{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"none","dataSources":[{"sourceId":91249,"databundleVersionId":11294684,"sourceType":"competition"}],"dockerImageVersionId":30918,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"import torch\nimport random","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2025-03-20T13:42:22.094830Z","iopub.execute_input":"2025-03-20T13:42:22.095211Z","iopub.status.idle":"2025-03-20T13:42:22.099568Z","shell.execute_reply.started":"2025-03-20T13:42:22.095184Z","shell.execute_reply":"2025-03-20T13:42:22.098385Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def random_crop_patch_around_target(volume, target_coord, patch_size):\n \"\"\"\n Randomly crop a 3D patch ensuring the target is inside.\n\n Args:\n volume (torch.Tensor): Input volume of shape (C, D_full, H_full, W_full).\n target_coord (tuple): (z, y, x) coordinates of the object in the full volume.\n patch_size (tuple): (D, H, W) size of the patch.\n\n Returns:\n cropped_patch (torch.Tensor): Cropped volume of shape (C, D, H, W).\n crop_start (tuple): (z_start, y_start, x_start) of the cropped patch.\n \"\"\"\n C, D_full, H_full, W_full = volume.shape\n D, H, W = patch_size\n z_target, y_target, x_target = target_coord\n\n # Define valid range for cropping\n z_min = max(0, z_target - D) \n z_max = min(D_full - D, z_target) \n y_min = max(0, y_target - H) \n y_max = min(H_full - H, y_target) \n x_min = max(0, x_target - W) \n x_max = min(W_full - W, x_target) \n\n # Randomly select a valid crop start position\n z_start = random.randint(z_min, z_max)\n y_start = random.randint(y_min, y_max)\n x_start = random.randint(x_min, x_max)\n\n # Crop the patch\n cropped_patch = volume[:, z_start:z_start + D, y_start:y_start + H, x_start:x_start + W]\n\n return cropped_patch, (z_start, y_start, x_start)\n \n\ndef generate_3d_labels(global_target, roi_start, roi_size, stride=2):\n \"\"\"\n Generate class_map and offset_map for 3D object detection.\n\n Args:\n global_target (tuple): (z, y, x) coordinate in the full volume.\n roi_start (tuple): (z_start, y_start, x_start) of the extracted patch in full volume.\n roi_size (tuple): (depth, height, width) of the extracted patch.\n class_map_size (tuple): (depth, height, width) of the class_map output.\n stride (int): The stride factor, default is 2.\n\n Returns:\n class_map (torch.Tensor): Binary tensor of shape class_map_size.\n offset_map (torch.Tensor): Offset tensor of shape (3, *class_map_size).\n \"\"\"\n #roi coordinate space\n z_local = global_target[0] - roi_start[0]\n y_local = global_target[1] - roi_start[1]\n x_local = global_target[2] - roi_start[2]\n\n #shrinkage coordinate\n z_shrink = z_local//stride\n y_shrink = y_local//stride\n x_shrink = x_local//stride\n \n #relative coordinate \n z_ratio = z_local/roi_size[0]\n y_ratio = y_local/roi_size[1]\n x_ratio = x_local/roi_size[2]\n\n #\n D, H, W = roi_size[0]//2, roi_size[1]//2, roi_size[2]//2\n class_map = torch.zeros(1, D, H, W)\n offset_map = torch.zeros(3, D, H, W)\n\n for dz in range(2):\n for dy in range(2):\n for dx in range(2):\n z_idx = z_shrink + dz\n y_idx = y_shrink + dy\n x_idx = x_shrink + dx\n if z_idx