{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.10","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"gpu","dataSources":[{"sourceId":13333,"databundleVersionId":862146,"sourceType":"competition"},{"sourceId":6014291,"sourceType":"datasetVersion","datasetId":3443156}],"dockerImageVersionId":30498,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# Table of Contents \n* [Imports and Installations](#imports)\n* [Read the Data](#read_data)\n * [Preaparing Train Dataframe](#prep_train)\n * [Preaparing Test Dataframe](#prep_test)\n* [Visualizing the Images](#viz_img)\n* [Visualizing the Segmentation Masks](#viz_mask)\n* [Visualizing the Images with Segmentation Masks](#viz_img_mask)\n* [Custom Data Generator](#data_gen)\n* [Train Test Split](#data_split)\n* [Model Definitions](#model_def)\n* [Custom Loss, metric and Optimizer classes](#loss_optim)\n* [Setting up wandb for model logging](#wandb)\n* [Defining Callbacks](#model_callbacks)\n* [Training the Model](#model_train)\n* [Saving and Downloading the trained Model](#model_save)\n* [Evaluate model on Test Set and make submission](#model_eval_test)\n* [Exploring Predicticted Mask on Validation Set](#model_eval_val)\n* [Acknowledgements](#ack)","metadata":{}},{"cell_type":"markdown","source":"# Imports and Installations \n[Go back to the Table of Contents](#contents)","metadata":{}},{"cell_type":"code","source":"import os\nimport io\nimport cv2\nimport time\nimport math\nimport random\nimport logging\nimport warnings\n\nwarnings.filterwarnings(\"ignore\")\n#logging.getLogger(\"tensorflow\").setLevel(logging.WARNING)\n#os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","execution":{"iopub.status.busy":"2025-04-17T16:15:32.702453Z","iopub.execute_input":"2025-04-17T16:15:32.702712Z","iopub.status.idle":"2025-04-17T16:15:32.929680Z","shell.execute_reply.started":"2025-04-17T16:15:32.702690Z","shell.execute_reply":"2025-04-17T16:15:32.929039Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import numpy as np\nimport pandas as pd\nimport tensorflow as tf\nimport albumentations as A\nimport matplotlib.pyplot as plt\nfrom PIL import Image\nfrom sklearn.model_selection import train_test_split","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:32.931760Z","iopub.execute_input":"2025-04-17T16:15:32.932519Z","iopub.status.idle":"2025-04-17T16:15:41.995141Z","shell.execute_reply.started":"2025-04-17T16:15:32.932483Z","shell.execute_reply":"2025-04-17T16:15:41.994229Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(\"GPU available!\" if tf.test.is_gpu_available() else \"GPU is not available\")","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:41.996312Z","iopub.execute_input":"2025-04-17T16:15:41.996657Z","iopub.status.idle":"2025-04-17T16:15:44.061402Z","shell.execute_reply.started":"2025-04-17T16:15:41.996624Z","shell.execute_reply":"2025-04-17T16:15:44.060142Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"SEED = 42\nrandom.seed(SEED)\nnp.random.seed(SEED)\ntf.random.set_seed(SEED)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:44.062632Z","iopub.execute_input":"2025-04-17T16:15:44.062979Z","iopub.status.idle":"2025-04-17T16:15:44.077319Z","shell.execute_reply.started":"2025-04-17T16:15:44.062937Z","shell.execute_reply":"2025-04-17T16:15:44.076617Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"SEED = 42\n\nIMG_WIDTH = 2100\nIMG_HEIGHT = 1400\n\n# training params\nR_WIDTH = 576\nR_HEIGHT = 384\nNUM_CHANNELS = 3\nNUM_CLASSES = 4\nNUM_EPOCHS = 20 \nBATCH_SIZE = 4\nTEST_BATCH_SIZE = 32\n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:44.078418Z","iopub.execute_input":"2025-04-17T16:15:44.078712Z","iopub.status.idle":"2025-04-17T16:15:44.089114Z","shell.execute_reply.started":"2025-04-17T16:15:44.078688Z","shell.execute_reply":"2025-04-17T16:15:44.088380Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(tf.__version__)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:44.090007Z","iopub.execute_input":"2025-04-17T16:15:44.090248Z","iopub.status.idle":"2025-04-17T16:15:44.100666Z","shell.execute_reply.started":"2025-04-17T16:15:44.090227Z","shell.execute_reply":"2025-04-17T16:15:44.099830Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"- pip install segmentation-models \nhttps://github.com/qubvel/segmentation_models","metadata":{}},{"cell_type":"code","source":"!pip install segmentation-models","metadata":{"_kg_hide-input":false,"_kg_hide-output":true,"execution":{"iopub.status.busy":"2025-04-17T16:15:44.103870Z","iopub.execute_input":"2025-04-17T16:15:44.104103Z","iopub.status.idle":"2025-04-17T16:15:53.369029Z","shell.execute_reply.started":"2025-04-17T16:15:44.104083Z","shell.execute_reply":"2025-04-17T16:15:53.368156Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"https://stackoverflow.com/questions/75433717/module-keras-utils-generic-utils-has-no-attribute-get-custom-objects-when-im","metadata":{}},{"cell_type":"code","source":"import os\nos.environ[\"SM_FRAMEWORK\"] = \"tf.keras\"\n\nfrom tensorflow import keras\nimport segmentation_models as sm","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:53.370262Z","iopub.execute_input":"2025-04-17T16:15:53.370527Z","iopub.status.idle":"2025-04-17T16:15:53.392891Z","shell.execute_reply.started":"2025-04-17T16:15:53.370501Z","shell.execute_reply":"2025-04-17T16:15:53.392088Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Read the Data \n[Go back to the Table of Contents](#contents)","metadata":{}},{"cell_type":"code","source":"data_dir = '../input/understanding_cloud_organization'\ntrain_csv_path = os.path.join(data_dir,'train.csv')\ntest_csv_path = os.path.join(data_dir,\"sample_submission.csv\")\ntrain_image_path = os.path.join(data_dir,'train_images')\ntest_image_path = os.path.join(data_dir,'test_images')","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:53.393779Z","iopub.execute_input":"2025-04-17T16:15:53.394016Z","iopub.status.idle":"2025-04-17T16:15:53.398471Z","shell.execute_reply.started":"2025-04-17T16:15:53.393995Z","shell.execute_reply":"2025-04-17T16:15:53.397519Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":" ### Preaparing Train Dataframe \n [Go back to the Table of Contents](#contents)","metadata":{}},{"cell_type":"code","source":"train_df = pd.read_csv(train_csv_path).fillna(-1)\ntrain_df.head()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:53.399442Z","iopub.execute_input":"2025-04-17T16:15:53.399701Z","iopub.status.idle":"2025-04-17T16:15:57.974724Z","shell.execute_reply.started":"2025-04-17T16:15:53.399668Z","shell.execute_reply":"2025-04-17T16:15:57.973848Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"train_df['Image_Id'] = train_df['Image_Label'].apply(lambda x: x.split('_')[0])\ntrain_df['Label'] = train_df['Image_Label'].apply(lambda x: x.split('_')[1])\ntrain_df.head()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:57.975554Z","iopub.execute_input":"2025-04-17T16:15:57.975792Z","iopub.status.idle":"2025-04-17T16:15:58.010340Z","shell.execute_reply.started":"2025-04-17T16:15:57.975772Z","shell.execute_reply":"2025-04-17T16:15:58.008940Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"train_df['Label_EncodedPixels'] = train_df.apply(lambda row: (row['Label'], row['EncodedPixels']), axis = 1)\ntrain_df.head()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:58.011845Z","iopub.execute_input":"2025-04-17T16:15:58.012172Z","iopub.status.idle":"2025-04-17T16:15:58.196445Z","shell.execute_reply.started":"2025-04-17T16:15:58.012142Z","shell.execute_reply":"2025-04-17T16:15:58.195633Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"grouped_EncodedPixels = train_df.groupby('Image_Id')['Label_EncodedPixels'].apply(list)\ngrouped_EncodedPixels.head()\ngrouped_EncodedPixels.info()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:58.197627Z","iopub.execute_input":"2025-04-17T16:15:58.199314Z","iopub.status.idle":"2025-04-17T16:15:58.305312Z","shell.execute_reply.started":"2025-04-17T16:15:58.199290Z","shell.execute_reply":"2025-04-17T16:15:58.304649Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"train_df = grouped_EncodedPixels.to_frame().reset_index()\ntrain_df.head()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:58.307734Z","iopub.execute_input":"2025-04-17T16:15:58.308020Z","iopub.status.idle":"2025-04-17T16:15:58.321220Z","shell.execute_reply.started":"2025-04-17T16:15:58.307998Z","shell.execute_reply":"2025-04-17T16:15:58.320400Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"labels = ['Fish', 'Flower', 'Gravel', 'Sugar']\n\nfor label in labels:\n train_df = train_df.assign(**{label: 0})\nfor index, row in train_df.iterrows():\n for item in row['Label_EncodedPixels']:\n label, value = item\n if value == -1:\n bool_value = 0\n else:\n bool_value = 1\n train_df.loc[index, label] = bool_value\n\ntrain_df['classes'] = train_df.apply(lambda row: [col for col in labels if row[col] == 1], axis=1)\n\ntrain_df.head()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:15:58.322114Z","iopub.execute_input":"2025-04-17T16:15:58.322309Z","iopub.status.idle":"2025-04-17T16:16:00.408236Z","shell.execute_reply.started":"2025-04-17T16:15:58.322291Z","shell.execute_reply":"2025-04-17T16:16:00.407380Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"train_df[labels].sum().plot(kind='bar')\nplt.xlabel('Columns')\nplt.ylabel('Frequency')","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:00.409401Z","iopub.execute_input":"2025-04-17T16:16:00.409992Z","iopub.status.idle":"2025-04-17T16:16:00.661347Z","shell.execute_reply.started":"2025-04-17T16:16:00.409957Z","shell.execute_reply":"2025-04-17T16:16:00.660491Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"train_df.info()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:00.662442Z","iopub.execute_input":"2025-04-17T16:16:00.662702Z","iopub.status.idle":"2025-04-17T16:16:00.675420Z","shell.execute_reply.started":"2025-04-17T16:16:00.662679Z","shell.execute_reply":"2025-04-17T16:16:00.674629Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Finding the index of images having all the masks\nfor ix,item in enumerate(train_df['Label_EncodedPixels'][:100]):\n c1=item[0][-1]!=-1\n c2=item[1][-1]!=-1\n c3=item[2][-1]!=-1\n c4=item[3][-1]!=-1\n if c1 and c2 and c3 and c4:\n print(ix)\n ","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:00.676271Z","iopub.execute_input":"2025-04-17T16:16:00.676475Z","iopub.status.idle":"2025-04-17T16:16:00.686800Z","shell.execute_reply.started":"2025-04-17T16:16:00.676456Z","shell.execute_reply":"2025-04-17T16:16:00.685984Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"train_df.loc[28]","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:00.687818Z","iopub.execute_input":"2025-04-17T16:16:00.688115Z","iopub.status.idle":"2025-04-17T16:16:00.700004Z","shell.execute_reply.started":"2025-04-17T16:16:00.688085Z","shell.execute_reply":"2025-04-17T16:16:00.699306Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"for ix,item in enumerate(os.listdir(train_image_path)):\n if item == \"015aa06.jpg\":\n print(ix)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:00.700851Z","iopub.execute_input":"2025-04-17T16:16:00.701108Z","iopub.status.idle":"2025-04-17T16:16:00.890171Z","shell.execute_reply.started":"2025-04-17T16:16:00.701075Z","shell.execute_reply":"2025-04-17T16:16:00.889463Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### Preaparing Test Dataframe \n[Go back to the Table of Contents](#contents)","metadata":{}},{"cell_type":"code","source":"test_df = pd.read_csv(test_csv_path).fillna(-1)\ntest_df.head()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:00.891040Z","iopub.execute_input":"2025-04-17T16:16:00.891270Z","iopub.status.idle":"2025-04-17T16:16:00.923669Z","shell.execute_reply.started":"2025-04-17T16:16:00.891249Z","shell.execute_reply":"2025-04-17T16:16:00.922865Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"test_df['Image_Id'] = test_df['Image_Label'].apply(lambda x: x.split('_')[0])\ntest_df['Label'] = test_df['Image_Label'].apply(lambda x: x.split('_')[1])\ntest_df.head()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:00.929259Z","iopub.execute_input":"2025-04-17T16:16:00.929500Z","iopub.status.idle":"2025-04-17T16:16:00.949646Z","shell.execute_reply.started":"2025-04-17T16:16:00.929479Z","shell.execute_reply":"2025-04-17T16:16:00.948659Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"test_df['Label_EncodedPixels'] = test_df.apply(lambda row: (row['Label'], row['EncodedPixels']), axis = 1)\ntest_df.head()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:00.950751Z","iopub.execute_input":"2025-04-17T16:16:00.951019Z","iopub.status.idle":"2025-04-17T16:16:01.066936Z","shell.execute_reply.started":"2025-04-17T16:16:00.950988Z","shell.execute_reply":"2025-04-17T16:16:01.066074Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"grouped_EncodedPixels = test_df.groupby('Image_Id')['Label_EncodedPixels'].apply(list)\ngrouped_EncodedPixels.head()\ngrouped_EncodedPixels.info()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:01.067956Z","iopub.execute_input":"2025-04-17T16:16:01.068173Z","iopub.status.idle":"2025-04-17T16:16:01.137708Z","shell.execute_reply.started":"2025-04-17T16:16:01.068154Z","shell.execute_reply":"2025-04-17T16:16:01.136836Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"test_df = grouped_EncodedPixels.to_frame().reset_index()\ntest_df.head()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:01.138724Z","iopub.execute_input":"2025-04-17T16:16:01.138967Z","iopub.status.idle":"2025-04-17T16:16:01.151619Z","shell.execute_reply.started":"2025-04-17T16:16:01.138946Z","shell.execute_reply":"2025-04-17T16:16:01.150815Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Visualizing the Images \n[Go back to the Table of Contents](#contents)
