While reviewing this OpenVINO notebook, I noticed it’s the only one using OpenVINO for inference in BirdCLEF 2025. Since there's no discussion around it yet, here’s a quick guide on how to use OpenVINO (or similar frameworks) to speed up your model inference - especially useful if you're using ensembles or want to deploy post-competition.
\nModel optimization frameworks like:
\ncan help you:
\n torch\n\nmodel.() \ndummy_input = torch.randn(, , , ) \n\ntorch.onnx.export(\n model, \n dummy_input, \n ,\n input_names=[], \n output_names=[],\n opset_version=,\n dynamic_axes={: {: }, : {: }}\n)\nKaggle does not support the full OpenVINO CLI (Model Optimizer), so you must convert the ONNX model on your local machine:
\nmo --input_model model.onnx --output_dir openvino_model\nThis will generate:
\nmodel.xml (model architecture) model.bin (model weights)Create a Kaggle Dataset and upload:
\nmodel.xml model.bin model.onnx, requirements.txt, OpenVINO .whl filesExample model uploaded: https://www.kaggle.com/models/kurisew/efficientnet_b0/PyTorch/openvino/1
\nSince internet access is disabled, use:
\n! python -m pip install --no-index \\\n --find-links=../input/openvino-wheels \\\n -r ../input/openvino-wheels/requirements.txt\nUpload your .whl files and requirements.txt to the ../input/openvino-wheels directory.
openvino.runtime Core\n\ncore = Core()\nmodel = core.read_model()\ncompiled_model = core.compile_model(model, device_name=)\n\n\nresult = compiled_model(inputs={: preprocessed_mel})[]\nMake sure your input tensor shape matches what was used in the ONNX export step.
\nUsed in the LB 0.874 baseline notebook.
\nWhile reviewing this OpenVINO notebook, I noticed it’s the only one using OpenVINO for inference in BirdCLEF 2025. Since there's no discussion around it yet, here’s a quick guide on how to use OpenVINO (or similar frameworks) to speed up your model inference - especially useful if you're using ensembles or want to deploy post-competition.
\nModel optimization frameworks like:
\ncan help you:
\n torch\n\nmodel.() \ndummy_input = torch.randn(, , , ) \n\ntorch.onnx.export(\n model, \n dummy_input, \n ,\n input_names=[], \n output_names=[],\n opset_version=,\n dynamic_axes={: {: }, : {: }}\n)\nKaggle does not support the full OpenVINO CLI (Model Optimizer), so you must convert the ONNX model on your local machine:
\nmo --input_model model.onnx --output_dir openvino_model\nThis will generate:
\nmodel.xml (model architecture) model.bin (model weights)Create a Kaggle Dataset and upload:
\nmodel.xml model.bin model.onnx, requirements.txt, OpenVINO .whl filesExample model uploaded: https://www.kaggle.com/models/kurisew/efficientnet_b0/PyTorch/openvino/1
\nSince internet access is disabled, use:
\n! python -m pip install --no-index \\\n --find-links=../input/openvino-wheels \\\n -r ../input/openvino-wheels/requirements.txt\nUpload your .whl files and requirements.txt to the ../input/openvino-wheels directory.
openvino.runtime Core\n\ncore = Core()\nmodel = core.read_model()\ncompiled_model = core.compile_model(model, device_name=)\n\n\nresult = compiled_model(inputs={: preprocessed_mel})[]\nMake sure your input tensor shape matches what was used in the ONNX export step.
\nUsed in the LB 0.874 baseline notebook.
