Asslamualim i run code with yolov8n-obb model and error generate Error: Shape of pred_cls before reshaping: torch.Size([5, 3, 1, 80, 80]) Reshaped pred_cls to: [5, 3, 80, 80] --------------------------------------------------------------------------- RuntimeError Traceback (most recent call last) <ipython-input-2-cb847e3ac8a0> in <cell line: 460>() 533 # Reshape pred_cls to have spatial dimensions 534 num_anchors = 3 --> 535 pred_cls = pred_cls.view(batch_size, num_classes, num_anchors, grid_size, grid_size) 536 pred_cls = pred_cls.permute(0, 1, 3, 4, 2) # permute channels for interpolation 537 RuntimeError: shape '[5, 1, 3, 20, 20]' is invalid for input of size 96000 code: for epoch in range(epochs): model.train() # Set the model to training mode for images, targets in train_dataloader: # Debugging code to inspect the targets print(f"Targets: {targets}") # Inspect the full targets structure print(f"Type of targets: {type(targets)}") # Check if targets is a dictionary # Now iterate over each target and inspect it for target in targets: print(f"Target: {target}") # Print the full target print(f"Type of target: {type(target)}") # Print the type of target # Ensure target is a dictionary before accessing its items if isinstance(target, dict): # Check if target is a dictionary print(f"Labels: {target['labels']}") # Access labels print(f"Boxes: {target['boxes']}") # Access boxes else: print("Target is not a dictionary. Investigate this structure.") # Assuming the target is now properly formatted and accessible images = images.to(device) # Ensure device is set (CPU or GPU) # Make sure to handle targets correctly for target in targets: if isinstance(target, dict): # Check if it's a valid dictionary true_boxes = target['boxes'].to(device) # Ground truth boxes true_cls = target['labels'].to(device) # Ground truth class labels else: print("Skipping invalid target structure.") continue # Skip invalid targets # Get image dimensions from a sample batch batch_size, channels, height, width = images.shape # Dynamic extraction from batch # Define grid_size dynamically based on the input image size and downsampling factor downsampling_factor = 32 # This is typical for YOLO models grid_size = height // downsampling_factor # For example, for 640x640 input, grid_size would be 20 optimizer.zero_grad() # Clear previous gradients # Initialize true_cls with zeros (absence of plane) true_cls_onehot = torch.zeros(batch_size, num_classes, grid_size, grid_size, device=device) # Loop over each target to assign the class labels based on the ground truth bounding boxes for i, target in enumerate(targets): if isinstance(target, dict): # Check if it's a valid dictionary print(f"Processing target: {target}") # Add this line to inspect the target structure if target['labels'] == 0: # If the label is 0 (plane) # Convert bounding box to grid cell indices x_center = (target['boxes'][0] / width) * grid_size # x_center to grid x-axis y_center = (target['boxes'][1] / height) * grid_size # y_center to grid y-axis grid_x = int(x_center) # Grid index for x-axis grid_y = int(y_center) # Grid index for y-axis # Set the class at the corresponding grid cell to 1 (plane) true_cls_onehot[i, 0, grid_y, grid_x] = 1 # plane class id is 0, mark it as 1 in the grid else: print("Skipping invalid target structure.") continue # Skip invalid targets # Forward pass through the model x_original, pred_boxes, pred_cls, pred_confidence = model(images) # Get image dimensions from a sample batch batch_size, channels, height, width = images.shape # Dynamic extraction from batch # Define grid_size dynamically based on the input image size and downsampling factor downsampling_factor = 32 # This is typical for YOLO models grid_size = height // downsampling_factor # For example, for 640x640 input, grid_size would be 20 # ... (true_cls_onehot generation logic remains the same) ... # Reshape pred_cls to have spatial dimensions num_anchors = 3 pred_cls = pred_cls.view(batch_size, num_classes, num_anchors, grid_size, grid_size) pred_cls = pred_cls.permute(0, 1, 3, 4, 2) # permute channels for interpolation # Before calculating focal loss, adjust pred_cls to match the spatial dimensions of true_cls_onehot pred_cls = F.interpolate(pred_cls, size=true_cls_onehot.shape[-2:], mode='bilinear', align_corners=False) pred_cls = pred_cls.permute(0, 1, 4, 2, 3) # permute back to original channel order pred_cls = pred_cls.reshape(batch_size, num_classes, -1) # flatten to original shape # Compute the classification loss (assuming focal loss is defined) try: loss_cls = focal_loss(pred_cls, true_cls_onehot) except ValueError as e: print(f"Error: {e}") print(f"Prediction shape: {pred_cls.shape}, Target shape: {true_cls_onehot.shape}") raise # Re-raise the exception to stop training and debug further # Compute bounding box regression loss (replace with your loss function) loss_bbox = bbox_loss(pred_boxes, true_boxes, width, height) # Assuming bbox_loss is defined # Compute confidence loss (optional, replace with your loss function) loss_conf = confidence_loss(pred_confidence, targets['conf'].to(device)) # Optional confidence loss # Total combined loss loss = combined_loss(pred_boxes, true_boxes, pred_cls, true_cls_onehot) # Backpropagation loss.backward() # Compute gradients optimizer.step() # Update model parameters torch.cuda.empty_cache() # Clear memory after training step to prevent overflow # Print the loss and shapes for debugging print(f"Epoch [{epoch}/{epochs}], Loss: {loss.item():.4f}, Loss_cls: {loss_cls.item():.4f}, Loss_bbox: {loss_bbox.item():.4f}, Loss_conf: {loss_conf.item():.4f}") print(f"x_original shape: {x_original.shape}") print(f"Predicted boxes shape: {pred_boxes.shape}") print(f"Predicted classes shape: {pred_cls.shape}") print(f"Confidence scores shape: {pred_confidence.shape}") # Save the trained model after training is complete torch.save(model.state_dict(), "yolov8n_obb_from_scratch.pth") the generate in this code and line is here and please give me a solution pred_cls = pred_cls.view(batch_size, num_classes, num_anchors, grid_size, grid_size) Continue reading...
