{"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":"gpu","dataSources":[{"sourceId":36363,"databundleVersionId":4050810,"sourceType":"competition"},{"sourceId":6403207,"sourceType":"datasetVersion","datasetId":3692048},{"sourceId":6665364,"sourceType":"datasetVersion","datasetId":3846147},{"sourceId":7229250,"sourceType":"datasetVersion","datasetId":4185447},{"sourceId":7659622,"sourceType":"datasetVersion","datasetId":3607309},{"sourceId":8258878,"sourceType":"datasetVersion","datasetId":4898906}],"dockerImageVersionId":30559,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"!pip install --upgrade pip --quiet\n!pip install nibabel --quiet\n!pip install torch torchvision torchaudio --quiet","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:53:09.947763Z","iopub.execute_input":"2024-04-29T04:53:09.948028Z","iopub.status.idle":"2024-04-29T04:53:57.212327Z","shell.execute_reply.started":"2024-04-29T04:53:09.948002Z","shell.execute_reply":"2024-04-29T04:53:57.211091Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import os\nimport pandas as pd\nimport numpy as np\nimport nibabel as nib\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torch.utils.data import Dataset, DataLoader\nfrom torchvision import transforms\nfrom sklearn.metrics import (accuracy_score, precision_recall_fscore_support, roc_auc_score)\nfrom sklearn.model_selection import train_test_split\nfrom torchvision import models\nfrom scipy.ndimage import zoom\nimport torch.nn.functional as F\nfrom PIL import Image\n\n# Constants and configuration settings\nsegmentation_dir = '/kaggle/input/rsna-2022-cervical-spine-fracture-detection/segmentations'\ncsv_file = '/kaggle/input/file-mask-path/train_file_mask_path.csv'\nbatch_size = 4\nnum_workers = 4\nnum_classes = 8\ndesired_shape = (128, 128, 128)\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")","metadata":{"_kg_hide-output":true,"execution":{"iopub.status.busy":"2024-04-29T04:53:57.214757Z","iopub.execute_input":"2024-04-29T04:53:57.215162Z","iopub.status.idle":"2024-04-29T04:54:00.05242Z","shell.execute_reply.started":"2024-04-29T04:53:57.215117Z","shell.execute_reply":"2024-04-29T04:54:00.051421Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Specify the path to your CSV file containing data\ncsv_file_path = '/kaggle/input/file-mask-path/train_file_mask_path.csv'\n\n# Load the CSV data using pandas\ndata_frame = pd.read_csv(csv_file_path)\n\n# Extract relevant information from the data\nimage_paths = data_frame['file_path'].values\nmask_paths = data_frame['mask_path'].values\nlabels = data_frame[['C1', 'C2', 'C3', 'C4', 'C5', 'C6', 'C7', 'patient_overall']].values\n","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:00.053884Z","iopub.execute_input":"2024-04-29T04:54:00.054353Z","iopub.status.idle":"2024-04-29T04:54:00.105797Z","shell.execute_reply.started":"2024-04-29T04:54:00.054326Z","shell.execute_reply":"2024-04-29T04:54:00.10499Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import nibabel as nib\nimport pandas as pd\nimport torch\nfrom torch.utils.data import Dataset\n\nclass CustomDataset(Dataset):\n    def __init__(self, image_paths, mask_paths, labels, transform=None):\n        self.image_paths = image_paths\n        self.mask_paths = mask_paths\n        self.labels = labels\n        self.transform = transform\n\n    def __len__(self):\n        return len(self.image_paths)\n\n    def __getitem__(self, idx):\n        image_path = self.image_paths[idx]\n        mask_path = self.mask_paths[idx]\n\n        image = nib.load(image_path).get_fdata()\n        segmentation_mask = None\n\n        if pd.notna(mask_path):\n            segmentation_mask = nib.load(mask_path)\n            segmentation_mask_data = segmentation_mask.get_fdata()\n            resized_data = resize_nifti(segmentation_mask_data, desired_shape)\n            segmentation_mask_data_affine = segmentation_mask.affine\n            resized_affine = segmentation_mask_data_affine\n            segmentation_mask = nib.Nifti1Image(resized_data, affine=resized_affine).get_fdata()\n\n        if self.transform:\n            image = self.transform(image)\n\n        if segmentation_mask is not None and self.transform:\n            \n            segmentation_mask = self.transform(segmentation_mask)\n        else:\n            segmentation_mask = torch.zeros_like(image)\n\n        label = torch.tensor(self.labels[idx], dtype=torch.float32)\n        \n        return image, segmentation_mask, label\n","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:00.107777Z","iopub.execute_input":"2024-04-29T04:54:00.108048Z","iopub.status.idle":"2024-04-29T04:54:00.117792Z","shell.execute_reply.started":"2024-04-29T04:54:00.108024Z","shell.execute_reply":"2024-04-29T04:54:00.11686Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Create an instance of the custom dataset\ndataset = CustomDataset(image_paths, mask_paths, labels, transform=None)\n\n# Define the index of the sample you want to access\nsample_index = 0  # You can change this to any index you're interested in\n\n# Access the image and mask paths for the specific sample\nimage_path = dataset.image_paths[sample_index]\nmask_path = dataset.mask_paths[sample_index]","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:00.118952Z","iopub.execute_input":"2024-04-29T04:54:00.119299Z","iopub.status.idle":"2024-04-29T04:54:00.128924Z","shell.execute_reply.started":"2024-04-29T04:54:00.119274Z","shell.execute_reply":"2024-04-29T04:54:00.127967Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import torch\nimport torch.nn as nn\nimport torchvision.models as models\nfrom scipy.ndimage import zoom\n# Function to resize NIfTI data\ndef resize_nifti(nifti_data, target_shape):\n    factors = (target_shape[0] / nifti_data.shape[0],\n               target_shape[1] / nifti_data.shape[1],\n               target_shape[2] / nifti_data.shape[2])\n    resized_data = zoom(nifti_data, factors, order=3)  # Cubic interpolation (higher quality)\n    return resized_data\n","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:00.130111Z","iopub.execute_input":"2024-04-29T04:54:00.130475Z","iopub.status.idle":"2024-04-29T04:54:00.137077Z","shell.execute_reply.started":"2024-04-29T04:54:00.130444Z","shell.execute_reply":"2024-04-29T04:54:00.136056Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torchvision import models\nfrom torchvision.models.video import swin3d_b, Swin3D_B_Weights\n\nclass MultiLabel3DAttentionModel(nn.Module):\n    def __init__(self, num_classes, num_classes_segmentation):\n        super(MultiLabel3DAttentionModel, self).__init__()\n\n        # Load a pre-trained ResNet3D backbone\n        self.backbone = models.video.swin3d_b(pretrained=True)\n\n        # Attention block\n        self.attention = nn.Sequential(\n            nn.Conv3d(1, 128, kernel_size=1),\n            nn.ReLU(inplace=True),\n            nn.Conv3d(128, 1, kernel_size=1),\n            nn.Sigmoid()\n        )\n        \n        # Classification head\n        self.classification_head = nn.Sequential(\n            nn.AdaptiveAvgPool3d(1),\n            nn.Flatten(),\n            nn.Linear(1, 64),\n            nn.ReLU(inplace=True),\n            nn.Linear(64, num_classes),\n            nn.Sigmoid()\n        )\n        \n        # Segmentation head\n        self.segmentation_head = nn.Sequential(\n            nn.Conv3d(1, 128, kernel_size=1),\n            nn.ReLU(inplace=True),\n            nn.Conv3d(128, num_classes_segmentation, kernel_size=1),\n            nn.Sigmoid()\n        )\n        \n    def forward(self, x, segmentation_mask):\n        # Feature extraction with the backbone\n#         print(x.shape, \"........................\")\n        transforms = Swin3D_B_Weights.KINETICS400_V1.transforms()\n        #x = transforms(torch.rand(4, 16, 3, 128, 128))\n        x = transforms(x)\n#         print(x.shape,\"...........after transformation...........\")\n        features = self.backbone(x)\n        \n#         print(\"features shape = \",features.shape)\n\n        # Apply attention to features\n        features = features.view(features.size(0), 1, 1, 1, features.size(1))\n        attention_weights = self.attention(features)\n        attended_features = features * attention_weights\n\n        # Classification branch\n        classification_output = self.classification_head(attended_features)\n        \n        # Initialize segmentation_output as None\n        segmentation_output = None\n\n        # Check if segmentation_mask is None\n        if segmentation_mask is not None:\n            \n            # Segmentation branch\n            segmentation_output = self.segmentation_head(attended_features)\n            segmentation_output = F.interpolate(segmentation_output, size=segmentation_mask.shape[2:], mode='trilinear')\n            segmentation_output = segmentation_output * segmentation_mask\n            \n        else:\n            # No segmentation branch in the test phase\n            segmentation_mask = torch.zeros_like(x)\n\n        return classification_output, segmentation_output","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:00.138544Z","iopub.execute_input":"2024-04-29T04:54:00.13892Z","iopub.status.idle":"2024-04-29T04:54:00.152096Z","shell.execute_reply.started":"2024-04-29T04:54:00.138888Z","shell.execute_reply":"2024-04-29T04:54:00.151231Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"transform = transforms.Compose([\n    transforms.ToTensor(),  # Convert to tensor\n    # Add more transformations if necessary\n])","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:00.153086Z","iopub.execute_input":"2024-04-29T04:54:00.153366Z","iopub.status.idle":"2024-04-29T04:54:00.162679Z","shell.execute_reply.started":"2024-04-29T04:54:00.153343Z","shell.execute_reply":"2024-04-29T04:54:00.162002Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import pandas as pd\n\n# Load the CSV file\ndata = pd.read_csv(csv_file)\ndata_length = len(data)\nprint(\"Length of DataFrame:\", data_length)\n\n# Remove leading and trailing whitespaces from column names\ndata.columns = data.columns.str.strip()\n\n# Split the data into training, validation, and test sets\n# train_data, temp_data = train_test_split(data, test_size=0.3, random_state=42)\n# val_data, test_data = train_test_split(temp_data, test_size=0.5, random_state=42)\n# train_data.to_csv('exp_train.csv', index=False)\n# test_data.to_csv('exp_test.csv', index=False)\n# val_data.to_csv('exp_val.csv', index=False)\ntrain_data = pd.read_csv('/kaggle/input/data-divided/exp_train.csv')\ntest_data = pd.read_csv('/kaggle/input/data-divided/exp_test.csv')\nval_data = pd.read_csv('/kaggle/input/data-divided/exp_val.csv')\n\n# Limit the number of samples for testing purposes\n# train_data = train_data[:12]\n# val_data = val_data[:12]\n# test_data = test_data[:100]\n\n# Extract file paths and labels from the data\ntrain_paths = train_data['file_path'].values\ntrain_mask_paths = train_data['mask_path'].values\ntrain_labels = train_data[['C1', 'C2', 'C3', 'C4', 'C5', 'C6', 'C7', 'patient_overall']].values\n\nval_paths = val_data['file_path'].values\nval_mask_paths = val_data['mask_path'].values\nval_labels = val_data[['C1', 'C2', 'C3', 'C4', 'C5', 'C6', 'C7', 'patient_overall']].values\n\ntest_paths = test_data['file_path'].values\ntest_mask_paths = test_data['mask_path'].values\ntest_labels = test_data[['C1', 'C2', 'C3', 'C4', 'C5', 'C6', 'C7', 'patient_overall']].values\n\nprint(\"Length of train_data:\", len(train_data))\nprint(\"Length of val_data:\", len(val_data))\nprint(\"Length of test_data:\", len(test_data))\n\n","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:00.163785Z","iopub.execute_input":"2024-04-29T04:54:00.164073Z","iopub.status.idle":"2024-04-29T04:54:00.219329Z","shell.execute_reply.started":"2024-04-29T04:54:00.16404Z","shell.execute_reply":"2024-04-29T04:54:00.218359Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Instantiate the datasets\ntrain_dataset = CustomDataset(train_paths, train_mask_paths, train_labels, transform=transform)\nval_dataset = CustomDataset(val_paths, val_mask_paths, val_labels, transform=transform)\ntest_dataset = CustomDataset(test_paths, test_mask_paths, test_labels, transform=transform)","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:00.222171Z","iopub.execute_input":"2024-04-29T04:54:00.222471Z","iopub.status.idle":"2024-04-29T04:54:00.227382Z","shell.execute_reply.started":"2024-04-29T04:54:00.222447Z","shell.execute_reply":"2024-04-29T04:54:00.226327Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Instantiate the data loaders\ntrain_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers, drop_last=True)\nval_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers)\ntest_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers)","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:00.228449Z","iopub.execute_input":"2024-04-29T04:54:00.228722Z","iopub.status.idle":"2024-04-29T04:54:00.236329Z","shell.execute_reply.started":"2024-04-29T04:54:00.228698Z","shell.execute_reply":"2024-04-29T04:54:00.235348Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"\n# Instantiate the model with the appropriate number of classes for both classification and segmentation\nnum_classes_classification = 8  # Number of classes for classification\nnum_classes_segmentation = 1    # Number of classes for segmentation (change this according to your task)\nmodel = MultiLabel3DAttentionModel(num_classes_classification, num_classes_segmentation)\nprint(model)","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:00.237498Z","iopub.execute_input":"2024-04-29T04:54:00.237783Z","iopub.status.idle":"2024-04-29T04:54:04.944915Z","shell.execute_reply.started":"2024-04-29T04:54:00.23776Z","shell.execute_reply":"2024-04-29T04:54:04.943856Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from tqdm import tqdm\nimport pickle \nimport random","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:04.946041Z","iopub.execute_input":"2024-04-29T04:54:04.946334Z","iopub.status.idle":"2024-04-29T04:54:04.950392Z","shell.execute_reply.started":"2024-04-29T04:54:04.946309Z","shell.execute_reply":"2024-04-29T04:54:04.949525Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import torch\n\ndef focal_loss(predicted, label, weights, gamma=2, epsilon=1e-7):\n    # Calculate the probability of the positive class (pt) for each sample and class\n    pt = torch.where(label == 1, predicted, 1 - predicted)\n    \n    # Calculate the loss components for each class\n    loss = -weights * ((1 - pt) ** gamma) * torch.log(pt + epsilon)\n\n    # Sum the loss components for each class\n    final_loss = torch.sum(loss, dim=1) / torch.sum(weights, dim=1)\n    \n    return final_loss.mean()\n","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:04.951466Z","iopub.execute_input":"2024-04-29T04:54:04.951725Z","iopub.status.idle":"2024-04-29T04:54:04.965676Z","shell.execute_reply.started":"2024-04-29T04:54:04.951703Z","shell.execute_reply":"2024-04-29T04:54:04.96493Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import torch\n\ndef binary_cross_entropy_loss(segmentation_outputs, batch_segmentation_masks):\n    # Ensure that both batch_segmentation_masks and segmentation_outputs are tensors of type torch.float32\n    batch_segmentation_masks = batch_segmentation_masks.to(torch.float32)\n    segmentation_outputs = segmentation_outputs.to(torch.float32)\n\n    # Create an instance of BCELoss\n    criterion = nn.BCELoss()\n\n    # Compute the BCE loss\n    loss = criterion(segmentation_outputs, batch_segmentation_masks)\n    \n    # Clamp the loss to ensure it's within the range [0, 1]\n    loss = torch.clamp(loss, min=0, max=1)\n\n    return loss\n\nimport torch\n\ndef weighted_cross_entropy(predicted, label, weights):\n    num_samples = predicted.shape[0]\n    \n    # Calculate element-wise losses\n    losses = weights * (-label * torch.log(predicted) - (1 - label) * torch.log(1 - predicted))\n    \n    # Sum the total_loss for all inputs and divide by the sum of weights\n    total_loss = torch.sum(losses, dim=1) / torch.sum(weights, dim=1)\n    \n    # Sum the total_loss for all inputs and divide by the number of samples\n    final_loss = torch.sum(total_loss) / num_samples\n    \n    # Clamp the final_loss to ensure it's within the range [0, 1]\n    final_loss = torch.clamp(final_loss, min=0, max=1)\n\n    return final_loss","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:04.966626Z","iopub.execute_input":"2024-04-29T04:54:04.966919Z","iopub.status.idle":"2024-04-29T04:54:04.976093Z","shell.execute_reply.started":"2024-04-29T04:54:04.966896Z","shell.execute_reply":"2024-04-29T04:54:04.97523Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import torch\nimport torch.nn as nn\nimport torch.optim as optim\nimport pickle\nfrom tqdm import tqdm\n\n# Function to calculate weights based on class labels\ndef weight_calculate(class_labels):\n    weights = []\n    total_weight = 0\n    count = 1\n\n    for label in class_labels:\n        if count % 8 == 0:\n            if label == 0:\n                weights.append(7)\n                total_weight += 7\n            else:\n                weights.append(14)\n                total_weight += 14\n        else:\n            if label == 0:\n                weights.append(1)\n                total_weight += 1\n            else:\n                weights.append(2)\n                total_weight += 2\n        count += 1\n\n    return torch.tensor(weights, dtype=torch.float32)\n\ndef cervical_spine_fracture_detection(num_epochs=10, initial_epoch=0, ran_once=False, model_path='sample.pth', history_path='sample_history.pkl'):\n\n    if ran_once:\n        # Load the existing model if ran_once is True\n        model = MultiLabel3DAttentionModel(num_classes_classification, num_classes_segmentation)\n        model.load_state_dict(torch.load(model_path))\n        model = model.to(device)\n\n        # Load the existing metrics history for plotting\n        with open(history_path, 'rb') as f:\n            metrics_history = pickle.load(f)\n\n        val_loss_history = metrics_history['val_loss_history']\n        val_acc_history = metrics_history['val_acc_history']\n        train_loss_history = metrics_history['train_loss_history']\n        train_acc_history = metrics_history['train_acc_history']\n    else:\n        # Create a new model if ran_once is False\n        model = MultiLabel3DAttentionModel(num_classes_classification, num_classes_segmentation)\n        model.to(device)\n\n        # Initialize empty lists for metrics history\n        val_loss_history = []\n        val_acc_history = []\n        train_loss_history = []\n        train_acc_history = []\n        \n#     total_start=time.time()\n    optimizer = optim.Adam(model.parameters(), lr=0.001)\n    best_metric = -1\n    best_metric_epoch = -1\n    best_metrics_epochs_and_time = [[], [], []]\n    \n    for epoch in range(num_epochs):\n        model.train()\n        running_loss = 0.0\n        correct_train = 0\n        total_train = 0\n\n        for batch_images, batch_segmentation_masks, batch_labels in tqdm(val_loader, desc=f\"Epoch {epoch+initial_epoch+1}/{num_epochs+initial_epoch} Validation:\"):\n            optimizer.zero_grad()\n\n            # Move data to the GPU if available\n            batch_images = batch_images.to(torch.float32).to(device)\n            batch_segmentation_masks = batch_segmentation_masks.to(torch.float32).to(device)\n            batch_labels = batch_labels.to(torch.float32).to(device)\n\n            # Assuming batch_images has shape (batch_size, num_frames, num_channels, height, width)\n            batch_images = batch_images.unsqueeze(1)  # Add a singleton dimension for channels\n            batch_segmentation_masks = batch_segmentation_masks.unsqueeze(1)\n\n            num_frames_to_add = 16\n\n            tensor_list = [batch_images] * num_frames_to_add\n            # Concatenate along the frames dimension (dimension 1)\n            nbatch_images = torch.cat(tensor_list, dim=1)\n            nbatch_segmentation_masks = torch.cat(tensor_list, dim=1)\n\n            newbatch_images = nbatch_images[:, :, :3, ...].contiguous()\n            newbatch_segmentation_masks = nbatch_segmentation_masks[:, :, :3, ...].contiguous()\n            # Check the current shape\n#                 print(\"Current shape:\", newbatch_images.shape)\n#                 print(\"Current shape:\", newbatch_segmentation_masks.shape)\n\n\n            # Forward pass\n            classification_outputs, segmentation_outputs = model(newbatch_images, newbatch_segmentation_masks)\n\n            # Apply sigmoid activation to the classification outputs\n            classification_outputs = torch.sigmoid(classification_outputs)\n\n            # Calculate weights for each sample based on class labels\n            weights = torch.stack([weight_calculate(labels) for labels in batch_labels]).to(device)\n\n            # Calculate binary cross-entropy loss with weights\n            weighted_loss = weighted_cross_entropy(classification_outputs, batch_labels, weights)\n\n            # Check if segmentation mask is available\n            if newbatch_segmentation_masks is not None and (newbatch_segmentation_masks != 0).any():\n                # Ensure that both input and target tensors are of type torch.float32\n                newbatch_segmentation_masks = newbatch_segmentation_masks.to(torch.float32)\n\n                # Apply sigmoid activation to segmentation_outputs\n                segmentation_outputs = torch.sigmoid(segmentation_outputs)\n                segmentation_outputs = segmentation_outputs.to(torch.float32)\n\n                # Calculate segmentation loss\n                segmentation_loss = binary_cross_entropy_loss(segmentation_outputs, newbatch_segmentation_masks)\n\n                # Add segmentation loss to the weighted loss\n                weighted_loss += segmentation_loss\n\n            running_loss += weighted_loss.item()\n\n            # Calculate accuracy for each class separately\n            accuracies = []\n            for class_index in range(num_classes_classification):\n                class_labels = batch_labels[:, class_index]\n                class_outputs = classification_outputs[:, class_index]\n\n                # Calculate binary predictions based on a threshold (e.g., 0.5)\n                predicted = (class_outputs > 0.5).float()\n\n                class_accuracy = accuracy_score(class_labels.cpu(), predicted.cpu())\n                accuracies.append(class_accuracy)\n\n            # Calculate overall accuracy\n            batch_accuracy = sum(accuracies) /(num_classes_classification)\n            correct_train += batch_accuracy\n            total_train += 1\n\n            # Backpropagation and optimization\n            weighted_loss.backward()\n            optimizer.step()\n\n        # Calculate and print average training accuracy and loss\n        avg_train_accuracy = correct_train / total_train\n        avg_train_loss = running_loss / len(train_loader)\n        train_acc_history.append(avg_train_accuracy)\n        train_loss_history.append(avg_train_loss)\n\n        print(f\"Epoch [{epoch + initial_epoch + 1}/{num_epochs + initial_epoch}]\")\n        print(f\"Train Accuracy: {avg_train_accuracy:.4f} | Train Loss: {avg_train_loss:.4f}\")\n\n        # Validation loop\n        model.eval()\n        total_val_loss = 0.0\n        correct_val = 0\n        total_val = 0\n\n        with torch.no_grad():\n            for batch_images, batch_segmentation_masks, batch_labels in tqdm(val_loader, desc=f\"Epoch {epoch+initial_epoch+1}/{num_epochs+initial_epoch} Validation:\"):\n                optimizer.zero_grad()\n\n                # Move data to the GPU if available\n                batch_images = batch_images.to(torch.float32).to(device)\n                batch_segmentation_masks = batch_segmentation_masks.to(torch.float32).to(device)\n                batch_labels = batch_labels.to(torch.float32).to(device)\n\n                # Assuming batch_images has shape (batch_size, num_frames, num_channels, height, width)\n                batch_images = batch_images.unsqueeze(1)  # Add a singleton dimension for channels\n                batch_segmentation_masks = batch_segmentation_masks.unsqueeze(1)\n                \n                num_frames_to_add = 16\n            \n                tensor_list = [batch_images] * num_frames_to_add\n                # Concatenate along the frames dimension (dimension 1)\n                nbatch_images = torch.cat(tensor_list, dim=1)\n                nbatch_segmentation_masks = torch.cat(tensor_list, dim=1)\n\n                newbatch_images = nbatch_images[:, :, :3, ...].contiguous()\n                newbatch_segmentation_masks = nbatch_segmentation_masks[:, :, :3, ...].contiguous()\n                # Check the current shape\n#                 print(\"Current shape:\", newbatch_images.shape)\n#                 print(\"Current shape:\", newbatch_segmentation_masks.shape)\n\n\n                # Forward pass\n                classification_outputs, segmentation_outputs = model(newbatch_images, newbatch_segmentation_masks)\n                \n                # Apply sigmoid activation to the classification outputs\n                classification_outputs = torch.sigmoid(classification_outputs)\n\n                # Calculate weights for each sample based on class labels\n                weights = torch.stack([weight_calculate(labels) for labels in batch_labels]).to(device)\n\n                # Calculate binary cross-entropy loss with weights\n                weighted_loss = weighted_cross_entropy(classification_outputs, batch_labels, weights)\n\n                # Check if segmentation mask is available\n                if newbatch_segmentation_masks is not None and (newbatch_segmentation_masks != 0).any():\n                    # Ensure that both input and target tensors are of type torch.float32\n                    newbatch_segmentation_masks = newbatch_segmentation_masks.to(torch.float32)\n\n                    # Apply sigmoid activation to segmentation_outputs\n                    segmentation_outputs = torch.sigmoid(segmentation_outputs)\n                    segmentation_outputs = segmentation_outputs.to(torch.float32)\n\n                    # Calculate segmentation loss\n                    segmentation_loss = binary_cross_entropy_loss(segmentation_outputs, newbatch_segmentation_masks)\n\n                    # Add segmentation loss to the weighted loss\n                    weighted_loss += segmentation_loss\n\n                total_val_loss += weighted_loss.item()\n\n                # Calculate accuracy for each class separately\n                accuracies = []\n                for class_index in range(num_classes_classification):\n                    class_labels = batch_labels[:, class_index]\n                    class_outputs = classification_outputs[:, class_index]\n\n                    # Calculate binary predictions based on a threshold (e.g., 0.5)\n                    predicted = (class_outputs > 0.5).float()\n\n                    class_accuracy = accuracy_score(class_labels.cpu(), predicted.cpu())\n                    accuracies.append(class_accuracy)\n\n                batch_accuracy = sum(accuracies) /(num_classes_classification)\n                correct_val += batch_accuracy\n                total_val += 1\n\n        val_accuracy = correct_val / total_val\n        avg_val_loss = total_val_loss / len(val_loader)\n        val_loss_history.append(avg_val_loss)\n        val_acc_history.append(val_accuracy)\n\n        print(f\"Epoch [{epoch + 1 + initial_epoch}/{num_epochs + initial_epoch}]\")\n        print(f\"Validation Accuracy: {val_accuracy:.4f} | Validation Loss: {avg_val_loss:.4f}\")\n        \n        if val_accuracy > best_metric:\n            best_metric = val_accuracy\n            best_metric_epoch = epoch + 1\n#             best_metrics_epochs_and_time[0].append(best_metric)\n#             best_metrics_epochs_and_time[1].append(best_metric_epoch)\n#             best_metrics_epochs_and_time[2].append(time.time() - total_start)\n            torch.save(model.state_dict(), f\"best_model_SwinTrans_wcel{epoch + 1 + initial_epoch}.pth\")\n        else:\n        # Check if validation loss has improved\n        # Reset the count since there was an improvement\n            if (epoch + 1 + initial_epoch) % 5 == 0:\n                torch.save(model.state_dict(), f\"SwinTrans_wcel{epoch + 1 + initial_epoch}.pth\")\n\n            # Save the lists of metrics to a file for later plotting\n        metrics_history = {\n            'val_loss_history': val_loss_history,\n            'val_acc_history': val_acc_history,\n            'train_loss_history': train_loss_history,\n            'train_acc_history': train_acc_history,\n        }\n\n        with open(f\"SwinTrans_wcel_metrics_history.pkl\", 'wb') as f:\n            # This will be a single file, containing all the history\n            pickle.dump(metrics_history, f)","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:04.977302Z","iopub.execute_input":"2024-04-29T04:54:04.977595Z","iopub.status.idle":"2024-04-29T04:54:05.014051Z","shell.execute_reply.started":"2024-04-29T04:54:04.977572Z","shell.execute_reply":"2024-04-29T04:54:05.013076Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# cervical_spine_fracture_detection(num_epochs=30, initial_epoch=0, ran_once=False)","metadata":{"execution":{"iopub.status.busy":"2024-04-27T11:11:12.863871Z","iopub.execute_input":"2024-04-27T11:11:12.864537Z","iopub.status.idle":"2024-04-27T11:13:12.876882Z","shell.execute_reply.started":"2024-04-27T11:11:12.864503Z","shell.execute_reply":"2024-04-27T11:13:12.875368Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"cervical_spine_fracture_detection(num_epochs=50, initial_epoch=100, ran_once=True, model_path='/kaggle/input/swintrans-after-100eps/sample_100.pth',history_path='/kaggle/input/swintrans-after-100eps/sample_metrics_history.pkl')","metadata":{"execution":{"iopub.status.busy":"2024-04-27T09:46:54.733508Z","iopub.execute_input":"2024-04-27T09:46:54.733907Z","iopub.status.idle":"2024-04-27T09:49:50.165786Z","shell.execute_reply.started":"2024-04-27T09:46:54.733876Z","shell.execute_reply":"2024-04-27T09:49:50.164715Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# cervical_spine_fracture_detection(num_epochs=10, initial_epoch=10, ran_once=True, model_path='/kaggle/input/adam-wcel-after-20-epoch/adam_wcel20.pth',history_path='/kaggle/input/adam-wcel-after-20-epoch/adam_wcel_metrics_history.pkl')","metadata":{"execution":{"iopub.status.busy":"2024-04-27T09:22:26.207679Z","iopub.execute_input":"2024-04-27T09:22:26.208046Z","iopub.status.idle":"2024-04-27T09:22:26.212941Z","shell.execute_reply.started":"2024-