\n`PIL or CV2 Image (image_width, image_height) == Numpy Array (image_height, image_width)`","metadata":{}},{"cell_type":"code","source":"def batchDataLoader(image_dir,img_w= 512, img_h=512, num_channel =4, Batch_Size=32):\n \n while True:\n k=0\n image_ids = os.listdir(image_dir)\n num_batches = math.ceil(len(image_ids)/Batch_Size)\n \n for batch_no in range(1,num_batches+1): \n if batch_no < num_batches:\n batch_size = Batch_Size\n batch_image_ids = image_ids[k:k+batch_size]\n image_batch = np.zeros((batch_size, img_h, img_w, num_channel),dtype=np.uint8)\n for i in range(batch_size):\n path = os.path.join(image_dir, image_ids[i])\n img = cv2.imread(path)\n img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n image_batch[i] = img\n # for the last batch which could be fractional\n if batch_no == num_batches:\n batch_image_ids = image_ids[k:]\n batch_size = len(batch_image_ids)\n image_batch = np.zeros((batch_size, img_h, img_w, num_channel),dtype=np.uint8)\n for i in range(batch_size):\n path = os.path.join(image_dir, image_ids[i])\n img = cv2.imread(path)\n img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n image_batch[i] = img\n \n k = k+batch_size\n print(f\"batch_no = {batch_no}\")\n yield image_batch","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:01.152695Z","iopub.execute_input":"2025-04-17T16:16:01.152940Z","iopub.status.idle":"2025-04-17T16:16:01.163538Z","shell.execute_reply.started":"2025-04-17T16:16:01.152919Z","shell.execute_reply":"2025-04-17T16:16:01.162848Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"img_width = 2100\nimg_height = 1400\nnum_channel = 3\nBATCH_SIZE = 32\ncurrent_batch = batchDataLoader(train_image_path,img_width,img_height, num_channel, BATCH_SIZE)\n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:01.164433Z","iopub.execute_input":"2025-04-17T16:16:01.164899Z","iopub.status.idle":"2025-04-17T16:16:01.176357Z","shell.execute_reply.started":"2025-04-17T16:16:01.164877Z","shell.execute_reply":"2025-04-17T16:16:01.175615Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"images = next(current_batch)\nprint(images.shape)\nplt.figure(figsize=(24,8))\nfor i in range(8):\n ax = plt.subplot(2,4, i+1)\n plt.imshow(images[i])\n #plt.title(labels[i])\n plt.axis(\"off\")","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:01.177398Z","iopub.execute_input":"2025-04-17T16:16:01.177746Z","iopub.status.idle":"2025-04-17T16:16:07.236429Z","shell.execute_reply.started":"2025-04-17T16:16:01.177713Z","shell.execute_reply":"2025-04-17T16:16:07.235536Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Visualizing the Segmentation Masks \n[Go back to the Table of Contents](#contents)
","metadata":{}},{"cell_type":"code","source":"def rle_to_mask(rle_string, height, width):\n '''\n convert RLE(run length encoding) string to numpy array\n\n Parameters: \n rle_string (str): string of rle encoded mask\n height (int): height of the mask\n width (int): width of the mask \n\n Returns: \n numpy.array: numpy array of the mask\n '''\n \n rows, cols = height, width\n \n if rle_string == -1:\n return np.zeros((height,width))\n else:\n rle_numbers = [int(num_string) for num_string in rle_string.split(' ')]\n rle_pairs = np.array(rle_numbers).reshape(-1,2)\n img = np.zeros(rows*cols, dtype=np.uint8)\n for index, length in rle_pairs:\n index -= 1\n img[index:index+length] = 255\n img = img.reshape(cols,rows)\n img = img.T\n img = img/255.0\n return img\n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:07.238011Z","iopub.execute_input":"2025-04-17T16:16:07.238327Z","iopub.status.idle":"2025-04-17T16:16:07.245802Z","shell.execute_reply.started":"2025-04-17T16:16:07.238299Z","shell.execute_reply":"2025-04-17T16:16:07.244984Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def get_masks_by_img_id(dataframe, image_id):\n masks = np.zeros((img_height,img_width,4))\n rle_masks = list(dataframe[dataframe['Image_Id'] == image_id]['Label_EncodedPixels'])[0]\n fish_mask = rle_to_mask(rle_masks[0][1], img_height, img_width)\n flower_mask = rle_to_mask(rle_masks[1][1], img_height, img_width)\n gravel_mask = rle_to_mask(rle_masks[2][1], img_height, img_width)\n sugar_mask = rle_to_mask(rle_masks[3][1], img_height, img_width)\n mask_list = [fish_mask,flower_mask,gravel_mask,sugar_mask]\n for ix, mask in enumerate(mask_list):\n masks[:,:,ix] = mask\n return masks","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:07.246924Z","iopub.execute_input":"2025-04-17T16:16:07.247552Z","iopub.status.idle":"2025-04-17T16:16:07.262633Z","shell.execute_reply.started":"2025-04-17T16:16:07.247521Z","shell.execute_reply":"2025-04-17T16:16:07.261807Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"image_id = '0011165.jpg'\npath = os.path.join(train_image_path,image_id)\nimg = cv2.imread(path)\nimg = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\nrle = list(train_df[train_df['Image_Id'] == image_id]['Label_EncodedPixels'])[0][0][1]\n\nm = rle_to_mask(rle,img_height,img_width)\nm = cv2.resize(m, (384,256),interpolation=cv2.INTER_LINEAR)\nm = (m>0).astype(int)\nplt.imshow(m)\nprint(m.shape)\nprint(np.unique(m))\nprint(np.argwhere(m==1)[0])","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:07.263673Z","iopub.execute_input":"2025-04-17T16:16:07.263915Z","iopub.status.idle":"2025-04-17T16:16:07.622119Z","shell.execute_reply.started":"2025-04-17T16:16:07.263895Z","shell.execute_reply":"2025-04-17T16:16:07.621291Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"image_id = os.listdir(train_image_path)[1]\n#image_id = 'f516a20.jpg'\nmasks = get_masks_by_img_id(train_df, image_id)\nmasks.shape","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:07.623209Z","iopub.execute_input":"2025-04-17T16:16:07.623461Z","iopub.status.idle":"2025-04-17T16:16:07.778239Z","shell.execute_reply.started":"2025-04-17T16:16:07.623438Z","shell.execute_reply":"2025-04-17T16:16:07.777379Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"image_id = os.listdir(train_image_path)[5390]\n#image_id = 'f516a20.jpg'\nmasks = get_masks_by_img_id(train_df, image_id)\nprint(image_id)\nplt.figure(figsize=(24,4))\nfor ix in range(masks.shape[-1]):\n ax = plt.subplot(1,4, ix+1)\n plt.imshow(masks[:,:,ix],cmap=None)\n plt.axis(\"off\");","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:07.779267Z","iopub.execute_input":"2025-04-17T16:16:07.779516Z","iopub.status.idle":"2025-04-17T16:16:09.249956Z","shell.execute_reply.started":"2025-04-17T16:16:07.779494Z","shell.execute_reply":"2025-04-17T16:16:09.249064Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Visualizing the Images with Segmentation Masks \n[Go back to the Table of Contents](#contents)
","metadata":{}},{"cell_type":"code","source":"def draw_label_on_mask(mask, label, obj=plt):\n '''\n Function to add labels to the image.\n '''\n if np.sum(mask) > 0:\n y,x = 0,0\n y,x = np.argwhere(mask==1)[0]\n y,x = y+50,x+20 \n obj.text(x,y,label,color='white',)\n return None","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:09.251133Z","iopub.execute_input":"2025-04-17T16:16:09.251439Z","iopub.status.idle":"2025-04-17T16:16:09.256343Z","shell.execute_reply.started":"2025-04-17T16:16:09.251415Z","shell.execute_reply":"2025-04-17T16:16:09.255433Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"image_id = os.listdir(train_image_path)[5390]\npath = os.path.join(train_image_path,image_id)\nimg = cv2.imread(path)\nimg = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\nimg = cv2.resize(img,(384,256))\nimg = img.astype(np.float32)\nimg = img/255.0\n#img -= img.mean()\n#img /= img.std()\n#standarization changes the color\nprint(img.shape)\nplt.imshow(img);","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:09.257312Z","iopub.execute_input":"2025-04-17T16:16:09.257530Z","iopub.status.idle":"2025-04-17T16:16:09.645734Z","shell.execute_reply.started":"2025-04-17T16:16:09.257501Z","shell.execute_reply":"2025-04-17T16:16:09.644856Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"image_id = os.listdir(train_image_path)[5390]\npath = os.path.join(train_image_path,image_id)\nimg = cv2.imread(path)\nimg = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\nmasks = get_masks_by_img_id(train_df, image_id)\nmask = masks[:,:,1]\nmask = np.clip(mask,0,1)\nmask = np.ma.masked_where(mask == 0, mask)\nplt.imshow(img)\nplt.imshow(mask,alpha=0.7,cmap='PuRd_r')\ndraw_label_on_mask(mask,\"Flower\")\nplt.axis('off');","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:09.646836Z","iopub.execute_input":"2025-04-17T16:16:09.647098Z","iopub.status.idle":"2025-04-17T16:16:10.860290Z","shell.execute_reply.started":"2025-04-17T16:16:09.647075Z","shell.execute_reply":"2025-04-17T16:16:10.859418Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"image_id = os.listdir(train_image_path)[5390]\npath = os.path.join(train_image_path,image_id)\nimg = cv2.imread(path)\nimg = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\nmasks = get_masks_by_img_id(train_df, image_id)\nmask = masks[:,:,1]\nmask = np.clip(mask,0,1)\nmask = np.ma.masked_where(mask == 0, mask)\nbbox = cv2.boundingRect(mask.astype(np.uint8))\ncv2.rectangle(img, bbox, (0, 255, 0), 5)\nplt.imshow(img)\n#plt.imshow(mask,alpha=0.7,cmap='PuRd_r')\ndraw_label_on_mask(mask,\"Flower\")\nplt.axis('off');","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:10.861386Z","iopub.execute_input":"2025-04-17T16:16:10.861642Z","iopub.status.idle":"2025-04-17T16:16:12.000416Z","shell.execute_reply.started":"2025-04-17T16:16:10.861619Z","shell.execute_reply":"2025-04-17T16:16:11.999618Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"image_id = os.listdir(train_image_path)[5390]\nmasks = get_masks_by_img_id(train_df, image_id)\nmasks = (masks[:,:,0], masks[:,:,1],masks[:,:,2],masks[:,:,3])\n\npath = os.path.join(train_image_path,image_id)\nimg = cv2.imread(path)\nimg = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\ncolormaps = ['PuRd_r', 'Blues_r', 'Purples_r','winter_r'] # colormap_r = inverse colormap\nmask_labels = ['Fish', 'Flower', 'Gravel', 'Sugar']\nplt.figure(figsize=(15,10))\nfor i,(mask,cmap,label)in enumerate(zip(masks,colormaps,mask_labels)):\n mask = np.clip(mask,0,1)\n mask = np.ma.masked_where(mask == 0, mask)\n ax = plt.subplot(2,2, i+1)\n plt.imshow(img)\n plt.imshow(mask,alpha=0.7,cmap=cmap) \n draw_label_on_mask(mask,label)\n plt.axis(\"off\")","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:12.001508Z","iopub.execute_input":"2025-04-17T16:16:12.001800Z","iopub.status.idle":"2025-04-17T16:16:15.810653Z","shell.execute_reply.started":"2025-04-17T16:16:12.001777Z","shell.execute_reply":"2025-04-17T16:16:15.809626Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"image_id = os.listdir(train_image_path)[5390]\nmasks = get_masks_by_img_id(train_df, image_id)\nmasks = (masks[:,:,0], masks[:,:,1],masks[:,:,2],masks[:,:,3])\n\npath = os.path.join(train_image_path,image_id)\nimg = cv2.imread(path)\nimg = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n\ncolormaps = ['PuRd_r', 'Blues_r', 'Purples_r','winter_r']\nmask_labels = ['Fish', 'Flower', 'Gravel', 'Sugar']\nplt.figure(figsize=(32,8))\nplt.imshow(img)\nfor i,(mask,cmap,label) in enumerate(zip(masks,colormaps,mask_labels)):\n mask = np.clip(mask,0,1)\n mask = np.ma.masked_where(mask == 0, mask)\n plt.imshow(mask,alpha=0.7,cmap=cmap) # colormap_r = inverse colormap\n draw_label_on_mask(mask,label)\n plt.axis(\"off\")","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:15.811780Z","iopub.execute_input":"2025-04-17T16:16:15.812036Z","iopub.status.idle":"2025-04-17T16:16:18.839919Z","shell.execute_reply.started":"2025-04-17T16:16:15.812013Z","shell.execute_reply":"2025-04-17T16:16:18.839053Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import cv2\nimport numpy as np\nimport matplotlib\n\ndef show_bounding_boxes(image, mask, labels, colors):\n \"\"\"Shows the bounding boxes surrounding the polygon in the image, and\n adds labels to the bounding boxes.\n\n Args:\n image: The image.\n mask: The binary polygon mask.\n labels: The labels of the objects in the mask.\n colors: A list of colors to use for the bounding boxes.\n\n Returns:\n The image with the bounding boxes and labels drawn on it.\n \"\"\"\n\n # Find the bounding boxes of the polygon.\n bounding_boxes = []\n for i in range(mask.shape[-1]):\n bbox = cv2.boundingRect(mask[:, :, i])\n bounding_boxes.append(bbox)\n\n # Draw the bounding boxes on the image.