Training yolov4-csp on a custom dataset starting with coco pretrained convolutional layers

[LukeAI]

@WongKinYiu

I'm trying to work out how I can train yolov4-csp on a custom dataset starting with coco pretrained convolutional layers.

1) Does training yolov4-csp with this repository require pytorch format .pt weights?

2) AlexeyAB provides coco-pretrained partial weights for transfer learning https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-csp.conv.142 Is it possible to convert darknet format partial weights to .pt ? If so, how?

3) If not, are pretrained weights in .pt available?

I tried training like this:

python train.py --device 0 --batch-size 14 --data data/custom.yaml --cfg models/yolov4-csp-custom.cfg --weights 'weights/yolov4-csp.conv.142' --name yolov4-csp

but the training curve looks the same as training from scratch, and https://github.com/WongKinYiu/ScaledYOLOv4/blob/4dfcec67f8b7db4893ed66000dd1b317691373a4/train.py#L60 it appears that only .pt weights can be used to initialise the model for training.

[WongKinYiu]

both of .pt and .weights could be pretrained weights.

modify the pretrained = pretrained = weights.endswith('.pt') or pretrained = weights.endswith('.weights'). and add code for loading .weights file as pretrained weights. for your reference: https://github.com/WongKinYiu/ScaledYOLOv4/blob/yolov4-csp/detect.py#L44-L48

[LukeAI]

thanks for your response! if I add this code to make it work with .weights as well as .pt would you accept a PR?

[WongKinYiu]

https://github.com/WongKinYiu/PyTorch_YOLOv4/blob/u3_preview/train.py#L116-L142

[LukeAI]

here I'd like to add the right cutoff for yolov4-csp.conv.142 - I guess it should be 143?

Would a good generic rule be to set cutoff=X+1 for weights.conv.X and cutoff=X for weights-tiny.conv.X ?

def load_darknet_weights(self, weights, cutoff=-1):
    # Parses and loads the weights stored in 'weights'

    # Establish cutoffs (load layers between 0 and cutoff. if cutoff = -1 all are loaded)
    file = Path(weights).name
    if file == 'darknet53.conv.74':
        cutoff = 75
    elif file == 'yolov3-tiny.conv.15':
        cutoff = 15

[huyhoangle86]

@LukeAI hey are u using .pt or .weights for pretrained ?

[LukeAI]

I actually ended up using scaled-yolo

[huyhoangle86]

@LukeAI but Did you use pretrained weights? If so what version .weights or .pt?