learning model for **histopathologic cancer classification** as a production-ready REST API.\n\nThe service provides:\n\n- Single image prediction\n- Batch prediction\n- Health monitoring endpoints\n- Docker support (CPU & GPU)\n- Interactive Swagger documentation\n- Production-oriented project structure\n\n> **Research Use Only β€” Not intended for clinical diagnosis.**\n\n---\n\n## Project Structure\n\n```text\nhistopath-api/\nβ”‚\nβ”œβ”€β”€ app/\nβ”‚ β”œβ”€β”€ main.py\nβ”‚ └── config.py\nβ”‚\nβ”œβ”€β”€ models/\nβ”‚ └── VGGNet.keras\nβ”‚\nβ”œβ”€β”€ client/\nβ”‚ └── client.py\nβ”‚\nβ”œβ”€β”€ docs/\nβ”‚ β”œβ”€β”€ architecture.png\nβ”‚ β”œβ”€β”€ api_examples.md\nβ”‚ β”œβ”€β”€ deployment.md\nβ”‚ └── screenshots/\nβ”‚\nβ”œβ”€β”€ tests/\nβ”‚\nβ”œβ”€β”€ Dockerfile.cpu\nβ”œβ”€β”€ Dockerfile.gpu\nβ”œβ”€β”€ docker-compose.yml\nβ”‚\nβ”œβ”€β”€ requirements.txt\nβ”œβ”€β”€ requirements-dev.txt\nβ”‚\nβ”œβ”€β”€ .dockerignore\nβ”œβ”€β”€ .gitignore\nβ”œβ”€β”€ .env.example\nβ”‚\nβ”œβ”€β”€ LICENSE\nβ”œβ”€β”€ CHANGELOG.md\nβ”œβ”€β”€ README.md\n```\n\n---\n\n## Tech Stack\n\n- Python\n- TensorFlow / Keras\n- FastAPI\n- Docker\n- Uvicorn\n- PyTest\n\n---\n\n## API Endpoints\n\n| Method | Endpoint | Description |\n|---------|----------|-------------|\n| GET | `/health` | Health check |\n| GET | `/health/live` | Liveness probe |\n| GET | `/health/ready` | Readiness probe |\n| POST | `/predict` | Single image prediction |\n| POST | `/predict-batch` | Batch prediction |\n| GET | `/model-info` | Model information |\n\nInteractive documentation:\n\n```\nhttp://localhost:8000/docs\n```\n\n---\n\n## Run with Docker\n\nCPU\n\n```bash\ndocker build -f Dockerfile.cpu -t histopath-api .\ndocker run -p 8000:8000 histopath-api\n```\n\nGPU\n\n```bash\ndocker build -f Dockerfile.gpu -t histopath-api-gpu .\ndocker run --gpus all -p 8000:8000 histopath-api-gpu\n```\n\n---\n\n## Local Installation\n\n```bash\npip install -r requirements.txt\n\nuvicorn app.main:app --reload\n```\n\n---\n\n## Example Response\n\n```json\n{\n \"filename\": \"sample.png\",\n \"probability\": 0.982531,\n \"prediction\": \"Cancer\",\n \"confidence\": 0.982531,\n \"confidence_level\": \"high_confidence\",\n \"inference_time_ms\": 18.42\n}\n```\n\n---\n\n## Features\n\n- Production-oriented FastAPI application\n- TensorFlow model warm-up\n- GPU memory growth configuration\n- Request validation\n- Structured response models\n- Health monitoring endpoints\n- Batch inference\n- Dockerized deployment\n- API testing with PyTest\n\n---\n\n## Future Improvements\n\n- Authentication\n- Rate limiting\n- CI/CD\n- Model versioning\n- Hugging Face deployment\n- Kubernetes deployment\n\n---\n\n## License\n\nReleased under the MIT License.\n\n---\n\n## Author\n\n**Sepehr Khalili**\n\nAI Engineer | Deep Learning | Computer Vision | Medical AI\n\nLinkedIn:\n\nhttps://www.linkedin.com/in/sepehr-khalilib/\n\n---\n\n## Disclaimer\n\nThis software is provided exclusively for research and educational purposes.\n\nIt must not be used for medical diagnosis or clinical decision-making.