The pretrained weights on ImageNet

[libzzluo]

@glenn-jocher
Hi, First of all, thank you very much for your code. I have started training for 24 hours (10 epochs) on my GTX1080, but it seems that I can't load the pre-trained weights on ImageNet (the darknet53.comv). And it takes long time to train from scratch. TAT

[glenn-jocher]

@libzzluo yes it takes a long time to train from scratch. I've not trained ImageNet, only COCO and xView, and in general if takes about 100-200 epochs to fully train (before overtraining begins). On COCO I get about 10-15 epochs per day on a 1080 Ti, depending on the augmentation used on the input images (augmentation is generally slow, though I've sped it up as fast as I could using the latest opencv functions).

About your second question I'm not sure I understand. The only pretrained weights I've used with the model are the yolov3 weights you can download by:

wget https://pjreddie.com/media/files/yolov3.weights

These darknet weights are loaded with load_weights() in models.py. They can be converted to PyTorch weights and loaded instead with the default PyTorch methods. These two options are both available in detect.py, where the code takes the appropriate action based on the file extension (*.pt for PyTorch, *.weights for darknet):

    weights_path = 'checkpoints/yolov3.weights'
    if weights_path.endswith('.weights'):  # saved in darknet format
        load_weights(model, weights_path)
    else:  # endswith('.pt'), saved in pytorch format
        checkpoint = torch.load(weights_path, map_location='cpu')
        model.load_state_dict(checkpoint['model'])
        del checkpoint

[jaelim]

@glenn-jocher If I understand correctly, your train.py trains a model (in this case a darknet on coco) from scratch, but not with an option to start the training with loading a pre-trained weight?