04-27T09:22:26.208019Z","shell.execute_reply":"2024-04-27T09:22:26.211924Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# cervical_spine_fracture_detection(num_epochs=10, initial_epoch=30, ran_once=True, model_path='/kaggle/input/adam-wcel-after-30-epoch/adam_wcel30.pth',history_path='/kaggle/input/adam-wcel-after-30-epoch/adam_wcel_metrics_history.pkl')","metadata":{"execution":{"iopub.status.busy":"2024-04-27T09:22:26.875489Z","iopub.execute_input":"2024-04-27T09:22:26.875855Z","iopub.status.idle":"2024-04-27T09:22:26.880427Z","shell.execute_reply.started":"2024-04-27T09:22:26.875828Z","shell.execute_reply":"2024-04-27T09:22:26.879132Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# cervical_spine_fracture_detection(num_epochs=10, initial_epoch=200, ran_once=True, model_path='/kaggle/input/adam-wcl-100-130/adam_wcel200.pth',history_path='/kaggle/input/adam-wcl-100-130/adam_wcel_metrics_history.pkl')","metadata":{"execution":{"iopub.status.busy":"2024-04-27T09:24:26.496183Z","iopub.execute_input":"2024-04-27T09:24:26.496563Z","iopub.status.idle":"2024-04-27T09:24:26.501028Z","shell.execute_reply.started":"2024-04-27T09:24:26.496532Z","shell.execute_reply":"2024-04-27T09:24:26.500024Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import pickle\nimport matplotlib.pyplot as plt\n\n# history_path = '/kaggle/input/train-run-1/metrics_history.pkl'\nhistory_path = '/kaggle/input/swintrans-after29epoch/SwinTrans_wcel_metrics_history.pkl'\n\n# Load the metrics history from the saved file\nwith open(history_path, 'rb') as f:\n    metrics_history = pickle.load(f)\n\n# Extract the lists of metrics\ntrain_accuracies = metrics_history['train_acc_history']\nval_accuracies = metrics_history['val_acc_history']\ntrain_losses = metrics_history['train_loss_history']\nval_losses = metrics_history['val_loss_history']\n\n# Create a list of epoch numbers for the x-axis\nepochs = list(range(1, len(train_accuracies) + 1))\n\n# Plot training and validation accuracies\nplt.figure(figsize=(12, 5))\nplt.subplot(1, 2, 1)\nplt.plot(epochs, train_accuracies, label='Train Accuracy', marker='o')\nplt.plot(epochs, val_accuracies, label='Validation Accuracy', marker='o')\nplt.xlabel('Epoch')\nplt.ylabel('Accuracy')\nplt.title('Training and Validation Accuracies')\nplt.legend()\n\n# Plot training and validation losses\nplt.subplot(1, 2, 2)\nplt.plot(epochs, train_losses, label='Train Loss', marker='o')\nplt.plot(epochs, val_losses, label='Validation Loss', marker='o')\nplt.xlabel('Epoch')\nplt.ylabel('Loss')\nplt.title('Training and Validation Losses')\nplt.legend()\n\nplt.tight_layout()\nplt.show()\n","metadata":{"execution":{"iopub.status.busy":"2024-04-29T04:54:38.838586Z","iopub.execute_input":"2024-04-29T04:54:38.839331Z","iopub.status.idle":"2024-04-29T04:54:39.517734Z","shell.execute_reply.started":"2024-04-29T04:54:38.839302Z","shell.execute_reply":"2024-04-29T04:54:39.516867Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from tqdm import tqdm\nimport torch\nimport pandas as pd\nfrom torch.utils.data import DataLoader\nfrom sklearn.metrics import accuracy_score\n\n# Define your test data loader here\ntest_data = pd.read_csv('/kaggle/input/data-divided/exp_test.csv')\n# Placeholder for DataLoader and model creation to suit your actual use case\ntest_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers)\n\n# Instantiate your model\nmodel = MultiLabel3DAttentionModel(num_classes_classification, num_classes_segmentation)\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nmodel.to(device)\n\n# Load the saved model checkpoint\nmodel.load_state_dict(torch.load('/kaggle/input/swintrans-after-100eps/sample_100.pth'))\n\n# Switch to evaluation mode\nmodel.eval()\n\n# Initialize variables for calculating test loss and accuracy\ntest_loss = 0.0\ncorrect_test = 0\ntotal_test = 0\nall_labels = []\nall_predictions = []\n\nwith torch.no_grad():\n    for batch_images, batch_segmentation_masks, batch_labels in tqdm(test_loader, desc=\"Evaluating\", leave=False):\n        # Move data to the GPU if available\n        batch_images = batch_images.to(device, dtype=torch.float32)\n        batch_labels = batch_labels.to(device, dtype=torch.float32)\n\n        # Assuming batch_images has shape (batch_size, num_frames, num_channels, height, width)\n        batch_images = batch_images.unsqueeze(1)\n        num_frames_to_add = 16\n            \n        tensor_list = [batch_images] * num_frames_to_add\n        # Concatenate along the frames dimension (dimension 1)\n        nbatch_images = torch.cat(tensor_list, dim=1)\n        nbatch_segmentation_masks = torch.cat(tensor_list, dim=1)\n\n        newbatch_images = nbatch_images[:, :, :3, ...].contiguous()\n        newbatch_segmentation_masks = nbatch_segmentation_masks[:, :, :3, ...].contiguous()# Add a singleton dimension for channels\n\n        # Forward pass\n        classification_outputs, _ = model(newbatch_images, None)  # No need for segmentation in the test phase\n\n        # Apply sigmoid activation to the classification outputs\n        classification_outputs = torch.sigmoid(classification_outputs)\n        weights = torch.stack([weight_calculate(labels) for labels in batch_labels]).to(device)\n\n        # Calculate test loss using your defined criterion (focal_loss)\n        weighted_loss = weighted_cross_entropy(classification_outputs, batch_labels, weights)\n        test_loss += weighted_loss.item()\n\n        # Calculate accuracy for each class separately (similar to training)\n        accuracies = []\n        for class_index in range(num_classes_classification):\n            class_labels = batch_labels[:, class_index]\n            class_outputs = classification_outputs[:, class_index]\n\n            # Calculate binary predictions based on a threshold (e.g., 0.5)\n            predicted = (class_outputs > 0.5).float()\n\n            class_accuracy = accuracy_score(class_labels.cpu(), predicted.cpu())\n            accuracies.append(class_accuracy)\n\n        # Calculate overall accuracy for this batch\n        batch_accuracy = sum(accuracies) / num_classes_classification\n        correct_test += batch_accuracy\n        total_test += 1\n\n# Calculate and print average test accuracy and loss\navg_test_accuracy = correct_test / total_test\navg_test_loss = test_loss / len(test_loader)\nprint(\"Test Accuracy: {:.4f} | Test Loss: {:.4f}\".format(avg_test_accuracy, avg_test_loss))\n","metadata":{"execution":{"iopub.status.busy":"2024-04-29T05:21:22.211448Z","iopub.execute_input":"2024-04-29T05:21:22.212356Z","iopub.status.idle":"2024-04-29T05:22:30.336151Z","shell.execute_reply.started":"2024-04-29T05:21:22.212322Z","shell.execute_reply":"2024-04-29T05:22:30.335109Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from tqdm import tqdm\nimport torch\nimport pandas as pd\nimport numpy as np\n\n# Define your test data loader here\ntest_data = pd.read_csv('/kaggle/input/data-divided/exp_test.csv')\n\ntest_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers)\n\n# Instantiate your model\nmodel = MultiLabel3DAttentionModel((num_classes_classification), num_classes_segmentation)\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nmodel.to(device)\n\n# Load the saved model checkpoint\nmodel.load_state_dict(torch.load('/kaggle/input/swintrans-after-100eps/sample_100.pth'))\n\n# Switch to evaluation mode\nmodel.eval()\n\nall_actual_labels = []\nall_predicted_labels = []\n\nwith torch.no_grad():\n    for batch_images, batch_segmentation_masks, batch_labels in tqdm(test_loader):\n        # Move data to the GPU if available\n        batch_images = batch_images.to(device, dtype=torch.float32)\n        batch_labels = batch_labels.to(device, dtype=torch.float32)\n\n        # Assuming batch_images has shape (batch_size, num_frames, num_channels, height, width)\n        batch_images = batch_images.unsqueeze(1)\n        num_frames_to_add = 16\n\n        tensor_list = [batch_images] * num_frames_to_add\n        # Concatenate along the frames dimension (dimension 1)\n        nbatch_images = torch.cat(tensor_list, dim=1)\n        nbatch_segmentation_masks = torch.cat(tensor_list, dim=1)\n\n        newbatch_images = nbatch_images[:, :, :3, ...].contiguous()\n        newbatch_segmentation_masks = nbatch_segmentation_masks[:, :, :3, ...].contiguous()# Add a singleton dimension for channels\n\n        # Forward pass\n        classification_outputs, _ = model(newbatch_images, None)  # No need for segmentation in the test phase\n\n        # Apply sigmoid activation to the classification outputs\n        classification_outputs = torch.sigmoid(classification_outputs)\n\n        # Append actual and predicted labels for this batch\n        all_actual_labels.append(batch_labels.cpu().numpy())\n#         print(all_actual_labels)\n        all_predicted_labels.append(classification_outputs.cpu().numpy())\n#         print(all_predicted_labels)\n\n# Concatenate all actual and predicted labels\nall_actual_labels = np.concatenate(all_actual_labels, axis=0)\n# print(all_actual_labels)\nall_predicted_labels = np.concatenate(all_predicted_labels, axis=0)\n# print(all_predicted_labels)\n\n\n# Now, you can use all_actual_labels and all_predicted_labels to calculate various metrics:\n# e.g., precision, recall, F1-score, hamming loss, etc.\nfrom sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score\n\n# Assuming you have all_actual_labels and all_predicted_labels as numpy arrays\n\n# Calculate accuracy\n# Calculate precision, recall, and F1 score\n\nprecision = precision_score(all_actual_labels, (all_predicted_labels > 0.5).astype(int), average='micro')\nrecall = recall_score(all_actual_labels, (all_predicted_labels > 0.5).astype(int), average='micro')\nf1 = f1_score(all_actual_labels, (all_predicted_labels > 0.5).astype(int), average='micro')\n\nprint(\"Precision: {:.4f}\".format(precision))\nprint(\"Recall: {:.4f}\".format(recall))\nprint(\"F1 Score: {:.4f}\".format(f1))","metadata":{"execution":{"iopub.status.busy":"2024-04-29T05:19:06.570583Z","iopub.execute_input":"2024-04-29T05:19:06.570976Z","iopub.status.idle":"2024-04-29T05:20:20.950761Z","shell.execute_reply.started":"2024-04-29T05:19:06.570948Z","shell.execute_reply":"2024-04-29T05:20:20.949505Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}