\n\"\"\"\n\nwith open(\"README.md\", \"w\") as f:\n f.write(readme.strip())\n\nprint(\"βœ… README.md created\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.726008Z","iopub.execute_input":"2026-07-02T07:13:02.726558Z","iopub.status.idle":"2026-07-02T07:13:02.744031Z","shell.execute_reply.started":"2026-07-02T07:13:02.726523Z","shell.execute_reply":"2026-07-02T07:13:02.743016Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## 🧱 Final Repository Structure (on Github)\n\n","metadata":{}},{"cell_type":"markdown","source":"```\nhistopath-api/\nβ”‚\nβ”œβ”€β”€ app/\nβ”‚ β”œβ”€β”€ main.py\nβ”‚ └── config.py\nβ”‚\nβ”œβ”€β”€ models/\nβ”‚ └── VGGNet.keras\nβ”‚\nβ”œβ”€β”€ client/\nβ”‚ └── client.py\nβ”‚\nβ”œβ”€β”€ docs/\nβ”‚ β”œβ”€β”€ architecture.png\nβ”‚ β”œβ”€β”€ api_examples.md\nβ”‚ β”œβ”€β”€ deployment.md\nβ”‚ └── screenshots/\nβ”‚\nβ”œβ”€β”€ tests/\nβ”‚\nβ”œβ”€β”€ Dockerfile.cpu\nβ”œβ”€β”€ Dockerfile.gpu\nβ”œβ”€β”€ docker-compose.yml\nβ”‚\nβ”œβ”€β”€ requirements.txt\nβ”œβ”€β”€ requirements-dev.txt\nβ”‚\nβ”œβ”€β”€ .dockerignore\nβ”œβ”€β”€ .gitignore\nβ”œβ”€β”€ .env.example\nβ”‚\nβ”œβ”€β”€ LICENSE\nβ”œβ”€β”€ CHANGELOG.md\nβ”œβ”€β”€ README.md\n```","metadata":{}},{"cell_type":"markdown","source":"# πŸ‘¨β€πŸ’» Phase 10 β€” Dockerization & Deployment","metadata":{}},{"cell_type":"markdown","source":"## πŸ“ Phase 10 β€” Dockerization & Deployment\n\nIn this phase, we transform our trained deep learning model into a **production-ready AI service** using FastAPI and Docker.\n\nThis enables:\n\n* Reproducible environments\n* Easy deployment across platforms\n* Real-world usability\n\n---\n\n### 🧱 Project Structure\n```\nhistopathology-api/\nβ”‚\nβ”œβ”€β”€ app/\nβ”‚ β”œβ”€β”€ main.py\nβ”‚ β”œβ”€β”€ config.py\nβ”‚\nβ”œβ”€β”€ models/\nβ”‚ └── VGGNet.keras\nβ”‚\nβ”œβ”€β”€ client/\nβ”‚ └── client.py\nβ”‚\nβ”œβ”€β”€ Dockerfile.cpu\nβ”œβ”€β”€ Dockerfile.gpu\nβ”œβ”€β”€ requirements.txt\nβ”œβ”€β”€ .dockerignore\nβ”œβ”€β”€ README.md\n```\n\n---\n\n### 🐳 Dockerization\n\n#### Why Docker?\n\n* Eliminates environment issues\n* Enables portability\n* Simplifies deployment\n\n---\n\n#### πŸ”Ή Build & Run (CPU)\n\n```bash\ndocker build -f Dockerfile.cpu -t histopath-api .\ndocker run -p 8000:8000 histopath-api\n```\n\n---\n\n#### πŸ”Ή Build & Run (GPU)\n\n```bash\ndocker build -f Dockerfile.gpu -t histopath-api-gpu .\ndocker run --gpus all -p 8000:8000 histopath-api-gpu\n```\n\n---\n\n### πŸ§ͺ API Testing\n\n#### Swagger UI\n\n```\nhttp://localhost:8000/docs\n```\n\n* Upload image\n* Get prediction\n* Interactive testing\n\n---\n\n#### cURL\n\n```bash\ncurl -X POST \"http://127.0.0.1:8000/predict\" \\\n -F \"file=@image.png\"\n```\n\n---\n\n### 🌍 Deployment Options\n\n| Platform | Type | Use Case |\n| ---------------- | ------- | ---------------- |\n| Render | CPU | Demo / MVP |\n| HuggingFace | CPU/GPU | Interactive demo |\n| AWS / GCP | GPU | Production |\n| Dedicated Server | GPU | SaaS |\n\n---\n\n### 🏁 Conclusion\n\nThis phase elevates the project from:\n\n> Research Prototype β†’ Production-Ready AI System\n\nIt demonstrates:\n\n* Engineering maturity\n* Deployment skills\n* Real-world applicability\n\n---\n\n### 🌠 Future Steps\n\n* SaaS productization\n* Authentication system\n* Monitoring & scaling\n* Frontend integration\n","metadata":{}},{"cell_type":"markdown","source":"## 🐳 Dockerfile (CPU Version β€” Professional)","metadata":{}},{"cell_type":"code","source":"\"\"\"\nLABEL maintainer=\"Sepehr | linkedin.com/in/sepehr-khalilib\"\nLABEL project=\"Histopathologic Cancer Classification API\"\nLABEL version=\"1.0.0\"\nLABEL description=\"Production-ready FastAPI for Histopathologic Cancer Classification\"\n\nFROM tensorflow/tensorflow:2.16.1\n\nWORKDIR /app\n\nENV PYTHONUNBUFFERED=1\n\nCOPY requirements.txt .