Enable training with DocLayout-YOLO

[agombert]

This PR addresses several issues encountered when running the training command:

python train.py --data data/config --model m-doclayout --epoch 10 --image-size 1024 --batch-size 4 --project finetuning --pretrain models/doclayout_yolo_docstructbench_imgsz1024.pt --optimizer Adam --lr0 0.001 --warmup-epochs 1.0 --patience 2 --device 0

Changes Made

1. Added Missing Dependency

Added huggingface_hub to project dependencies in pyproject.toml:

dependencies = [
    # ... existing dependencies ...
    "huggingface_hub>=0.26.2",
]

2. Fixed AMP Check Issue

Following the solution from issue #32, commented out the AMP check line in utils/checks.py to avoid the UnpicklingError: invalid load key, '<' error.

3. Fixed Dilated Convolution Implementation

Modified the dilated_conv method in g2l_crm.py to properly handle batch normalization and activation functions:

def dilated_conv(self, x, dilation):
    weight = self.dcv.conv.weight
    padding = dilation * (self.k//2)
    x = F.conv2d(x, weight, stride=1, padding=padding, dilation=dilation)
    if hasattr(self.dcv, 'bn'):
        x = self.dcv.bn(x)
    if hasattr(self.dcv, 'act'):
        x = self.dcv.act(x)
    return x

This change fixes the AttributeError: 'Conv' object has no attribute 'bn' by properly checking for the existence of batch normalization and activation layers before applying them.

Testing

• Successfully ran training for 10 epochs
• No more AttributeError or UnpicklingError issues

Note: There are still some FutureWarning messages about torch.load with weights_only=False, but these don't affect functionality and can be addressed in a future update.

[JulioZhao97]

Thanks for your contribution! We can discuss further in #32

[JulioZhao97]

@agombert Hello, please check issue in #72

[dinhanhx]

I just try out this code, it works flawlessly, why hasn't this got merged into main yet???