\n for bbox, label, color_name in zip(bounding_boxes, labels, colors):\n rgb_color = matplotlib.colors.to_rgb(color_name)\n rgb_color = tuple(value * 255 for value in rgb_color)\n cv2.rectangle(image, bbox, rgb_color, 10)\n cv2.putText(image, label, (bbox[0], bbox[1] + 50),\n cv2.FONT_HERSHEY_SIMPLEX, 2, (255,255,255), 6)\n\n return image\n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:18.841152Z","iopub.execute_input":"2025-04-17T16:16:18.841482Z","iopub.status.idle":"2025-04-17T16:16:18.851051Z","shell.execute_reply.started":"2025-04-17T16:16:18.841452Z","shell.execute_reply":"2025-04-17T16:16:18.850076Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import matplotlib\nrgb_color = matplotlib.colors.to_rgb('darkblue')\nrgb_color = tuple(value * 255 for value in rgb_color)\nrgb_color","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:18.852209Z","iopub.execute_input":"2025-04-17T16:16:18.852849Z","iopub.status.idle":"2025-04-17T16:16:18.869331Z","shell.execute_reply.started":"2025-04-17T16:16:18.852817Z","shell.execute_reply":"2025-04-17T16:16:18.868422Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"image_id = os.listdir(train_image_path)[5390]\nmasks = get_masks_by_img_id(train_df, image_id)\nmasks = masks.astype(np.uint8)\n\n\npath = os.path.join(train_image_path,image_id)\nimg = cv2.imread(path)\nimg = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n\ncolors = ['maroon', 'darkblue', 'purple','teal']\nlabels = ['Fish', 'Flower', 'Gravel', 'Sugar']\n\nimg = show_bounding_boxes(img,masks,labels,colors)\n\nplt.figure(figsize=(32,8))\nplt.imshow(img)\nplt.axis(\"off\");","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:18.870722Z","iopub.execute_input":"2025-04-17T16:16:18.871322Z","iopub.status.idle":"2025-04-17T16:16:19.860498Z","shell.execute_reply.started":"2025-04-17T16:16:18.871285Z","shell.execute_reply":"2025-04-17T16:16:19.859637Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def show_img_with_masks(img,masks,comment=\"\"):\n \n colormaps = ['PuRd_r', 'Blues_r', 'Purples_r','winter_r']\n mask_labels = ['Fish', 'Flower', 'Gravel', 'Sugar']\n \n fig, axes = plt.subplots(1,6,figsize=(36,4))\n axes = axes.ravel()\n \n if img.shape[-1]!=3:\n img_cmap = 'gray'\n else:\n img_cmap=None\n \n for ix,axis in enumerate(axes):\n ix = ix%6\n axis.imshow(img,cmap=img_cmap)\n axis.axis('off')\n if ix==0:\n axis.set_title(\"Main Image\")\n elif ix==1:\n for i,(mask,cmap,label) in enumerate(zip(masks,colormaps,mask_labels)):\n mask = np.clip(mask,0,1)\n mask = np.ma.masked_where(mask == 0, mask)\n axis.imshow(mask,alpha=0.7,cmap=cmap)\n axis.set_title(f\"All the mask {comment}\")\n draw_label_on_mask(mask,label,axis)\n elif ix>=2:\n for i,(mask,cmap,label) in enumerate(zip(masks,colormaps,mask_labels)):\n mask = np.clip(mask,0,1)\n mask = np.ma.masked_where(mask == 0, mask)\n axis = axes[2+i]\n axis.imshow(mask,alpha=0.4,cmap=cmap)\n axis.set_title(f\"{label} {comment}\")\n draw_label_on_mask(mask,label,axis)\n plt.show()\n \n return None","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:19.861803Z","iopub.execute_input":"2025-04-17T16:16:19.862337Z","iopub.status.idle":"2025-04-17T16:16:19.873006Z","shell.execute_reply.started":"2025-04-17T16:16:19.862307Z","shell.execute_reply":"2025-04-17T16:16:19.872097Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"image_id = os.listdir(train_image_path)[5390]\npath = os.path.join(train_image_path,image_id)\nimg = cv2.imread(path)\nimg = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\nmasks = get_masks_by_img_id(train_df, image_id)\nmasks = (masks[:,:,0], masks[:,:,1],masks[:,:,2],masks[:,:,3])\nshow_img_with_masks(img,masks)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:19.874160Z","iopub.execute_input":"2025-04-17T16:16:19.874840Z","iopub.status.idle":"2025-04-17T16:16:28.063456Z","shell.execute_reply.started":"2025-04-17T16:16:19.874810Z","shell.execute_reply":"2025-04-17T16:16:28.062642Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"image_ids = os.listdir(train_image_path)[13:16]\nfor image_id in image_ids:\n path = os.path.join(train_image_path,image_id)\n img = cv2.imread(path)\n img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n masks = get_masks_by_img_id(train_df, image_id)\n masks = (masks[:,:,0], masks[:,:,1],masks[:,:,2],masks[:,:,3])\n show_img_with_masks(img,masks)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:28.064810Z","iopub.execute_input":"2025-04-17T16:16:28.065559Z","iopub.status.idle":"2025-04-17T16:16:51.852842Z","shell.execute_reply.started":"2025-04-17T16:16:28.065527Z","shell.execute_reply":"2025-04-17T16:16:51.851967Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Custom Data Generator \n[Go back to the Table of Contents](#contents)
","metadata":{"execution":{"iopub.status.busy":"2023-06-03T00:01:18.601333Z","iopub.execute_input":"2023-06-03T00:01:18.602029Z","iopub.status.idle":"2023-06-03T00:01:18.606731Z","shell.execute_reply.started":"2023-06-03T00:01:18.602004Z","shell.execute_reply":"2023-06-03T00:01:18.605097Z"}}},{"cell_type":"code","source":"def rle_to_mask(rle_string, height, width):\n '''\n convert RLE(run length encoding) string to numpy array\n\n Parameters: \n rle_string (str): string of rle encoded mask\n height (int): height of the mask\n width (int): width of the mask \n\n Returns: \n numpy.array: numpy array of the mask\n '''\n \n rows, cols = height, width\n \n if rle_string == -1:\n return np.zeros((height,width))\n else:\n rle_numbers = [int(num_string) for num_string in rle_string.split(' ')]\n rle_pairs = np.array(rle_numbers).reshape(-1,2)\n img = np.zeros(rows*cols, dtype=np.uint8)\n for index, length in rle_pairs:\n index -= 1\n img[index:index+length] = 255\n img = img.reshape(cols,rows)\n img = img.T\n img = img/255.0\n return img\n \ndef get_masks_by_img_id(dataframe, image_id):\n masks = np.zeros((img_height,img_width,4))\n rle_masks = list(dataframe[dataframe['Image_Id'] == image_id]['Label_EncodedPixels'])[0]\n fish_mask = rle_to_mask(rle_masks[0][1], img_height, img_width)\n flower_mask = rle_to_mask(rle_masks[1][1], img_height, img_width)\n gravel_mask = rle_to_mask(rle_masks[2][1], img_height, img_width)\n sugar_mask = rle_to_mask(rle_masks[3][1], img_height, img_width)\n mask_list = [fish_mask,flower_mask,gravel_mask,sugar_mask]\n for ix, mask in enumerate(mask_list):\n masks[:,:,ix] = mask\n return masks\n\ndef draw_label_on_mask(mask, label, obj=plt):\n '''\n Function to add labels to the image.\n '''\n if np.sum(mask) > 0:\n y,x = 0,0\n y,x = np.argwhere(mask==1)[0]\n y,x = y+10,x+5 \n obj.text(x,y,label,color='white',)\n return None\n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:51.853978Z","iopub.execute_input":"2025-04-17T16:16:51.854229Z","iopub.status.idle":"2025-04-17T16:16:51.863060Z","shell.execute_reply.started":"2025-04-17T16:16:51.854206Z","shell.execute_reply":"2025-04-17T16:16:51.862318Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import tensorflow as tf\nimport numpy as np\nimport math\n\nclass DataGenerator(tf.keras.utils.Sequence):\n \n def __init__(self,\n dataframe=None,\n root_dir=\".\",\n img_width=2100,\n img_height=1400,\n resize = False,\n resize_width=384,\n resize_height=256,\n mode='fit',\n augmentations=None,\n num_channels = 3,\n num_classes = 4,\n batch_size=32,\n shuffle=True, \n random_state=42): \n \n self.dataframe = dataframe\n self.filenames = list(dataframe['Image_Id']) if not dataframe is None else os.listdir(root_dir)\n self.root_dir = root_dir\n self.img_width = img_width\n self.img_height = img_height\n self.resize = resize\n self.resize_width = resize_width\n self.resize_height = resize_height\n self.mode = mode\n self.augmentations = augmentations\n self.num_channels = num_channels\n self.num_classes = num_classes\n self.total_samples = len(self.filenames)\n self.indexes = np.arange(len(self.filenames))\n self.batch_size = batch_size\n self.shuffle = shuffle\n self.random_state = random_state\n \n self.on_epoch_end()\n \n @property\n def image_shape(self):\n if not self.resize:\n img_shape = (self.img_height, self.img_width, self.num_channels)\n else:\n img_shape = (self.resize_height, self.resize_width, self.num_channels)\n return img_shape\n \n \n def __len__(self):\n num_batches_total = math.ceil(self.total_samples / self.batch_size)\n return num_batches_total\n \n def __getitem__(self,index):\n low = index * self.batch_size\n high = min(low + self.batch_size, self.total_samples)\n batch_files = self.filenames[low:high]\n \n if self.mode == 'fit':\n batch_X = self.__generate_X(batch_files)\n batch_y = self.__generate_y(batch_files)\n if self.augmentations is not None:\n batch_X, batch_y = self.augment_batch(batch_X, batch_y)\n batch_X, batch_y = batch_X.astype(np.float32), batch_y.astype(np.float32)\n return batch_X, batch_y\n elif self.mode == 'predict':\n batch_X = self.__generate_X(batch_files)\n batch_X = batch_X.astype(np.float32)\n return batch_X\n else:\n raise AttributeError('The mode parameter should be set to \"fit\" or \"predict\".')\n \n def __generate_X(self,batch_files):\n if self.resize:\n img_size = (self.resize_height, self.resize_width)\n else:\n img_size = (self.img_height, self.img_width)\n \n batch_images = np.zeros((len(batch_files),*img_size, self.num_channels), dtype=np.uint8)\n\n for ix,filename in enumerate(batch_files):\n img_path = os.path.join(self.root_dir, filename)\n if self.num_channels==3:\n img = cv2.imread(img_path)\n img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n else:\n img = cv2.imread(img_path,0)\n #img = np.expand_dims(img, axis=-1)\n if self.resize:\n img = cv2.resize(img,tuple(reversed(img_size)))\n if self.num_channels==1:\n img = np.expand_dims(img, axis=-1)\n \n img = img.astype(np.uint8)\n #img = img/255\n batch_images[ix] = img\n \n return batch_images\n \n def __generate_y(self,batch_files):\n if self.resize:\n mask_size = (self.resize_height, self.resize_width)\n else:\n mask_size = (self.img_height, self.img_width)\n \n batch_masks = np.zeros((len(batch_files),*mask_size, self.num_classes), dtype = np.uint8)\n \n for ix,filename in enumerate(batch_files):\n masks = self.get_masks_by_img_id(self.dataframe, filename)\n batch_masks[ix] = masks\n \n batch_masks = (batch_masks > 0).astype(int)\n batch_masks = batch_masks.astype(np.uint8)\n return batch_masks\n \n \n def on_epoch_end(self):\n self.indexes = np.arange(self.total_samples)\n if self.shuffle:\n np.random.seed(self.random_state)\n np.random.shuffle(self.indexes)\n \n\n def get_masks_by_img_id(self,dataframe,image_id):\n rle_masks = list(dataframe[dataframe['Image_Id'] == image_id]['Label_EncodedPixels'])[0]\n fish_mask = rle_to_mask(rle_masks[0][1], self.img_height, self.img_width)\n flower_mask = rle_to_mask(rle_masks[1][1], self.img_height, self.img_width)\n gravel_mask = rle_to_mask(rle_masks[2][1], self.img_height, self.img_width)\n sugar_mask = rle_to_mask(rle_masks[3][1], self.img_height, self.img_width)\n mask_list = [fish_mask,flower_mask,gravel_mask,sugar_mask]\n if self.resize:\n resized_mask_list = []\n for mask in mask_list:\n mask = cv2.resize(mask, (self.resize_width,self.resize_height))\n resized_mask_list.append(mask)\n masks = np.zeros((self.resize_height,self.resize_width,self.num_classes))\n \n for ix, mask in enumerate(resized_mask_list):\n masks[:,:,ix] = mask\n else:\n masks = np.zeros((self.img_height,self.img_width,self.num_classes)) \n for ix, mask in enumerate(mask_list):\n masks[:,:,ix] = mask\n \n return masks\n \n\n def augment_batch(self,batch_images, batch_masks):\n batch_X = np.zeros(batch_images.shape, dtype=np.float32)\n batch_y = np.zeros(batch_masks.shape, dtype=np.float32)\n for ix, (img,masks) in enumerate(zip(batch_images,batch_masks)):\n augmented = self.augmentations(image=img, mask=masks)\n img = augmented['image']\n masks = augmented['mask']\n batch_X[ix] = img\n batch_y[ix] = masks\n return batch_X, batch_y\n \n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:51.864250Z","iopub.execute_input":"2025-04-17T16:16:51.864562Z","iopub.status.idle":"2025-04-17T16:16:51.885252Z","shell.execute_reply.started":"2025-04-17T16:16:51.864533Z","shell.execute_reply":"2025-04-17T16:16:51.884501Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"augmentations = A.Compose([\n A.HorizontalFlip(p=0.5),\n A.VerticalFlip(p=0.5),\n A.ShiftScaleRotate(p=0.5),\n A.GridDistortion(p=0.5),\n A.OpticalDistortion(p=0.5),\n A.RandomGamma(),\n A.RandomBrightnessContrast(),\n A.CLAHE(),\n A.Equalize(),\n A.ToFloat()\n])","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:51.886224Z","iopub.execute_input":"2025-04-17T16:16:51.886776Z","iopub.status.idle":"2025-04-17T16:16:51.900873Z","shell.execute_reply.started":"2025-04-17T16:16:51.886744Z","shell.execute_reply":"2025-04-17T16:16:51.900139Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"data_generator = DataGenerator(train_df,\n train_image_path,\n resize=True,\n resize_width=576,\n resize_height=384,\n num_channels=3,\n augmentations=augmentations\n )\n\nprint(data_generator.total_samples)\nprint(len(data_generator.indexes))\nprint(data_generator.__len__())\nprint(data_generator.image_shape)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:51.901797Z","iopub.execute_input":"2025-04-17T16:16:51.902028Z","iopub.status.idle":"2025-04-17T16:16:51.915478Z","shell.execute_reply.started":"2025-04-17T16:16:51.902008Z","shell.execute_reply":"2025-04-17T16:16:51.914726Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"batch_X, batch_y = data_generator.__getitem__(1)\nprint(batch_X.shape)\nprint(batch_y.shape)\nprint(batch_X.dtype)\nprint(batch_y.dtype)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:51.916356Z","iopub.execute_input":"2025-04-17T16:16:51.916607Z","iopub.status.idle":"2025-04-17T16:16:55.851248Z","shell.execute_reply.started":"2025-04-17T16:16:51.916560Z","shell.execute_reply":"2025-04-17T16:16:55.850319Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"plt.imshow(batch_X[13])","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:55.852185Z","iopub.execute_input":"2025-04-17T16:16:55.852432Z","iopub.status.idle":"2025-04-17T16:16:56.225108Z","shell.execute_reply.started":"2025-04-17T16:16:55.852409Z","shell.execute_reply":"2025-04-17T16:16:56.224333Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"plt.imshow(batch_y[13][:,:,1])","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:56.226065Z","iopub.execute_input":"2025-04-17T16:16:56.226284Z","iopub.status.idle":"2025-04-17T16:16:56.489410Z","shell.execute_reply.started":"2025-04-17T16:16:56.226264Z","shell.execute_reply":"2025-04-17T16:16:56.488656Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"for img,masks in zip(batch_X[3:6],batch_y[3:6]):\n masks = (masks[:,:,0], masks[:,:,1],masks[:,:,2],masks[:,:,3])\n show_img_with_masks(img,masks)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:16:56.490709Z","iopub.execute_input":"2025-04-17T16:16:56.491329Z","iopub.status.idle":"2025-04-17T16:17:00.373641Z","shell.execute_reply.started":"2025-04-17T16:16:56.491295Z","shell.execute_reply":"2025-04-17T16:17:00.372849Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Train Test Split \n[Go back to the Table of Contents](#contents)