\n\nRUN pip install --no-cache-dir -r requirements.txt\n\nCOPY app/ app/\nCOPY models/ models/\n\nHEALTHCHECK --interval=30s --timeout=10s --start-period=20s --retries=3 \\\nCMD python -c \"import urllib.request; urllib.request.urlopen('http://localhost:8000/health')\"\n\nEXPOSE 8000\n\nCMD [\"uvicorn\", \"app.main:app\", \"--host\", \"0.0.0.0\", \"--port\", \"8000\"]\n\"\"\"","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.745541Z","iopub.execute_input":"2026-07-02T07:13:02.745996Z","iopub.status.idle":"2026-07-02T07:13:02.768317Z","shell.execute_reply.started":"2026-07-02T07:13:02.745966Z","shell.execute_reply":"2026-07-02T07:13:02.767448Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"dockerfile_cpu = \"\"\"\n\n# =========================\n# Base Image\n# =========================\nFROM tensorflow/tensorflow:2.16.1\n\n# =========================\n# LABEL\n# =========================\nLABEL maintainer=\"Sepehr | linkedin.com/in/sepehr-khalilib\"\nLABEL project=\"Histopathologic Cancer Classification API\"\nLABEL version=\"1.0.0\"\nLABEL description=\"Production-ready FastAPI for Histopathologic Cancer Classification\"\n\n# =========================\n# Set working directory\n# =========================\nWORKDIR /app\n\nENV PYTHONUNBUFFERED=1\nENV PYTHONDONTWRITEBYTECODE=1\n\n# =========================\n# Copy requirements\n# =========================\nCOPY requirements.txt .\n\n# =========================\n# Install Python dependencies\n# =========================\nRUN pip install --no-cache-dir -r requirements.txt\n\n# =========================\n# Copy project files\n# =========================\nCOPY app/ app/\nCOPY models/ models/\n\n# =========================\n# HEALTHCHECK\n# =========================\nHEALTHCHECK --interval=30s --timeout=10s --start-period=20s --retries=3 \\\nCMD python -c \"import urllib.request,sys; urllib.request.urlopen('http://localhost:8000/health'); sys.exit(0)\"\n# =========================\n# Expose port\n# =========================\nEXPOSE 8000\n\n# =========================\n# Run FastAPI\n# =========================\nCMD [\"uvicorn\", \"app.main:app\", \"--host\", \"0.0.0.0\", \"--port\", \"8000\"]\n\n\"\"\"\n\nwith open(\"Dockerfile.cpu\", \"w\") as f:\n f.write(dockerfile_cpu)\n\nprint(\"βœ… Dockerfile.cpu created\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.769524Z","iopub.execute_input":"2026-07-02T07:13:02.769894Z","iopub.status.idle":"2026-07-02T07:13:02.792615Z","shell.execute_reply.started":"2026-07-02T07:13:02.769854Z","shell.execute_reply":"2026-07-02T07:13:02.791483Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"\n## ⚑ Dockerfile (GPU Version β€” Professional)","metadata":{}},{"cell_type":"code","source":"dockerfile_gpu = \"\"\"\n\n# =========================\n# NVIDIA CUDA Base Image\n# =========================\nFROM tensorflow/tensorflow:2.16.1-gpu\n\n# =========================\n# LABEL\n# =========================\nLABEL maintainer=\"Sepehr | linkedin.com/in/sepehr-khalilib\"\nLABEL project=\"Histopathologic Cancer Classification API\"\nLABEL version=\"1.0.0\"\nLABEL description=\"Production-ready FastAPI for Histopathologic Cancer Classification\"\n\n# =========================\n# Set working directory\n# =========================\n\nWORKDIR /app\n\nENV PYTHONUNBUFFERED=1\nENV PYTHONDONTWRITEBYTECODE=1\n\n# =========================\n# Copy requirements\n# =========================\n\nCOPY requirements.txt .