\nShoutout and thanks to https://www.kaggle.com/code/kurisew/bird25-openvino-ensemble-infer-baseline-lb-874/notebook author: @kurisew
", "rawMarkdown": "Shoutout and thanks to https://www.kaggle.com/code/kurisew/bird25-openvino-ensemble-infer-baseline-lb-874/notebook author: @kurisew", "votes": null } ], "comments": [ { "id": 3216648, "author_name": "aayush26", "author_url": "", "post_date": "06/03/2025 22:17:55", "content": "Shoutout and thanks to https://www.kaggle.com/code/kurisew/bird25-openvino-ensemble-infer-baseline-lb-874/notebook author: @kurisew
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3216647": "While reviewing [this OpenVINO notebook](https://www.kaggle.com/code/kurisew/bird25-openvino-ensemble-infer-baseline-lb-874/notebook), I noticed it’s the only one using **OpenVINO** for inference in BirdCLEF 2025. Since there's no discussion around it yet, here’s a quick guide on how to use OpenVINO (or similar frameworks) to **speed up your model inference** - especially useful if you're using ensembles or want to deploy post-competition.\n\n---\n\n## ⚡ Why OpenVINO (and Similar Frameworks)?\n\nModel optimization frameworks like:\n\n- **OpenVINO** (Intel CPUs, edge devices) \n- **ONNX Runtime** (cross-platform) \n- **TensorRT** (NVIDIA GPUs) \n- **TFLite** (mobile & embedded)\n\ncan help you:\n\n- Speed up inference \n- Reduce memory use \n- Run on CPU-only environments (like Kaggle notebooks) \n- Accelerate ensemble inference\n\n---\n\n## 🛠️ Step-by-Step Guide: Using OpenVINO in Kaggle\n\n### ✅ Step 1: Convert PyTorch Model to ONNX\n\n```python\nimport torch\n\nmodel.eval() # set model to inference mode\ndummy_input = torch.randn(1, 1, 128, 500) # match your spectrogram input shape\n\ntorch.onnx.export(\n model, \n dummy_input, \n \"model.onnx\",\n input_names=[\"input\"], \n output_names=[\"output\"],\n opset_version=12,\n dynamic_axes={\"input\": {0: \"batch\"}, \"output\": {0: \"batch\"}}\n)\n```\n\n---\n\n### ❌ Step 2: Convert to OpenVINO IR (Outside Kaggle)\n\nKaggle does **not** support the full OpenVINO CLI (Model Optimizer), so you must convert the ONNX model on your **local machine**:\n\n```bash\nmo --input_model model.onnx --output_dir openvino_model\n```\n\nThis will generate:\n\n- `model.xml` (model architecture) \n- `model.bin` (model weights)\n\n---\n\n### ⬆️ Step 3: Upload Files to Kaggle Dataset\n\nCreate a [Kaggle Dataset](https://www.kaggle.com/datasets) and upload:\n\n- `model.xml` \n- `model.bin` \n- Optional: `model.onnx`, `requirements.txt`, OpenVINO `.whl` files\n\nExample model uploaded: https://www.kaggle.com/models/kurisew/efficientnet_b0/PyTorch/openvino/1\n\n---\n\n### ✅ Step 4: Install OpenVINO in Kaggle (Offline Mode)\n\nSince internet access is disabled, use:\n\n```bash\n! python -m pip install --no-index \\\n --find-links=../input/openvino-wheels \\\n -r ../input/openvino-wheels/requirements.txt\n```\n\nUpload your `.whl` files and `requirements.txt` to the `../input/openvino-wheels` directory.\n\n---\n\n### ✅ Step 5: Inference Using OpenVINO in Kaggle\n\n```python\nfrom openvino.runtime import Core\n\ncore = Core()\nmodel = core.read_model(\"path_to/model.xml\")\ncompiled_model = core.compile_model(model, device_name=\"CPU\")\n\n# Run inference\nresult = compiled_model(inputs={\"input\": preprocessed_mel})[\"output\"]\n```\n\nMake sure your input tensor shape matches what was used in the ONNX export step.\n\n---\n\n## 📈 Why This Helps in BirdCLEF 2025\n\n- Makes ensembling heavy backbones like EfficientNetV2 or SEResNeXt faster \n- Saves time during final submission \n- Ideal for offline deployment (Raspberry Pi, mobile devices, etc.)\n\nUsed in the [LB 0.874 baseline notebook](https://www.kaggle.com/code/kurisew/bird25-openvino-ensemble-infer-baseline-lb-874/notebook).\n\n---", "3216648": "Shoutout and thanks to https://www.kaggle.com/code/kurisew/bird25-openvino-ensemble-infer-baseline-lb-874/notebook author: @kurisew" }, "source": "meta" }