","metadata":{}},{"cell_type":"code","source":"df_train, df_val = train_test_split(train_df,test_size=0.1,random_state=42, stratify=train_df['classes'])\nprint(df_train.shape)\nprint(df_val.shape)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:00.374685Z","iopub.execute_input":"2025-04-17T16:17:00.374944Z","iopub.status.idle":"2025-04-17T16:17:00.388937Z","shell.execute_reply.started":"2025-04-17T16:17:00.374921Z","shell.execute_reply":"2025-04-17T16:17:00.387974Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"df_train[labels].sum().plot(kind='bar')\nplt.xlabel('Columns')\nplt.ylabel('Frequency')","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:00.390048Z","iopub.execute_input":"2025-04-17T16:17:00.390320Z","iopub.status.idle":"2025-04-17T16:17:00.536088Z","shell.execute_reply.started":"2025-04-17T16:17:00.390297Z","shell.execute_reply":"2025-04-17T16:17:00.535182Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"df_val[labels].sum().plot(kind='bar')\nplt.xlabel('Columns')\nplt.ylabel('Frequency')","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:00.537446Z","iopub.execute_input":"2025-04-17T16:17:00.537797Z","iopub.status.idle":"2025-04-17T16:17:00.686423Z","shell.execute_reply.started":"2025-04-17T16:17:00.537765Z","shell.execute_reply":"2025-04-17T16:17:00.685624Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"R_WIDTH = 576\nR_HEIGHT = 384\nNUM_CHANNELS = 3\nBATCH_SIZE = 4\nTEST_BATCH_SIZE = 32","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:00.687394Z","iopub.execute_input":"2025-04-17T16:17:00.687732Z","iopub.status.idle":"2025-04-17T16:17:00.692409Z","shell.execute_reply.started":"2025-04-17T16:17:00.687699Z","shell.execute_reply":"2025-04-17T16:17:00.691402Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"train_generator = DataGenerator(dataframe=df_train,\n root_dir=train_image_path,\n mode=\"fit\",\n resize=True,\n shuffle=True,\n resize_width=R_WIDTH,\n resize_height=R_HEIGHT,\n num_channels=NUM_CHANNELS,\n batch_size=BATCH_SIZE,\n augmentations=augmentations)\n\nprint(train_generator.total_samples)\nprint(len(train_generator.indexes))\nprint(train_generator.__len__())\n\nbatch_X, batch_y = train_generator.__getitem__(4)\nprint(batch_X.shape)\nprint(batch_y.shape)\nprint(batch_X.dtype)\nprint(batch_y.dtype)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:00.693244Z","iopub.execute_input":"2025-04-17T16:17:00.693520Z","iopub.status.idle":"2025-04-17T16:17:01.317599Z","shell.execute_reply.started":"2025-04-17T16:17:00.693491Z","shell.execute_reply":"2025-04-17T16:17:01.316729Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"val_generator = DataGenerator(dataframe=df_val,\n root_dir=train_image_path,\n mode=\"fit\",\n resize=True,\n shuffle=True,\n resize_width=R_WIDTH,\n resize_height=R_HEIGHT,\n num_channels=NUM_CHANNELS,\n batch_size=BATCH_SIZE,\n augmentations=augmentations)\n\nprint(val_generator.total_samples)\nprint(len(val_generator.indexes))\nprint(val_generator.__len__())\n\nbatch_X, batch_y = val_generator.__getitem__(1)\nprint(batch_X.shape)\nprint(batch_y.shape)\nprint(batch_X.dtype)\nprint(batch_y.dtype)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:01.318773Z","iopub.execute_input":"2025-04-17T16:17:01.319087Z","iopub.status.idle":"2025-04-17T16:17:01.847259Z","shell.execute_reply.started":"2025-04-17T16:17:01.319055Z","shell.execute_reply":"2025-04-17T16:17:01.846415Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"test_generator = DataGenerator(dataframe=test_df,\n root_dir=test_image_path,\n mode=\"predict\",\n resize=True,\n shuffle=False,\n resize_width=R_WIDTH,\n resize_height=R_HEIGHT,\n num_channels=NUM_CHANNELS,\n batch_size=TEST_BATCH_SIZE)\n\n\nprint(test_generator.total_samples)\nprint(len(test_generator.indexes))\nprint(test_generator.__len__())\n\nbatch_X = test_generator.__getitem__(1)\nprint(batch_X.shape)\nprint(batch_X.dtype)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:01.848337Z","iopub.execute_input":"2025-04-17T16:17:01.848606Z","iopub.status.idle":"2025-04-17T16:17:03.550404Z","shell.execute_reply.started":"2025-04-17T16:17:01.848561Z","shell.execute_reply":"2025-04-17T16:17:03.549517Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"eval_generator = DataGenerator(dataframe=df_val,\n root_dir=train_image_path,\n mode=\"fit\",\n resize=True,\n shuffle=False,\n resize_width=R_WIDTH,\n resize_height=R_HEIGHT,\n num_channels=NUM_CHANNELS,\n batch_size=TEST_BATCH_SIZE,\n augmentations=augmentations)\n\nprint(eval_generator.total_samples)\nprint(len(eval_generator.indexes))\nprint(eval_generator.__len__())\n\nbatch_X, batch_y = eval_generator.__getitem__(1)\nprint(batch_X.shape)\nprint(batch_y.shape)\nprint(batch_X.dtype)\nprint(batch_y.dtype)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:03.551541Z","iopub.execute_input":"2025-04-17T16:17:03.551880Z","iopub.status.idle":"2025-04-17T16:17:08.273740Z","shell.execute_reply.started":"2025-04-17T16:17:03.551852Z","shell.execute_reply":"2025-04-17T16:17:08.272806Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Model Definitions \n[Go back to the Table of Contents](#contents)
","metadata":{}},{"cell_type":"code","source":"from PIL import Image\nimport io\n\ndef viz_model(model):\n tf.keras.backend.clear_session()\n model_plot = tf.keras.utils.plot_model(model, show_shapes=True, rankdir='TB')\n model_plot = io.BytesIO(model_plot.data)\n model_plot = Image.open(model_plot)\n model_plot = np.array(model_plot)\n model_plot = np.transpose(model_plot, (1, 0, 2))\n model_plot = Image.fromarray(model_plot)\n model_plot = model_plot.transpose(Image.FLIP_TOP_BOTTOM)\n model_plot = np.array(model_plot)\n fig = plt.figure(figsize=(20,10))\n ax = fig.add_subplot(111)\n ax.imshow(model_plot)\n ax.axis('off')\n plt.show()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:08.274799Z","iopub.execute_input":"2025-04-17T16:17:08.275129Z","iopub.status.idle":"2025-04-17T16:17:08.281245Z","shell.execute_reply.started":"2025-04-17T16:17:08.275099Z","shell.execute_reply":"2025-04-17T16:17:08.280343Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### 1. Vanilla Unet OOP implementation","metadata":{}},{"cell_type":"code","source":"import tensorflow as tf\n\nclass UNet:\n def __init__(self,\n input_shape=(256,384,1),\n num_classes=4,\n activation='relu',\n num_filters=(32,64,128,256,512)):\n \n self.input_shape = input_shape\n self.num_classes = num_classes\n self.activation = activation\n self.num_filters = num_filters\n self.model = self.build()\n\n def conv_block(self, inputs, filters, kernel_size=3, activation='relu', padding='same'):\n conv = tf.keras.layers.Conv2D(filters, kernel_size, padding=padding)(inputs)\n conv = tf.keras.layers.BatchNormalization()(conv)\n conv = tf.keras.layers.Activation(activation)(conv)\n conv = tf.keras.layers.Conv2D(filters, kernel_size, padding=padding)(conv)\n conv = tf.keras.layers.BatchNormalization()(conv)\n conv = tf.keras.layers.Activation(activation)(conv)\n return conv\n\n def upsample_block(self, inputs, skip_connection, filters, kernel_size=2, activation='relu', padding='same'):\n upsample = tf.keras.layers.Conv2DTranspose(filters, kernel_size, strides=(2, 2), padding=padding)(inputs)\n merge = tf.keras.layers.concatenate([skip_connection, upsample], axis=3)\n conv = self.conv_block(merge, filters, kernel_size, activation, padding)\n return conv\n\n def build(self):\n # Input layer\n inputs = tf.keras.layers.Input(self.input_shape)\n\n # Contracting path\n conv1 = self.conv_block(inputs, self.num_filters[0],activation=self.activation)\n pool1 = tf.keras.layers.MaxPooling2D(pool_size=(2, 2))(conv1)\n\n conv2 = self.conv_block(pool1, self.num_filters[1],activation=self.activation)\n pool2 = tf.keras.layers.MaxPooling2D(pool_size=(2, 2))(conv2)\n\n conv3 = self.conv_block(pool2, self.num_filters[2],activation=self.activation)\n pool3 = tf.keras.layers.MaxPooling2D(pool_size=(2, 2))(conv3)\n\n conv4 = self.conv_block(pool3, self.num_filters[3],activation=self.activation)\n pool4 = tf.keras.layers.MaxPooling2D(pool_size=(2, 2))(conv4)\n\n conv5 = self.conv_block(pool4, self.num_filters[4],activation=self.activation)\n \n\n # Expansive path\n up6 = self.upsample_block(conv5, conv4, self.num_filters[3], activation=self.activation)\n up7 = self.upsample_block(up6, conv3, self.num_filters[2], activation=self.activation)\n up8 = self.upsample_block(up7, conv2, self.num_filters[1], activation=self.activation)\n up9 = self.upsample_block(up8, conv1, self.num_filters[0], activation=self.activation)\n\n # Output layer\n outputs = tf.keras.layers.Conv2D(self.num_classes, 1, activation='sigmoid')(up9)\n\n # Create model\n model = tf.keras.Model(inputs=inputs, outputs=outputs)\n return model\n\n \n def get_model(self):\n return self.model","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:08.282347Z","iopub.execute_input":"2025-04-17T16:17:08.282630Z","iopub.status.idle":"2025-04-17T16:17:08.298918Z","shell.execute_reply.started":"2025-04-17T16:17:08.282560Z","shell.execute_reply":"2025-04-17T16:17:08.298235Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"#tf.keras.backend.clear_session()\n#UNet().get_model().summary()\nmodel = UNet().get_model()\ntotal_param = format(model.count_params(),\",\")\nprint(f\"Total no of parameters = {total_param}\")\nviz_model(model)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:08.299819Z","iopub.execute_input":"2025-04-17T16:17:08.300114Z","iopub.status.idle":"2025-04-17T16:17:11.373111Z","shell.execute_reply.started":"2025-04-17T16:17:08.300085Z","shell.execute_reply":"2025-04-17T16:17:11.372236Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### 2.Xception Style Unet from Keras Documentation OOP Implementation \nhttps://keras.io/examples/vision/oxford_pets_image_segmentation/","metadata":{}},{"cell_type":"code","source":"class UnetXception:\n def __init__(self,\n input_shape=(256,384, 1),\n num_classes=4,\n activation='relu',\n num_filters=(32,64,128,256)):\n \n self.input_shape = input_shape\n self.num_classes = num_classes\n self.activation = activation\n self.num_filters = num_filters\n self.model = self.build()\n\n def build(self):\n inputs = tf.keras.layers.Input(self.input_shape)\n x = self.downsampling(inputs)\n x = self.upsampling(x)\n outputs = tf.keras.layers.Conv2D(self.num_classes, 1, activation=\"sigmoid\", padding=\"same\")(x)\n model = tf.keras.Model(inputs, outputs)\n return model\n\n def downsampling(self, x):\n # Entry block\n x = tf.keras.layers.Conv2D(32, 3, strides=2, padding=\"same\")(x)\n x = tf.keras.layers.BatchNormalization()(x)\n x = tf.keras.layers.Activation(\"relu\")(x)\n previous_block_activation = x\n \n # Blocks 1, 2, 3 are identical apart from the feature depth.