\n\n# =========================\n# Install Python\n# =========================\n\nRUN pip install --no-cache-dir -r requirements.txt\n\n# =========================\n# Copy project\n# =========================\n\nCOPY app/ app/\nCOPY models/ models/\n\n# =========================\n# HEALTHCHECK\n# =========================\nHEALTHCHECK --interval=30s --timeout=10s --start-period=20s --retries=3 \\\nCMD python -c \"import urllib.request,sys; urllib.request.urlopen('http://localhost:8000/health'); sys.exit(0)\"\n\n# =========================\n# Expose port\n# =========================\nEXPOSE 8000\n\n# =========================\n# Run FastAPI\n# =========================\nCMD [\"uvicorn\", \"app.main:app\", \"--host\", \"0.0.0.0\", \"--port\", \"8000\"]\n\n\n\n\"\"\"\n\nwith open(\"Dockerfile.gpu\", \"w\") as f:\n f.write(dockerfile_gpu)\n\nprint(\"βœ… Dockerfile.gpu created\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.793815Z","iopub.execute_input":"2026-07-02T07:13:02.794257Z","iopub.status.idle":"2026-07-02T07:13:02.815765Z","shell.execute_reply.started":"2026-07-02T07:13:02.794224Z","shell.execute_reply":"2026-07-02T07:13:02.814430Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## 🚫 .dockerignore","metadata":{}},{"cell_type":"code","source":"dockerignore = \"\"\"\n# Git\n.git\n.gitignore\n\n# Environment\n.env\n.env.*\n\n# Python\n__pycache__/\n*.py[cod]\n*.pyo\n*.pyd\n\n# Virtual Environment\n.venv/\nvenv/\nenv/\n\n# IDE\n.vscode/\n.idea/\n\n# Jupyter\n.ipynb_checkpoints/\n\n# OS\n.DS_Store\nThumbs.db\n\n# Documentation (not needed in image)\ndocs/\n\n# Tests (not needed in image)\ntests/\n\"\"\"\n\nwith open(\".dockerignore\", \"w\") as f:\n f.write(dockerignore)\n\nprint(\"βœ… .dockerignore created\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.817003Z","iopub.execute_input":"2026-07-02T07:13:02.817536Z","iopub.status.idle":"2026-07-02T07:13:02.840207Z","shell.execute_reply.started":"2026-07-02T07:13:02.817484Z","shell.execute_reply":"2026-07-02T07:13:02.839098Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## ▢️ Build & Run\n\n### CPU:","metadata":{}},{"cell_type":"code","source":"# 1️⃣ First:\n'''\ndocker build -f Dockerfile.cpu -t histopath-api .\n'''\n\n# 2️⃣ In the continue:\n'''\ndocker run -p 8000:8000 histopath-api\n'''","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.841588Z","iopub.execute_input":"2026-07-02T07:13:02.842162Z","iopub.status.idle":"2026-07-02T07:13:02.859309Z","shell.execute_reply.started":"2026-07-02T07:13:02.842094Z","shell.execute_reply":"2026-07-02T07:13:02.858118Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### GPU","metadata":{}},{"cell_type":"code","source":"# 1️⃣ First:\n'''\ndocker build -f Dockerfile.gpu -t histopath-api-gpu .\n'''\n\n# 2️⃣ In the continue:\n'''\ndocker run --gpus all -p 8000:8000 histopath-api-gpu\n'''","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.860585Z","iopub.execute_input":"2026-07-02T07:13:02.861004Z","iopub.status.idle":"2026-07-02T07:13:02.882408Z","shell.execute_reply.started":"2026-07-02T07:13:02.860963Z","shell.execute_reply":"2026-07-02T07:13:02.881094Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# 🌟 Client Guide😊","metadata":{}},{"cell_type":"markdown","source":"## πŸ‘¨β€πŸ’» API Testing Guide\n\n### Swagger UI\n\nAfter starting the FastAPI server:\n\n```bash\nuvicorn app.main:app --host 0.0.0.0 --port 8000\n```\n\nOpen:\n\n```text\nhttp://localhost:8000/docs\n```\n\nSwagger UI is automatically generated by FastAPI.