Is the training.main() doing a training from scratch, or fine-tuned on pre-trained weights?

[wsong1106]

Hi all -

I am training open_clip on my custom datasets, following the [instructions here](but the training actually was conducted using GPU). After training for 30 epochs, the image_to_text_R@n and the text_to_image_R@n are pretty low. My question is: is the training.main() doing the training from scratch, or is it doing the training from a pre-trained weights?

How can I make training.main() to start training from a pre-trained weights, for example your huggingface checkpoints?

[mitchellnw]

It's training from scratch

[wsong1106]

Thanks for your response, is there a way that we can start training from CLIP's pre-trained weights?

[wsong1106]

I think I figured out, by adding --pretrained openai into the commands