\n for filters in self.num_filters[1:]:\n x = tf.keras.layers.Activation(\"relu\")(x)\n x = tf.keras.layers.SeparableConv2D(filters, 3, padding=\"same\")(x)\n x = tf.keras.layers.BatchNormalization()(x)\n\n x = tf.keras.layers.Activation(\"relu\")(x)\n x = tf.keras.layers.SeparableConv2D(filters, 3, padding=\"same\")(x)\n x = tf.keras.layers.BatchNormalization()(x)\n\n x = tf.keras.layers.MaxPooling2D(3, strides=2, padding=\"same\")(x)\n\n # Project residual\n residual = tf.keras.layers.Conv2D(filters, 1, strides=2, padding=\"same\")(\n previous_block_activation\n )\n x = tf.keras.layers.add([x, residual]) # Add back residual\n previous_block_activation = x # Set aside next residual\n\n return x\n\n \n def upsampling(self, x):\n num_filters = list(reversed(self.num_filters))\n previous_block_activation = x\n for filters in num_filters:\n x = tf.keras.layers.Activation(\"relu\")(x)\n x = tf.keras.layers.Conv2DTranspose(filters, 3, padding=\"same\")(x)\n x = tf.keras.layers.BatchNormalization()(x)\n\n x = tf.keras.layers.Activation(\"relu\")(x)\n x = tf.keras.layers.Conv2DTranspose(filters, 3, padding=\"same\")(x)\n x = tf.keras.layers.BatchNormalization()(x)\n\n x = tf.keras.layers.UpSampling2D(2)(x)\n\n # Project residual\n residual = tf.keras.layers.UpSampling2D(2)(previous_block_activation)\n residual = tf.keras.layers.Conv2D(filters, 1, padding=\"same\")(residual)\n x = tf.keras.layers.add([x, residual]) # Add back residual\n previous_block_activation = x # Set aside next residual\n\n return x\n\n\n def get_model(self):\n return self.model\n\n ","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:11.374291Z","iopub.execute_input":"2025-04-17T16:17:11.374518Z","iopub.status.idle":"2025-04-17T16:17:11.385514Z","shell.execute_reply.started":"2025-04-17T16:17:11.374498Z","shell.execute_reply":"2025-04-17T16:17:11.384728Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"#tf.keras.backend.clear_session()\n#UnetXception.get_model().summary()\nmodel = UnetXception().get_model()\ntotal_param = format(model.count_params(),\",\")\nprint(f\"Total no of parameters = {total_param}\")\nviz_model(model)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:11.386602Z","iopub.execute_input":"2025-04-17T16:17:11.386848Z","iopub.status.idle":"2025-04-17T16:17:13.923809Z","shell.execute_reply.started":"2025-04-17T16:17:11.386828Z","shell.execute_reply":"2025-04-17T16:17:13.922763Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### 3. Unet with ResNet50 backbone(encoder)","metadata":{}},{"cell_type":"code","source":"import tensorflow as tf\n\nclass UNetResNet50:\n def __init__(self,\n input_shape=(256,384, 3),\n num_classes=4,\n activation='relu'):\n self.input_shape = input_shape\n self.num_classes = num_classes\n self.activation = activation\n self.model = self.build()\n\n def build_encoder(self, input_tensor):\n base_model = tf.keras.applications.ResNet50(weights='imagenet', include_top=False, input_tensor=input_tensor)\n return base_model\n\n def build_decoder(self, encoder_output):\n decoder = tf.keras.layers.Conv2DTranspose(256, (2, 2), strides=(2, 2), padding='same')(encoder_output.output)\n decoder = tf.keras.layers.Concatenate()([decoder, encoder_output.layers[142].output]) # Skip connection\n decoder = tf.keras.layers.Conv2D(256, (3, 3), activation=self.activation, padding='same')(decoder)\n decoder = tf.keras.layers.Conv2D(256, (3, 3), activation=self.activation, padding='same')(decoder)\n\n decoder = tf.keras.layers.Conv2DTranspose(128, (2, 2), strides=(2, 2), padding='same')(decoder)\n decoder = tf.keras.layers.Concatenate()([decoder, encoder_output.layers[80].output]) # Skip connection\n decoder = tf.keras.layers.Conv2D(128, (3, 3), activation=self.activation, padding='same')(decoder)\n decoder = tf.keras.layers.Conv2D(128, (3, 3), activation=self.activation, padding='same')(decoder)\n\n decoder = tf.keras.layers.Conv2DTranspose(64, (2, 2), strides=(2, 2), padding='same')(decoder)\n decoder = tf.keras.layers.Concatenate()([decoder, encoder_output.layers[38].output]) # Skip connection\n decoder = tf.keras.layers.Conv2D(64, (3, 3), activation=self.activation, padding='same')(decoder)\n decoder = tf.keras.layers.Conv2D(64, (3, 3), activation=self.activation, padding='same')(decoder)\n\n decoder = tf.keras.layers.Conv2DTranspose(32, (2, 2), strides=(2, 2), padding='same')(decoder)\n decoder = tf.keras.layers.Concatenate()([decoder, encoder_output.layers[4].output]) # Skip connection\n decoder = tf.keras.layers.Conv2D(32, (3, 3), activation=self.activation, padding='same')(decoder)\n decoder = tf.keras.layers.Conv2D(32, (3, 3), activation=self.activation, padding='same')(decoder)\n\n return decoder\n\n def build(self):\n input_tensor = tf.keras.layers.Input(shape=self.input_shape)\n encoder_output = self.build_encoder(input_tensor)\n decoder_output = self.build_decoder(encoder_output)\n output = tf.keras.layers.Conv2D(self.num_classes, (1, 1), activation='sigmoid')(decoder_output)\n model = tf.keras.models.Model(inputs=input_tensor, outputs=output)\n return model\n\n \n def get_model(self):\n return self.model","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:13.936800Z","iopub.execute_input":"2025-04-17T16:17:13.937148Z","iopub.status.idle":"2025-04-17T16:17:13.952272Z","shell.execute_reply.started":"2025-04-17T16:17:13.937118Z","shell.execute_reply":"2025-04-17T16:17:13.951255Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"#tf.keras.backend.clear_session()\n#UnetXception.get_model().summary()\nmodel = UNetResNet50().get_model()\ntotal_param = format(model.count_params(),\",\")\nprint(f\"Total no of parameters = {total_param}\")\nviz_model(model)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:13.953273Z","iopub.execute_input":"2025-04-17T16:17:13.953615Z","iopub.status.idle":"2025-04-17T16:17:23.950315Z","shell.execute_reply.started":"2025-04-17T16:17:13.953553Z","shell.execute_reply":"2025-04-17T16:17:23.949452Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### 4. Unet with EfficientNetB2 backbone(encoder)","metadata":{}},{"cell_type":"code","source":"class UNetEfficientNetB2:\n def __init__(self,\n input_shape=(256,384, 3),\n num_classes=4,\n activation='relu'):\n self.input_shape = input_shape\n self.num_classes = num_classes\n self.activation = activation\n self.model = self.build_model()\n\n def build_encoder(self, input_tensor):\n base_model = tf.keras.applications.EfficientNetB2(weights='imagenet', include_top=False, input_shape=input_shape)\n return base_model\n\n def build_decoder(self, encoder_output):\n decoder = tf.keras.layers.Conv2DTranspose(256, (2, 2), strides=(2, 2), padding='same')(encoder_output.output)\n decoder = tf.keras.layers.Concatenate()([decoder, encoder_output.layers[254].output]) # Skip connection\n decoder = tf.keras.layers.Conv2D(256, (3, 3), activation=self.activation, padding='same')(decoder)\n decoder = tf.keras.layers.Conv2D(256, (3, 3), activation=self.activation, padding='same')(decoder)\n\n decoder = tf.keras.layers.Conv2DTranspose(128, (2, 2), strides=(2, 2), padding='same')(decoder)\n decoder = tf.keras.layers.Concatenate()([decoder, encoder_output.layers[122].output]) # Skip connection\n decoder = tf.keras.layers.Conv2D(128, (3, 3), activation=self.activation, padding='same')(decoder)\n decoder = tf.keras.layers.Conv2D(128, (3, 3), activation=self.activation, padding='same')(decoder)\n\n decoder = tf.keras.layers.Conv2DTranspose(64, (2, 2), strides=(2, 2), padding='same')(decoder)\n decoder = tf.keras.layers.Concatenate()([decoder, encoder_output.layers[50].output]) # Skip connection\n decoder = tf.keras.layers.Conv2D(64, (3, 3), activation=self.activation, padding='same')(decoder)\n decoder = tf.keras.layers.Conv2D(64, (3, 3), activation=self.activation, padding='same')(decoder)\n\n decoder = tf.keras.layers.Conv2DTranspose(32, (2, 2), strides=(2, 2), padding='same')(decoder)\n decoder = tf.keras.layers.Concatenate()([decoder, encoder_output.layers[7].output]) # Skip connection\n decoder = tf.keras.layers.Conv2D(32, (3, 3), activation=self.activation, padding='same')(decoder)\n decoder = tf.keras.layers.Conv2D(32, (3, 3), activation=self.activation, padding='same')(decoder)\n\n return decoder\n\n def build_model(self):\n input_tensor = tf.keras.layers.Input(shape=self.input_shape)\n encoder_output = self.build_encoder(input_tensor)\n decoder_output = self.build_decoder(encoder_output)\n output = tf.keras.layers.Conv2D(self.num_classes, (1, 1), activation='sigmoid')(decoder_output)\n model = Model(inputs=input_tensor, outputs=output)\n return model\n \n def get_model(self):\n return self.model\n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:23.951555Z","iopub.execute_input":"2025-04-17T16:17:23.951868Z","iopub.status.idle":"2025-04-17T16:17:23.964321Z","shell.execute_reply.started":"2025-04-17T16:17:23.951843Z","shell.execute_reply":"2025-04-17T16:17:23.963431Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"#tf.keras.backend.clear_session()\n#UnetXception.get_model().summary()\n#model = UNetEfficientNetB2().get_model()\n#total_param = format(model.count_params(),\",\")\n#print(f\"Total no of parameters = {total_param}\")\n#viz_model(model)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:23.966273Z","iopub.execute_input":"2025-04-17T16:17:23.966807Z","iopub.status.idle":"2025-04-17T16:17:23.979705Z","shell.execute_reply.started":"2025-04-17T16:17:23.966784Z","shell.execute_reply":"2025-04-17T16:17:23.979017Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### 5. Unet with different backbones using Segmentation-Models Library ","metadata":{}},{"cell_type":"code","source":"tf.keras.backend.clear_session()\n\nmodel = sm.Unet('efficientnetb3',\n input_shape=(256,384, 1),\n classes=4,\n activation='sigmoid',\n encoder_weights=None)\n\n\ntotal_param = format(model.count_params(),\",\")\nprint(f\"Total no of parameters = {total_param}\")\nviz_model(model)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:23.980498Z","iopub.execute_input":"2025-04-17T16:17:23.980760Z","iopub.status.idle":"2025-04-17T16:17:32.357053Z","shell.execute_reply.started":"2025-04-17T16:17:23.980739Z","shell.execute_reply":"2025-04-17T16:17:32.356151Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"tf.keras.backend.clear_session()\n\nmodel = sm.Unet('resnet50',\n input_shape=(256,384, 1),\n classes=4,\n activation='sigmoid',\n encoder_weights=None)\n\n\ntotal_param = format(model.count_params(),\",\")\nprint(f\"Total no of parameters = {total_param}\")\nviz_model(model)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:32.358153Z","iopub.execute_input":"2025-04-17T16:17:32.358402Z","iopub.status.idle":"2025-04-17T16:17:40.591392Z","shell.execute_reply.started":"2025-04-17T16:17:32.358379Z","shell.execute_reply":"2025-04-17T16:17:40.590566Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"tf.keras.backend.clear_session()\n\nmodel = sm.FPN('efficientnetb2',\n input_shape=(256,384, 1),\n classes=4,\n activation='sigmoid',\n encoder_weights=None)\n\n\ntotal_param = format(model.count_params(),\",\")\nprint(f\"Total no of parameters = {total_param}\")\nviz_model(model)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:40.592636Z","iopub.execute_input":"2025-04-17T16:17:40.592950Z","iopub.status.idle":"2025-04-17T16:17:55.387057Z","shell.execute_reply.started":"2025-04-17T16:17:40.592921Z","shell.execute_reply":"2025-04-17T16:17:55.386153Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### 6. DeepLabV3+ with self attention","metadata":{}},{"cell_type":"markdown","source":"1. https://github.com/taki0112/Self-Attention-GAN-Tensorflow/\n2. https://link.springer.com/article/10.1007/s10489-020-01917-8\n3. https://arxiv.org/abs/1905.08008","metadata":{}},{"cell_type":"code","source":"import tensorflow as tf","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:55.388098Z","iopub.execute_input":"2025-04-17T16:17:55.388326Z","iopub.status.idle":"2025-04-17T16:17:55.392463Z","shell.execute_reply.started":"2025-04-17T16:17:55.388306Z","shell.execute_reply":"2025-04-17T16:17:55.391384Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"base = tf.keras.applications.EfficientNetB0(include_top=False,\n weights=\"imagenet\",\n input_shape=(256,384, 3)\n )","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:55.393468Z","iopub.execute_input":"2025-04-17T16:17:55.393701Z","iopub.status.idle":"2025-04-17T16:17:57.632096Z","shell.execute_reply.started":"2025-04-17T16:17:55.393676Z","shell.execute_reply":"2025-04-17T16:17:57.631382Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"tf.keras.backend.clear_session()\n#base.summary()\ntotal_param = format(base.count_params(),\",\")\nprint(f\"Total no of parameters = {total_param}\")\nviz_model(model)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:17:57.633158Z","iopub.execute_input":"2025-04-17T16:17:57.633465Z","iopub.status.idle":"2025-04-17T16:18:10.062833Z","shell.execute_reply.started":"2025-04-17T16:17:57.633435Z","shell.execute_reply":"2025-04-17T16:18:10.062021Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"tf.keras.backend.clear_session()\n#tf.keras.utils.plot_model(base, show_shapes=True, rankdir='TB')","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:10.063919Z","iopub.execute_input":"2025-04-17T16:18:10.064178Z","iopub.status.idle":"2025-04-17T16:18:10.070685Z","shell.execute_reply.started":"2025-04-17T16:18:10.064156Z","shell.execute_reply":"2025-04-17T16:18:10.069874Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"base.get_layer('block7a_dwconv').output[0]\nbase.get_layer('block6d_project_conv').output[0]\nbase.get_layer('block7a_bn').output[0]\nbase.get_layer('block2a_dwconv').output[0]","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:10.071700Z","iopub.execute_input":"2025-04-17T16:18:10.072425Z","iopub.status.idle":"2025-04-17T16:18:10.102869Z","shell.execute_reply.started":"2025-04-17T16:18:10.072403Z","shell.execute_reply":"2025-04-17T16:18:10.102124Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def hw_flatten(x) :\n return tf.keras.layers.Reshape(target_shape=(-1,x.shape[-1]))(x)\n\n\ndef self_attention_cnn(x):\n q = tf.keras.layers.Conv2D(filters=x.shape[-1]//8, kernel_size=(1,1))(x) # [bs, h, w, c']\n k = tf.keras.layers.Conv2D(filters=x.shape[-1]//8, kernel_size=(1,1))(x) # [bs, h, w, c']\n v = tf.keras.layers.Conv2D(filters=x.shape[-1], kernel_size=(1,1))(x) # [bs, h, w, c]\n qk = tf.keras.layers.Dot(axes=2)([hw_flatten(q), hw_flatten(k)]) # [bs, n, n] ;n = h * w\n softmax = tf.keras.layers.Activation(\"softmax\")(qk) # [bs, n, n]\n attention = tf.keras.layers.Dot(axes=1)([softmax,hw_flatten(v)]) # [bs, N, c]\n attention = tf.keras.layers.Reshape(target_shape=tuple(x.shape[1:]))(attention) # [bs, h,w, c]\n y = tf.keras.layers.Add()([x,attention]) # [bs, h,w, c]\n return y\n\ndef refined_self_attention_cnn(x):\n q = tf.keras.layers.Conv2D(filters=x.shape[-1]//8, kernel_size=(1,1))(x) # [bs, h, w, c']\n k = tf.keras.layers.Conv2D(filters=x.shape[-1]//8, kernel_size=(1,1))(x) # [bs, h, w, c']\n v = tf.keras.layers.Conv2D(filters=x.shape[-1], kernel_size=(1,1))(x) # [bs, h, w, c]\n kv = tf.keras.layers.Dot(axes=1)([hw_flatten(k), hw_flatten(v)])\n norm = kv/hw_flatten(x).shape[1]\n norm = tf.keras.layers.Reshape((norm.shape[-1],norm.shape[1]))(norm)\n attention = tf.keras.layers.Dot(axes=2)([hw_flatten(q),norm])\n attention = tf.keras.layers.Reshape(target_shape=tuple(x.shape[1:]))(attention)\n y = tf.keras.layers.Add()([x,attention])\n return x\n\n\ndef ASPP(inputs):\n shape = inputs.shape\n\n y_pool = tf.keras.layers.AveragePooling2D(pool_size=(shape[1], shape[2]), name='average_pooling')(inputs)\n y_pool = tf.keras.layers.Conv2D(filters=256, kernel_size=1, padding='same', use_bias=False)(y_pool)\n y_pool = tf.keras.layers.BatchNormalization(name=f'bn_1')(y_pool)\n y_pool = tf.keras.layers.Activation('relu', name=f'relu_1')(y_pool)\n y_pool = tf.keras.layers.UpSampling2D((shape[1], shape[2]), interpolation=\"bilinear\")(y_pool)\n\n y_1 = tf.keras.layers.Conv2D(filters=256, kernel_size=1, dilation_rate=1, padding='same', use_bias=False)(inputs)\n y_1 = tf.keras.layers.BatchNormalization()(y_1)\n y_1 = tf.keras.layers.Activation('relu')(y_1)\n\n y_6 = tf.keras.layers.Conv2D(filters=256, kernel_size=3, dilation_rate=6, padding='same', use_bias=False)(inputs)\n y_6 = tf.keras.layers.BatchNormalization()(y_6)\n y_6 = tf.keras.layers.Activation('relu')(y_6)\n\n y_12 = tf.keras.layers.Conv2D(filters=256, kernel_size=3, dilation_rate=12, padding='same', use_bias=False)(inputs)\n y_12 = tf.keras.layers.BatchNormalization()(y_12)\n y_12 = tf.keras.layers.Activation('relu')(y_12)\n\n y_18 = tf.keras.layers.Conv2D(filters=256, kernel_size=3, dilation_rate=18, padding='same', use_bias=False)(inputs)\n y_18 = tf.keras.layers.BatchNormalization()(y_18)\n y_18 = tf.keras.layers.Activation('relu')(y_18)\n\n y = tf.keras.layers.Concatenate()([y_pool, y_1, y_6, y_12, y_18])\n\n y = tf.keras.layers.Conv2D(filters=256, kernel_size=1, dilation_rate=1, padding='same', use_bias=False)(y)\n y = tf.keras.layers.BatchNormalization()(y)\n y = tf.keras.layers.Activation('relu')(y)\n return y\n\ndef EfficientAttentionDeepLabV3Plus(num_classes,input_shape=(256,384, 3)):\n \n \"\"\" Inputs \"\"\"\n inputs = tf.keras.layers.Input(input_shape)\n \n \"\"\" Pre-trained Output \"\"\"\n backbone = tf.keras.applications.EfficientNetB0(include_top=False,weights=\"imagenet\",input_tensor=inputs)\n image_features = backbone.get_layer('block7a_dwconv').output\n x_a = ASPP(image_features)\n x_a = tf.keras.layers.UpSampling2D((8, 8), interpolation=\"bilinear\")(x_a)\n\n \n \"\"\" Get low-level features \"\"\"\n x_b = backbone.get_layer('block2a_dwconv').output\n x_b = self_attention_cnn(x_b)\n x_b = tf.keras.layers.Conv2D(filters=48, kernel_size=1, padding='same', use_bias=False)(x_b)\n x_b = tf.keras.layers.BatchNormalization()(x_b)\n x_b = tf.keras.layers.Activation('relu')(x_b)\n\n x = tf.keras.layers.Concatenate()([x_a, x_b])\n\n x = tf.keras.layers.Conv2D(filters=256, kernel_size=3, padding='same', activation='relu',use_bias=False)(x)\n x = tf.keras.layers.BatchNormalization()(x)\n x = tf.keras.layers.Activation('relu')(x)\n\n x = tf.keras.layers.Conv2D(filters=256, kernel_size=3, padding='same', activation='relu', use_bias=False)(x)\n x = tf.keras.layers.BatchNormalization()(x)\n x = tf.keras.layers.Activation('relu')(x)\n x = tf.keras.layers.UpSampling2D((4, 4), interpolation=\"bilinear\")(x)\n\n \"\"\" Outputs \"\"\"\n x = tf.keras.layers.Conv2D(num_classes, (1, 1), name='output_layer')(x)\n x = tf.keras.layers.Activation('sigmoid')(x)\n\n \"\"\" Model \"\"\"\n model = tf.keras.models.Model(inputs=inputs, outputs=x)\n return model\n \n\n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:10.104054Z","iopub.execute_input":"2025-04-17T16:18:10.104286Z","iopub.status.idle":"2025-04-17T16:18:10.122405Z","shell.execute_reply.started":"2025-04-17T16:18:10.104266Z","shell.execute_reply":"2025-04-17T16:18:10.121639Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"input_shape=(256,384, 3)\ninputs = tf.keras.layers.Input(input_shape)\nbackbone = tf.keras.applications.EfficientNetB0(include_top=False,weights=\"imagenet\",input_tensor=inputs)\n\n\nimage_features = backbone.get_layer('block7a_dwconv').output\nprint(image_features.shape)\nx_a = ASPP(image_features)\nx_a = tf.keras.layers.UpSampling2D((8, 8), interpolation=\"bilinear\")(x_a)\nprint(x_a.shape)\n\n\"\"\" Get low-level features \"\"\"\nx_b = backbone.get_layer('block2a_dwconv').output\nx_b = self_attention_cnn(x_b)\nprint(x_b.shape)\nx_b = tf.keras.layers.Conv2D(filters=48, kernel_size=1, padding='same', use_bias=False)(x_b)\nx_b = tf.keras.layers.BatchNormalization()(x_b)\nx_b = tf.keras.layers.Activation('relu')(x_b)\nprint(x_b.shape)\nx = tf.keras.layers.Concatenate()([x_a, x_b])\nprint(x.shape)\nx = tf.keras.layers.Conv2D(filters=256, kernel_size=3, padding='same', activation='relu',use_bias=False)(x)\nx = tf.keras.layers.BatchNormalization()(x)\nx = tf.keras.layers.Activation('relu')(x)\n\nx = tf.keras.layers.Conv2D(filters=256, kernel_size=3, padding='same', activation='relu', use_bias=False)(x)\nx = tf.keras.layers.BatchNormalization()(x)\nx = tf.keras.layers.Activation('relu')(x)\nprint(x.shape)\nx = tf.keras.layers.UpSampling2D((4, 4), interpolation=\"bilinear\")(x)\n\n\"\"\" Outputs \"\"\"\nx = tf.keras.layers.Conv2D(1, (1, 1), name='output_layer')(x)\nx = tf.keras.layers.Activation('sigmoid')(x)\n\n\"\"\" Model \"\"\"\nmodel = tf.keras.models.Model(inputs=inputs, outputs=x)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:10.123476Z","iopub.execute_input":"2025-04-17T16:18:10.123787Z","iopub.status.idle":"2025-04-17T16:18:12.166363Z","shell.execute_reply.started":"2025-04-17T16:18:10.123757Z","shell.execute_reply":"2025-04-17T16:18:12.165618Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"model = EfficientAttentionDeepLabV3Plus(num_classes=4,\n input_shape = (R_HEIGHT, R_WIDTH, NUM_CHANNELS))\ntotal_param = format(model.count_params(),\",\")\nprint(f\"Total no of parameters = {total_param}\")\nviz_model(model)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:12.167300Z","iopub.execute_input":"2025-04-17T16:18:12.167519Z","iopub.status.idle":"2025-04-17T16:18:29.341937Z","shell.execute_reply.started":"2025-04-17T16:18:12.167499Z","shell.execute_reply":"2025-04-17T16:18:29.341164Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"total_param = format(model.count_params(),\",\")\nprint(f\"Total no of parameters = {total_param}\")\nviz_model(model)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:29.343130Z","iopub.execute_input":"2025-04-17T16:18:29.343743Z","iopub.status.idle":"2025-04-17T16:18:44.357745Z","shell.execute_reply.started":"2025-04-17T16:18:29.343711Z","shell.execute_reply":"2025-04-17T16:18:44.356881Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class DeepLabV3Plus:\n def __init__(self,\n input_shape=(256,384, 3),\n num_classes=4,\n activation='relu'):\n self.input_shape = input_shape\n self.num_classes = num_classes\n self.activation = activation\n self.model = self.build_model()\n self.backbone = tf.keras.applications.EfficientNetB0(include_top=False,\n weights=\"imagenet\",\n input_shape=input_shape,\n )\n \n self.encoder = self.backbone.get_layer(\"block7a_dwconv\").output\n \n pass","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:44.358933Z","iopub.execute_input":"2025-04-17T16:18:44.359498Z","iopub.status.idle":"2025-04-17T16:18:44.365538Z","shell.execute_reply.started":"2025-04-17T16:18:44.359466Z","shell.execute_reply":"2025-04-17T16:18:44.364697Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Custom Loss, metric and Optimizer classes \n[Go back to the Table of Contents](#contents)
","metadata":{}},{"cell_type":"code","source":"class DiceScore(tf.keras.metrics.Metric):\n def __init__(self, name='dice_score', **kwargs):\n super(DiceScore, self).__init__(name=name, **kwargs)\n self.intersection = self.add_weight(name='intersection', initializer='zeros')\n self.union = self.add_weight(name='union', initializer='zeros')\n\n def update_state(self, y_true, y_pred, sample_weight=None):\n y_true = tf.cast(y_true, dtype=tf.float32)\n y_pred = tf.cast(y_pred, dtype=tf.float32)\n\n intersection = tf.reduce_sum(y_true * y_pred)\n union = tf.reduce_sum(y_true) + tf.reduce_sum(y_pred)\n\n self.intersection.assign_add(intersection)\n self.union.assign_add(union)\n\n def result(self):\n epsilon = tf.keras.backend.epsilon()\n epsilon = 1\n dice = (2.0 * self.intersection) / (self.union + epsilon)\n return dice\n\n def reset_state(self):\n self.intersection.assign(0.0)\n self.union.assign(0.0)\n \n\n \nclass IoUScore(tf.keras.metrics.Metric):\n def __init__(self, name='iou_score', **kwargs):\n super(IoUScore, self).__init__(name=name, **kwargs)\n self.intersection = self.add_weight(name='intersection', initializer='zeros')\n self.union = self.add_weight(name='union', initializer='zeros')\n\n def update_state(self, y_true, y_pred, sample_weight=None):\n y_true = tf.cast(y_true, dtype=tf.float32)\n y_pred = tf.cast(y_pred, dtype=tf.float32)\n\n intersection = tf.reduce_sum(y_true * y_pred)\n union = tf.reduce_sum(tf.maximum(y_true, y_pred))\n\n self.intersection.assign_add(intersection)\n self.union.assign_add(union)\n\n def result(self):\n epsilon = tf.keras.backend.epsilon()\n epsilon = 1\n iou = (self.intersection) / (self.union + epsilon)\n return iou\n\n def reset_state(self):\n self.intersection.assign(0.0)\n self.union.assign(0.0)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:44.366668Z","iopub.execute_input":"2025-04-17T16:18:44.366900Z","iopub.status.idle":"2025-04-17T16:18:44.384415Z","shell.execute_reply.started":"2025-04-17T16:18:44.366880Z","shell.execute_reply":"2025-04-17T16:18:44.383624Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"y_pred = np.array([0., 0., 1., 0.])\ny_true = np.array([[1., 1., 1., 0.]])\n\nd = IoUScore()\nd.update_state(y_true,y_pred)\nprint(d.result().numpy())\n\nd = DiceScore()\nd.update_state(y_true,y_pred)\nprint(d.result().numpy())","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:44.385530Z","iopub.execute_input":"2025-04-17T16:18:44.386272Z","iopub.status.idle":"2025-04-17T16:18:44.580762Z","shell.execute_reply.started":"2025-04-17T16:18:44.386242Z","shell.execute_reply":"2025-04-17T16:18:44.579981Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class DiceLoss(tf.keras.losses.Loss):\n def __init__(self, from_logits=False, reduction=tf.keras.losses.Reduction.AUTO, name='dice_loss'):\n super(DiceLoss, self).__init__(reduction=reduction, name=name)\n self.from_logits = from_logits\n\n def dice_loss(self, y_true, y_pred):\n epsilon = tf.keras.backend.epsilon()\n epsilon = 1\n intersection = tf.reduce_sum(y_true * y_pred)\n dice = (2.0 * intersection + epsilon) / (tf.reduce_sum(y_true) + tf.reduce_sum(y_pred) + epsilon)\n return 1.0 - dice\n\n def call(self, y_true, y_pred):\n if self.from_logits:\n y_pred = tf.nn.sigmoid(y_pred)\n dice_loss = self.dice_loss(y_true, y_pred)\n return dice_loss\n \nclass IoULoss(tf.keras.losses.Loss):\n def __init__(self, from_logits=False, reduction=tf.keras.losses.Reduction.AUTO, name='iou_loss'):\n super(IoULoss, self).__init__(reduction=reduction, name=name)\n self.from_logits = from_logits\n\n def iou_loss(self, y_true, y_pred):\n #epsilon = 2*tf.keras.backend.epsilon()\n epsilon = 1\n intersection = tf.reduce_sum(y_true * y_pred)\n union = tf.reduce_sum(tf.maximum(y_true, y_pred))\n iou = (intersection + epsilon) / (union + epsilon)\n return 1.0 - iou\n\n def call(self, y_true, y_pred):\n if self.from_logits:\n y_pred = tf.nn.sigmoid(y_pred)\n iou_loss = self.iou_loss(y_true, y_pred)\n return iou_loss\n\nclass BCELoss(tf.keras.losses.Loss):\n def __init__(self, from_logits=False, reduction=tf.keras.losses.Reduction.AUTO, name='bce_loss'):\n super(BCELoss, self).__init__(reduction=reduction, name=name)\n self.from_logits = from_logits\n \n def bce_loss(self,y_true,y_pred):\n epsilon = 2*tf.keras.backend.epsilon()\n y_pred = np.clip(y_pred, epsilon, 1.0 - epsilon) # Clip to avoid log(0) and log(1)\n bce_loss = - (y_true * np.log(y_pred) + (1 - y_true) * np.log(1 - y_pred))\n mean_bce_loss = tf.reduce_sum(bce_loss)/len(bce_loss.flatten())\n return mean_bce_loss\n\n def call(self, y_true, y_pred):\n if self.from_logits:\n y_pred = tf.nn.sigmoid(y_pred)\n bce_loss = self.bce_loss(y_true, y_pred)\n return bce_loss\n \nalpha = 0.8\ngamma = 2.0\n \nclass FocalLoss(tf.keras.losses.Loss):\n def __init__(self, from_logits=False, reduction=tf.keras.losses.Reduction.AUTO, name='focal_loss'):\n super(FocalLoss, self).__init__(reduction=reduction, name=name)\n self.from_logits = from_logits\n \n def focal_loss(self,y_true,y_pred):\n bce_loss = tf.keras.losses.BinaryCrossentropy(from_logits=self.from_logits,reduction=self.reduction)(y_true, y_pred)\n pt = np.exp(-bce_loss)\n focal_loss = np.power(alpha*(1-pt), gamma)*bce_loss\n return focal_loss\n\n def call(self, y_true, y_pred):\n if self.from_logits:\n y_pred = tf.nn.sigmoid(y_pred)\n focal_loss = self.focal_loss(y_true,y_pred)\n return focal_loss\n\n\nclass DiceBCELoss(tf.keras.losses.Loss):\n def __init__(self, from_logits=False, reduction=tf.keras.losses.Reduction.AUTO, name='dice_bce_loss'):\n super(DiceBCELoss, self).