\n\n#### Available Features\n\n* Interactive API testing\n* Request validation\n* Response inspection\n* Endpoint documentation\n\n#### Test Prediction Endpoint\n\n1. Open `/predict`\n2. Click **Try it out**\n3. Upload an image\n4. Click **Execute**\n5. Review the JSON response\n\n---\n\n### cURL Example\n\n```bash\ncurl -X POST \"http://127.0.0.1:8000/predict\" \\\n-H \"accept: application/json\" \\\n-F \"file=@image.png\"\n```\n\n### Example Response\n\n```json\n{\n \"filename\": \"Histology-Cancer.png\",\n \"probability\": 0.004296,\n \"prediction\": \"Normal\",\n \"confidence\": 0.995704,\n \"confidence_level\": \"high_confidence\",\n \"inference_time_ms\": 324.96\n}\n```\n\n---\n\n### Why Swagger and cURL?\n\nSwagger is ideal for interactive testing and demonstrations.\n\ncURL is ideal for scripting, automation, CI/CD pipelines, and production verification.\n\n```\n```\n","metadata":{}},{"cell_type":"markdown","source":"## 🏹 Client Setup Guide\n\n### Step 1 β€” Download the Project\n\n#### Option A β€” GitHub\n\n```bash\ngit clone \ncd project\n```\n\n#### Option B β€” Gumroad / Upwork Delivery\n\nDownload the provided ZIP package and extract it.\n\n---\n\n### Step 2 β€” Install Docker\n\nDownload Docker Desktop:\n\nhttps://www.docker.com/products/docker-desktop/\n\nVerify installation:\n\n```bash\ndocker --version\n```\n\n---\n\n### Step 3 β€” Build Docker Image\n\nCPU Version:\n\n```bash\ndocker build -f Dockerfile.cpu -t histopath-api .\n```\n\nGPU Version:\n\n```bash\ndocker build -f Dockerfile.gpu -t histopath-api-gpu .\n```\n\n---\n\n### Step 4 β€” Run the Container\n\nCPU:\n\n```bash\ndocker run -p 8000:8000 histopath-api\n```\n\nGPU:\n\n```bash\ndocker run --gpus all -p 8000:8000 histopath-api-gpu\n```\n\n---\n\n### Step 5 β€” Open Swagger UI\n\nOpen:\n\n```text\nhttp://localhost:8000/docs\n```\n\n---\n\n### Step 6 β€” Upload an Image\n\n1. Select `/predict`\n2. Click **Try it out**\n3. Upload a histopathology image\n4. Execute the request\n\n---\n\n### Step 7 β€” Review Results\n\nExample:\n\n```json\n{\n \"prediction\": \"Cancer\",\n \"confidence\": 0.92,\n \"confidence_level\": \"high\"\n}\n```\n\n---\n\n### Troubleshooting\n\n#### Docker not found\n\nInstall Docker Desktop.\n\n#### GPU unavailable\n\nInstall:\n\n* NVIDIA Driver\n* NVIDIA Container Toolkit\n\nThen retry the GPU container.\n\n#### Port already in use\n\nChange port mapping:\n\n```bash\ndocker run -p 8080:8000 histopath-api\n```\n\nThen access:\n\n```text\nhttp://localhost:8080/docs\n```\n","metadata":{}},{"cell_type":"markdown","source":"## πŸ—ƒ Download Kaggle Files","metadata":{}},{"cell_type":"code","source":"!zip -r histopath-api.zip app models requirements.txt Dockerfile.cpu Dockerfile.gpu README.md","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:02.883748Z","iopub.execute_input":"2026-07-02T07:13:02.884308Z","iopub.status.idle":"2026-07-02T07:13:15.620627Z","shell.execute_reply.started":"2026-07-02T07:13:02.884258Z","shell.execute_reply":"2026-07-02T07:13:15.619606Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from IPython.display import FileLink\n\nFileLink(\"histopath-api.zip\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-07-02T07:13:15.622460Z","iopub.execute_input":"2026-07-02T07:13:15.623548Z","iopub.status.idle":"2026-07-02T07:13:15.630336Z","shell.execute_reply.started":"2026-07-02T07:13:15.623508Z","shell.execute_reply":"2026-07-02T07:13:15.629239Z"}},"outputs":[],"execution_count":null}]}