自定义训练YOLOv4-tiny模型转换为TensorFlow模型时出现Reshape错误的解决方案咨询

The ValueError: cannot reshape array of size 607322 into shape (256,384,3,3) error occurs because the structure of your custom YOLOv4-Tiny model (defined in yolov4-tiny-device.cfg) doesn’t align with the hardcoded model architecture in the hunglc007/tensorflow-yolov4-tflite repository. Here’s how to fix it step by step:

1. Validate Your Custom YOLOv4-Tiny Config #

First, confirm critical parameters in your yolov4-tiny-device.cfg match your single-class setup:

• Check that the classes value under every [yolo] layer is set to 1 (since you have one custom class).
• For each [yolo] layer, the immediately preceding [convolutional] layer must have filters = (classes + 5) * 3. For your single-class model, this equals (1 + 5)*3 = 18—ensure both of these convolutional layers in your cfg use this filter count.
• Compare your cfg line-by-line with Darknet’s default yolov4-tiny.cfg to catch any other structural changes (like modified filter counts in intermediate layers) that could break weight loading.

2. Update TensorFlow Repository Configs #

Even though you updated __C.YOLO.CLASSES, you need to adjust additional settings to match your custom model:

• Open core/config.py and set __C.YOLO.NUM_CLASSES = 1—this tells the TensorFlow model how many classes to expect.
• If you modified any intermediate convolutional layers in your cfg (e.g., changed filter counts), update the corresponding layers in core/yolov4.py’s yolov4_tiny() function. For example, if you changed a conv layer’s filter count from 256 to 128, adjust the filters parameter in the matching conv2d call in the TF code.

3. Debug the Weight Mismatch #

The error points to a specific convolutional layer where the weight array size doesn’t fit the expected shape. To pinpoint the issue:

• Calculate the expected parameter count for each convolutional layer in your cfg:

  For a conv layer with filters=F, in_channels=I, kernel_size=K, total parameters = F * I * K * K (weights) + F (biases) = F*(I*K*K +1).

• Add print statements in core/utils.py’s load_weights function to log the expected shape and actual weight size for each layer, then cross-reference with your cfg’s parameters.
• Sync the TensorFlow code’s layer parameters with your cfg once you find the mismatched layer.

4. Re-run the Conversion Command #

Double-check your command (yours looks correct, but confirm all flags align):

py ./save_model.py --weights ./data/yolov4-tiny-device_last.weights --output ./checkpoint/yolov4-tiny-device-416 --input_size 416 --model yolov4 --tiny

After updating configs and model structure, this command should load weights without reshaping errors.

内容的提问来源于stack exchange，提问作者Daniel Botnik