__init__(reduction=reduction, name=name)\n self.from_logits = from_logits\n\n def dice_loss(self, y_true, y_pred):\n epsilon = tf.keras.backend.epsilon()\n epsilon = 1\n intersection = tf.reduce_sum(y_true * y_pred)\n dice = (2.0 * intersection + epsilon) / (tf.reduce_sum(y_true) + tf.reduce_sum(y_pred) + epsilon)\n return 1.0 - dice\n\n def call(self, y_true, y_pred):\n if self.from_logits:\n y_pred = tf.nn.sigmoid(y_pred)\n \n bce_loss = tf.keras.losses.BinaryCrossentropy(from_logits=self.from_logits, reduction=self.reduction)(y_true, y_pred)\n dice_loss = self.dice_loss(y_true, y_pred)\n\n return dice_loss + bce_loss\n \n\nclass DiceFocalLoss(tf.keras.losses.Loss):\n def __init__(self, from_logits=False, reduction=tf.keras.losses.Reduction.AUTO, name='dice_focal_loss'):\n super(DiceFocalLoss, self).__init__(reduction=reduction, name=name)\n self.from_logits = from_logits\n\n def dice_loss(self, y_true, y_pred):\n epsilon = tf.keras.backend.epsilon()\n epsilon = 1\n intersection = tf.reduce_sum(y_true * y_pred)\n dice = (2.0 * intersection + epsilon) / (tf.reduce_sum(y_true) + tf.reduce_sum(y_pred) + epsilon)\n return 1.0 - dice\n \n def focal_loss(self,y_true, y_pred):\n bce_loss = tf.keras.losses.BinaryCrossentropy(from_logits=self.from_logits,reduction=self.reduction)(y_true, y_pred)\n pt = np.exp(-bce_loss)\n focal_loss = np.mean(np.power(alpha*(1-pt), gamma)*bce_loss)\n return focal_loss\n\n def call(self, y_true, y_pred):\n if self.from_logits:\n y_pred = tf.nn.sigmoid(y_pred)\n \n dice_loss = self.dice_loss(y_true, y_pred)\n focal_loss = self.focal_loss(y_true, y_pred)\n\n return dice_loss + focal_loss\n \n \nclass DiceFocalBCELoss(tf.keras.losses.Loss):\n def __init__(self, from_logits=False, reduction=tf.keras.losses.Reduction.AUTO, name='dice_focal_bce_loss'):\n super(DiceFocalBCELoss, self).__init__(reduction=reduction, name=name)\n self.from_logits = from_logits\n\n def dice_loss(self, y_true, y_pred):\n epsilon = tf.keras.backend.epsilon()\n epsilon = 1\n intersection = tf.reduce_sum(y_true * y_pred)\n dice = (2.0 * intersection + epsilon) / (tf.reduce_sum(y_true) + tf.reduce_sum(y_pred) + epsilon)\n return 1.0 - dice\n \n def focal_loss(self,y_true,y_pred):\n bce_loss = tf.keras.losses.BinaryCrossentropy(from_logits=self.from_logits,reduction=self.reduction)(y_true, y_pred)\n pt = np.exp(-bce_loss)\n focal_loss = np.mean(np.power(alpha*(1-pt), gamma)*bce_loss)\n return focal_loss\n\n def call(self, y_true, y_pred):\n if self.from_logits:\n y_pred = tf.nn.sigmoid(y_pred)\n \n bce_loss = tf.keras.losses.BinaryCrossentropy(from_logits=self.from_logits, reduction=self.reduction)(y_true, y_pred)\n dice_loss = self.dice_loss(y_true, y_pred)\n focal_loss = self.focal_loss(y_true, y_pred)\n\n return dice_loss + focal_loss+ bce_loss\n \n \n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:44.582156Z","iopub.execute_input":"2025-04-17T16:18:44.582489Z","iopub.status.idle":"2025-04-17T16:18:44.604286Z","shell.execute_reply.started":"2025-04-17T16:18:44.582458Z","shell.execute_reply":"2025-04-17T16:18:44.603420Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"y_pred = np.array([0., 0., 1., 0.])\ny_true = np.array([[1., 1., 1., 0.]])\n\ncriterion = BCELoss()\nloss = criterion.call(y_true,y_pred)\nprint(f\"{criterion.name} = {loss.numpy()}\")\n\ncriterion = FocalLoss()\nloss = criterion.call(y_true,y_pred)\nprint(f\"{criterion.name} = {loss.numpy()}\")\n\n\ncriterion = DiceLoss()\nloss = criterion.call(y_true,y_pred)\nprint(f\"{criterion.name} = {loss.numpy()}\")\n\ncriterion = IoULoss()\nloss = criterion.call(y_true,y_pred)\nprint(f\"{criterion.name} = {loss.numpy()}\")\n\n\ncriterion = DiceBCELoss()\nloss = criterion.call(y_true,y_pred)\nprint(f\"{criterion.name} = {loss.numpy()}\")\n\n\ncriterion = DiceFocalLoss()\nloss = criterion.call(y_true,y_pred)\nprint(f\"{criterion.name} = {loss.numpy()}\")\n\ncriterion = DiceFocalBCELoss()\nloss = criterion.call(y_true,y_pred)\nprint(f\"{criterion.name} = {loss.numpy()}\")","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:44.605456Z","iopub.execute_input":"2025-04-17T16:18:44.605789Z","iopub.status.idle":"2025-04-17T16:18:44.685610Z","shell.execute_reply.started":"2025-04-17T16:18:44.605759Z","shell.execute_reply":"2025-04-17T16:18:44.684714Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import wandb\nfrom wandb.keras import WandbCallback,WandbModelCheckpoint\nfrom kaggle_secrets import UserSecretsClient\n\nwandb_api_key = user_secrets.get_secret(\"wandb_api_key\")\nwandb.login(key=wandb_api_key)\n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T17:26:08.646549Z","iopub.execute_input":"2025-04-17T17:26:08.646978Z","iopub.status.idle":"2025-04-17T17:26:12.531983Z","shell.execute_reply.started":"2025-04-17T17:26:08.646946Z","shell.execute_reply":"2025-04-17T17:26:12.530971Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"config = dict(\n input_shape = (R_HEIGHT, R_WIDTH, NUM_CHANNELS),\n batch_size = BATCH_SIZE,\n optimizer = tf.keras.optimizers.Adam(),\n loss = DiceBCELoss(),\n metrics = [DiceScore(), IoUScore()],\n epochs = 15,\n)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T18:35:00.554486Z","iopub.execute_input":"2025-04-17T18:35:00.555079Z","iopub.status.idle":"2025-04-17T18:35:00.565894Z","shell.execute_reply.started":"2025-04-17T18:35:00.555048Z","shell.execute_reply":"2025-04-17T18:35:00.565053Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"run = wandb.init(entity='nazish63663-bahauddin-zakariya-university',\n project = 'cloud_formation_segmentation',\n name = 'test_1e_Effiecientnetb2AttentionDeeplabV3Plus',\n config = config,\n mode=\"offline\",\n save_code = True,\n )","metadata":{"execution":{"iopub.status.busy":"2025-04-17T17:43:33.052346Z","iopub.execute_input":"2025-04-17T17:43:33.053075Z","iopub.status.idle":"2025-04-17T17:44:04.163759Z","shell.execute_reply.started":"2025-04-17T17:43:33.053043Z","shell.execute_reply":"2025-04-17T17:44:04.163045Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Defining Callbacks \n[Go back to the Table of Contents](#contents)
","metadata":{}},{"cell_type":"code","source":"work_dir = \"/kaggle/working/\"\nmodel_name = \"cloud_1e_Effiecientnetb2AttentionDeeplabV3Plus\"\n\n\nwandb_callback = WandbCallback(save_model=False)\nwandb_checkpoint = WandbModelCheckpoint(filepath=\"weights-improvment-{epoch:02d}-{val_dice_score:.4f}.h5\",\n monitor='val_dice_score',\n mode=\"max\",\n save_best_only=True,\n save_freq=\"epoch\",\n period=5,)\n\nreduce_lr = tf.keras.callbacks.ReduceLROnPlateau(monitor=\"val_dice_score\",\n mode=\"max\",\n factor=0.5,\n min_lr=1e-6,\n patience=1,\n verbose=0)\n\nearly_stop = tf.keras.callbacks.EarlyStopping(monitor=\"val_dice_score\",\n mode=\"max\",\n patience=3,\n verbose=0,\n restore_best_weights=True)\n\n\nckpt_filepath = os.path.join(work_dir,model_name,\"ckpts\",\"weights-improvement-{epoch:02d}-{val_loss:.4f}.hdf5\")\n\nmodel_checkpoint = tf.keras.callbacks.ModelCheckpoint(ckpt_filepath,\n monitor=\"val_loss\",\n mode=\"min\",\n save_best_only=False,\n save_freq=\"epoch\",\n period=2,\n )\n\ncsv_filepath = os.path.join(work_dir,model_name,\"logs\",\"training.csv\")\ncsv_log = tf.keras.callbacks.CSVLogger(csv_filepath)\n\ncallbacks = [reduce_lr,early_stop,wandb_callback, wandb_checkpoint]","metadata":{"execution":{"iopub.status.busy":"2025-04-17T17:54:06.740099Z","iopub.execute_input":"2025-04-17T17:54:06.740796Z","iopub.status.idle":"2025-04-17T17:55:06.760054Z","shell.execute_reply.started":"2025-04-17T17:54:06.740765Z","shell.execute_reply":"2025-04-17T17:55:06.759369Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Training the Model \n[Go back to the Table of Contents](#contents)
","metadata":{}},{"cell_type":"code","source":"tf.keras.backend.clear_session()\n\nmodel = sm.Unet('efficientnetb2',\n input_shape=config[\"input_shape\"],\n classes=4,\n activation='sigmoid',\n encoder_weights='imagenet')\n\n\nmodel.compile(optimizer=config['optimizer'], loss=config['loss'], metrics=config['metrics'])\n\ntotal_param = format(model.count_params(),\",\")\nprint(f\"Total no of parameters = {total_param}\")\nprint(model.name)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.151328Z","iopub.status.idle":"2025-04-17T16:18:46.151761Z","shell.execute_reply.started":"2025-04-17T16:18:46.151531Z","shell.execute_reply":"2025-04-17T16:18:46.151550Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"tf.keras.backend.clear_session()\n\nmodel = EfficientAttentionDeepLabV3Plus(num_classes=4,\n input_shape = (R_HEIGHT, R_WIDTH, NUM_CHANNELS))\n\n\nmodel.compile(optimizer=config['optimizer'], loss=config['loss'], metrics=config['metrics'])\n\n\ntotal_param = format(model.count_params(),\",\")\nprint(f\"Total no of parameters = {total_param}\")\n#viz_model(model)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T18:02:05.534206Z","iopub.execute_input":"2025-04-17T18:02:05.534966Z","iopub.status.idle":"2025-04-17T18:02:07.785061Z","shell.execute_reply.started":"2025-04-17T18:02:05.534934Z","shell.execute_reply":"2025-04-17T18:02:07.784191Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Test everything before training the model\nprint(model.input_shape,'\\t',train_generator.image_shape,'\\t',val_generator.image_shape)\nprint(\"GPU available!\" if tf.test.is_gpu_available() else \"GPU is not available\")\nprint(model.optimizer)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T18:03:09.611662Z","iopub.execute_input":"2025-04-17T18:03:09.612307Z","iopub.status.idle":"2025-04-17T18:03:09.621545Z","shell.execute_reply.started":"2025-04-17T18:03:09.612276Z","shell.execute_reply":"2025-04-17T18:03:09.620651Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Run this line if you used dropout layer or segmentation_model library\n# tf.config.optimizer.set_experimental_options({\"layout_optimizer\": False}) ","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.156018Z","iopub.status.idle":"2025-04-17T16:18:46.156278Z","shell.execute_reply.started":"2025-04-17T16:18:46.156151Z","shell.execute_reply":"2025-04-17T16:18:46.156163Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import multiprocessing\nnum_cores = multiprocessing.cpu_count()\nprint(num_cores)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T18:03:27.684805Z","iopub.execute_input":"2025-04-17T18:03:27.685136Z","iopub.status.idle":"2025-04-17T18:03:27.690463Z","shell.execute_reply.started":"2025-04-17T18:03:27.685113Z","shell.execute_reply":"2025-04-17T18:03:27.689635Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"start = time.perf_counter()\n\nhistory = model.fit(train_generator,\n validation_data=val_generator,\n epochs=config['epochs'],\n callbacks=callbacks,\n workers=num_cores,\n use_multiprocessing=False\n )\n\nend = time.perf_counter()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-04-17T18:03:44.110563Z","iopub.execute_input":"2025-04-17T18:03:44.111403Z","iopub.status.idle":"2025-04-17T18:14:42.511686Z","shell.execute_reply.started":"2025-04-17T18:03:44.111372Z","shell.execute_reply":"2025-04-17T18:14:42.510956Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"**Error** - `E tensorflow/core/grappler/optimizers/meta_optimizer.cc:954] layout failed: INVALID_ARGUMENT: Size of values 0 does not match size of permutation 4 @ fanin shape inmodel_1/dropout_9/dropout/SelectV2-2-TransposeNHWCToNCHW-LayoutOptimizer`\n\n**Reason** - `Dropout layer`\n\n**Solution** - `1. tf.config.optimizer.set_experimental_options({\"layout_optimizer\": False})`\n `2. Remove dropout layer`","metadata":{"execution":{"iopub.status.busy":"2023-06-10T12:27:08.580742Z","iopub.execute_input":"2023-06-10T12:27:08.58113Z","iopub.status.idle":"2023-06-10T12:27:08.586615Z","shell.execute_reply.started":"2023-06-10T12:27:08.581098Z","shell.execute_reply":"2023-06-10T12:27:08.585686Z"}}},{"cell_type":"code","source":"print(f\"Total training time = {(end - start)/3600:.2f} hours\")\n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T18:26:09.546639Z","iopub.execute_input":"2025-04-17T18:26:09.547763Z","iopub.status.idle":"2025-04-17T18:26:09.553126Z","shell.execute_reply.started":"2025-04-17T18:26:09.547728Z","shell.execute_reply":"2025-04-17T18:26:09.552122Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def plot_accuracy_loss(history):\n f,ax = plt.subplots(1,3,figsize=(16,4))\n ax = ax.ravel()\n \n ax[0].plot([None]+history.history['loss'],'o-')\n ax[0].plot([None]+history.history['val_loss'],'o-')\n ax[0].legend(['Train Loss','Validation Loss'],loc = 0)\n ax[0].set_title('Training & Validation Loss')\n ax[0].set_xlabel('Epoch')\n ax[0].set_ylabel('Loss')\n ax[0].grid(True)\n\n ax[1].plot([None]+history.history['dice_score'],'o-')\n ax[1].plot([None]+history.history['val_dice_score'],'o-')\n ax[1].legend(['Training Dice','Validation Dice'],loc = 0)\n ax[1].set_title('Training & Validation Dice Score')\n ax[1].set_xlabel('Epoch')\n ax[1].set_ylabel('Dice Score')\n ax[1].grid(True)\n \n ax[2].plot([None]+history.history['iou_score'],'o-')\n ax[2].plot([None]+history.history['val_iou_score'],'o-')\n ax[2].legend(['Train IOU','Validation IOU'],loc = 0)\n ax[2].set_title('Training & Validation IOU Score')\n ax[2].set_xlabel('Epoch')\n ax[2].set_ylabel('IOU Score')\n ax[2].grid(True)\n \n #plt.style.use('ggplot')\n plt.tight_layout()\n plt.show()\n \n return None\n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T18:30:42.594525Z","iopub.execute_input":"2025-04-17T18:30:42.594948Z","iopub.status.idle":"2025-04-17T18:30:42.603870Z","shell.execute_reply.started":"2025-04-17T18:30:42.594917Z","shell.execute_reply":"2025-04-17T18:30:42.602753Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"plot_accuracy_loss(history)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T18:35:12.649456Z","iopub.execute_input":"2025-04-17T18:35:12.650247Z","iopub.status.idle":"2025-04-17T18:35:13.292773Z","shell.execute_reply.started":"2025-04-17T18:35:12.650213Z","shell.execute_reply":"2025-04-17T18:35:13.291730Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Saving and Downloading the trained Model \n[Go back to the Table of Contents](#contents)
","metadata":{}},{"cell_type":"code","source":"work_dir = \"/kaggle/working/\"\nfilepath = os.path.join(work_dir,model_name,f\"{model_name}.h5\")\nmodel_path = model.save(filepath,save_format=\"h5\")\nwandb.save(filepath)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T18:32:41.584202Z","iopub.execute_input":"2025-04-17T18:32:41.585074Z","iopub.status.idle":"2025-04-17T18:32:41.748493Z","shell.execute_reply.started":"2025-04-17T18:32:41.585039Z","shell.execute_reply":"2025-04-17T18:32:41.747321Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"wandb.finish()","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.168195Z","iopub.status.idle":"2025-04-17T16:18:46.168612Z","shell.execute_reply.started":"2025-04-17T16:18:46.168386Z","shell.execute_reply":"2025-04-17T16:18:46.168405Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nimport subprocess\nfrom IPython.display import FileLink, display\n\ndef download_file(path, download_file_name):\n os.chdir('/kaggle/working/')\n zip_name = f\"/kaggle/working/{download_file_name}.zip\"\n command = f\"zip {zip_name} {path} -r\"\n result = subprocess.run(command, shell=True, capture_output=True, text=True)\n if result.returncode != 0:\n print(\"Unable to run zip command!\")\n print(result.stderr)\n return\n display(FileLink(f'{download_file_name}.zip'))\n","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.170243Z","iopub.status.idle":"2025-04-17T16:18:46.170664Z","shell.execute_reply.started":"2025-04-17T16:18:46.170433Z","shell.execute_reply":"2025-04-17T16:18:46.170453Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"download_file(filepath, \"cloud_model_20_effiUnet\")","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.172280Z","iopub.status.idle":"2025-04-17T16:18:46.172711Z","shell.execute_reply.started":"2025-04-17T16:18:46.172469Z","shell.execute_reply":"2025-04-17T16:18:46.172489Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Evaluate model on Test Set and make submission \n[Go back to the Table of Contents](#contents)
","metadata":{}},{"cell_type":"code","source":"work_dir = \"/kaggle/working/\"\nfilepath = os.path.join(work_dir,f\"{model_name}.h5\")\nfile = wandb.restore(f\"{model_name}.h5\",run_path='taki/cloud_formation_segmentation/runs/gcdsx2sl')\nmodel = tf.keras.saving.load_model(filepath,compile=False)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.173791Z","iopub.status.idle":"2025-04-17T16:18:46.174197Z","shell.execute_reply.started":"2025-04-17T16:18:46.173990Z","shell.execute_reply":"2025-04-17T16:18:46.174010Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"`Classwise best threshold and minsize params: {0: (0.5, 25000), 1: (0.6, 21000), 2: (0.3, 20000), 3: (0.5, 10000)}` `Collected from` https://www.kaggle.com/code/artgor/classification-in-catalyst-with-utility-scripts#Post-processing","metadata":{}},{"cell_type":"code","source":"#minsizes = [20000 ,20000, 22500, 10000]\nminsizes = [25000 ,21000, 2000, 10000]\nsigmoid = lambda x: 1 / (1 + np.exp(-x))\nthresholds = [0.5, 0.6, 0.3, 0.5]\n\ndef mask2rle(img):\n '''\n Convert mask to rle.\n img: numpy array, 1 - mask, 0 - background\n Returns run length as string formated\n '''\n pixels= img.T.flatten()\n pixels = np.concatenate([[0], pixels, [0]])\n runs = np.where(pixels[1:] != pixels[:-1])[0] + 1\n runs[1::2] -= runs[::2]\n return ' '.join(str(x) for x in runs)\n\ndef post_process(probability, threshold, min_size):\n \"\"\"\n Post processing of each predicted mask, components with lesser number of pixels\n than `min_size` are ignored\n \"\"\"\n \n mask = cv2.threshold(probability, threshold, 1, cv2.THRESH_BINARY)[1]\n \n num_component, component = cv2.connectedComponents(mask.astype(np.uint8))\n predictions = np.zeros((350, 525), np.float32)\n num = 0\n for c in range(1, num_component):\n p = (component == c)\n if p.sum() > min_size:\n predictions[p] = 1\n num += 1\n return predictions, num\n ","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.176548Z","iopub.status.idle":"2025-04-17T16:18:46.176991Z","shell.execute_reply.started":"2025-04-17T16:18:46.176768Z","shell.execute_reply":"2025-04-17T16:18:46.176788Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"https://www.kaggle.com/code/ratthachat/cloud-convexhull-polygon-postprocessing-no-gpu?scriptVersionId=20977692","metadata":{}},{"cell_type":"code","source":"def draw_convex_hull(mask, mode='approx'):\n \n img = np.zeros(mask.shape)\n contours, hier = cv2.findContours(mask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)\n \n for c in contours:\n if mode=='rect': # simple rectangle\n x, y, w, h = cv2.boundingRect(c)\n cv2.rectangle(img, (x, y), (x+w, y+h), (255, 255, 255), -1)\n elif mode=='convex': # minimum convex hull\n hull = cv2.convexHull(c)\n cv2.drawContours(img, [hull], 0, (255, 255, 255),-1)\n elif mode=='approx':\n epsilon = 0.02*cv2.arcLength(c,True)\n approx = cv2.approxPolyDP(c,epsilon,True)\n cv2.drawContours(img, [approx], 0, (255, 255, 255),-1)\n else: # minimum area rectangle\n rect = cv2.minAreaRect(c)\n box = cv2.boxPoints(rect)\n box = np.int0(box)\n cv2.drawContours(img, [box], 0, (255, 255, 255),-1)\n return img/255.","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.178680Z","iopub.status.idle":"2025-04-17T16:18:46.179090Z","shell.execute_reply.started":"2025-04-17T16:18:46.178879Z","shell.execute_reply":"2025-04-17T16:18:46.178899Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"batch_images = test_generator.__getitem__(0)\nlow = index * batch_Size\nhigh = min(low + batch_Size, total_samples)\nbatch_indexes = list(range(low,high))\nbatch_indexes_sub = range(low*4, high*4)\nbatch_size = batch_images.shape[0]\nbatch_predicted_masks = model.predict(batch_images)\nbatch_predicted_masks = batch_predicted_masks.round()\nbatch_predicted_masks[0]\nresized_batch_predicted_masks = np.zeros((32, 350, 525, 4))\nfor i in range(batch_predicted_masks.shape[0]):\n resized_batch_predicted_masks[i] = cv2.resize(batch_predicted_masks[i], (525,350))\nresized_batch_predicted_masks = resized_batch_predicted_masks.astype(np.int64)\n#print(resized_batch_predicted_masks[0]) ","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.180225Z","iopub.status.idle":"2025-04-17T16:18:46.180512Z","shell.execute_reply.started":"2025-04-17T16:18:46.180377Z","shell.execute_reply":"2025-04-17T16:18:46.180391Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"sub_df = pd.read_csv(\"/kaggle/input/understanding_cloud_organization/sample_submission.csv\")\n\ntotal_samples = test_generator.total_samples\nbatch_Size = test_generator.batch_size\nno_batches = test_generator.__len__() \n\nfor index in range(no_batches):\n batch_images = test_generator.__getitem__(index)\n low = index * batch_Size\n high = min(low + batch_Size, total_samples)\n batch_indexes = list(range(low,high))\n batch_indexes_sub = range(low*4, high*4)\n batch_size = batch_images.shape[0]\n batch_predicted_masks = model.predict(batch_images) \n batch_predicted_masks = batch_predicted_masks.round()\n\n resized_batch_predicted_masks = np.zeros((32, 350, 525, 4))\n for i in range(batch_predicted_masks.shape[0]):\n resized_batch_predicted_masks[i] = cv2.resize(batch_predicted_masks[i], (525,350))\n #resized_batch_predicted_masks = resized_batch_predicted_masks.astype(np.int32)\n \n for k in batch_indexes_sub:\n m = k%32\n if m<32:\n mask, num_predict = post_process(resized_batch_predicted_masks[m][:,:,m%4], thresholds[m%4], minsizes[m%4])\n mask = draw_convex_hull(mask.astype(np.uint8))\n rle = mask2rle(mask)\n sub_df.loc[k][\"EncodedPixels\"] = rle\n ","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.181871Z","iopub.status.idle":"2025-04-17T16:18:46.182153Z","shell.execute_reply.started":"2025-04-17T16:18:46.182020Z","shell.execute_reply":"2025-04-17T16:18:46.182034Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"batch_indexes_sub","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.183711Z","iopub.status.idle":"2025-04-17T16:18:46.184011Z","shell.execute_reply.started":"2025-04-17T16:18:46.183874Z","shell.execute_reply":"2025-04-17T16:18:46.183888Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"sub_df.head(30)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.185003Z","iopub.status.idle":"2025-04-17T16:18:46.185254Z","shell.execute_reply.started":"2025-04-17T16:18:46.185131Z","shell.execute_reply":"2025-04-17T16:18:46.185143Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"sub_df.loc[10]","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.186491Z","iopub.status.idle":"2025-04-17T16:18:46.186913Z","shell.execute_reply.started":"2025-04-17T16:18:46.186700Z","shell.execute_reply":"2025-04-17T16:18:46.186720Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"sub_df.to_csv('submission.csv', index=False)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.188261Z","iopub.status.idle":"2025-04-17T16:18:46.188675Z","shell.execute_reply.started":"2025-04-17T16:18:46.188449Z","shell.execute_reply":"2025-04-17T16:18:46.188468Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Exploring Predicticted Mask on Validation Set \n[Go back to the Table of Contents](#contents)
","metadata":{}},{"cell_type":"code","source":"def batch_post_process(batch_pred_masks,\n thresholds=[0.5, 0.6, 0.3, 0.5],\n minsizes=[20000 ,20000, 22500, 10000]):\n \n batch_processed_masks = np.zeros(batch_pred_masks.shape)\n for k in range(batch_pred_masks.shape[0]):\n for i in range(4):\n probability = batch_pred_masks[k][:,:,i]\n min_size = minsizes[i]\n threshold = thresholds[i]\n\n mask = cv2.threshold(probability, threshold, 1, cv2.THRESH_BINARY)[1]\n num_component, component = cv2.connectedComponents(mask.astype(np.uint8))\n predictions = np.zeros((R_HEIGHT, R_WIDTH), np.float32)\n num = 0\n for c in range(1, num_component):\n p = (component == c)\n if p.sum() > min_size:\n predictions[p] = 1\n num += 1\n mask = predictions\n mask = draw_convex_hull(mask.astype(np.uint8))\n batch_processed_masks[k][:,:,i] = mask.astype(np.uint8)\n return batch_processed_masks","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.189807Z","iopub.status.idle":"2025-04-17T16:18:46.190213Z","shell.execute_reply.started":"2025-04-17T16:18:46.190006Z","shell.execute_reply":"2025-04-17T16:18:46.190025Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"val_generator = DataGenerator(dataframe=df_val,\n root_dir=train_image_path,\n mode=\"fit\",\n resize=True,\n resize_width=R_WIDTH,\n resize_height=R_HEIGHT,\n num_channels=NUM_CHANNELS,\n batch_size=32,\n augmentations=None)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.191952Z","iopub.status.idle":"2025-04-17T16:18:46.192210Z","shell.execute_reply.started":"2025-04-17T16:18:46.192085Z","shell.execute_reply":"2025-04-17T16:18:46.192098Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"index = 2\nbatch_images,batch_masks = val_generator.__getitem__(index)\nbatch_predicted_masks = model.predict(batch_images)\nbatch_predicted_masks = batch_predicted_masks.round()\nprocessed_batch_predicted_masks = batch_post_process(batch_predicted_masks)\nprint(batch_images.shape)\nprint(batch_masks.shape)\nprint(batch_predicted_masks.shape)\nprint(processed_batch_predicted_masks.shape)","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.193407Z","iopub.status.idle":"2025-04-17T16:18:46.193725Z","shell.execute_reply.started":"2025-04-17T16:18:46.193553Z","shell.execute_reply":"2025-04-17T16:18:46.193567Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"for img,masks,pred_masks in zip(batch_images[3:6],batch_masks[3:6],processed_batch_predicted_masks[3:6]):\n masks = (masks[:,:,0], masks[:,:,1],masks[:,:,2],masks[:,:,3])\n pred_masks = (pred_masks[:,:,0], pred_masks[:,:,1],pred_masks[:,:,2],pred_masks[:,:,3])\n print(\"Image, Masks and Predicted Masks\")\n show_img_with_masks(img,masks,comment=\"(ground truth)\")\n show_img_with_masks(img,pred_masks,comment=\"(predicted)\")\n ","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.194975Z","iopub.status.idle":"2025-04-17T16:18:46.195228Z","shell.execute_reply.started":"2025-04-17T16:18:46.195104Z","shell.execute_reply":"2025-04-17T16:18:46.195116Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"## Something is wrong in post processing. Couldn't figure out! probably sigmoid function","metadata":{"execution":{"iopub.status.busy":"2025-04-17T16:18:46.196385Z","iopub.status.idle":"2025-04-17T16:18:46.196709Z","shell.execute_reply.started":"2025-04-17T16:18:46.196536Z","shell.execute_reply":"2025-04-17T16:18:46.196550Z"},"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Acknowledgements \n[Go back to the Table of Contents](#contents)
","metadata":{}},{"cell_type":"markdown","source":"* [Satellite Clouds: U-Net with ResNet Encoder](https://www.kaggle.com/code/xhlulu/satellite-clouds-u-net-with-resnet-encoder)\n* [Cloud: ConvexHull& Polygon PostProcessing (No GPU)](https://www.kaggle.com/code/ratthachat/cloud-convexhull-polygon-postprocessing-no-gpu?scriptVersionId=20977692)\n* [Jupyter Notebook Tricks](https://www.kaggle.com/code/tientd95/jupyter-notebook-tricks)","metadata":{}},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}