Why is the resulting PyTorch model so big?

#3
by Luigi - opened

Hi Wang Qi,

Very nice and exciting project! What I'm particularly curious about is the model size.
If it's a model with only around 58 million parameters. We may expect the resulting checkpoint might take 232 MB as you use float32 for parameter. Why is the resulting PyTorch model file so big, around 843 MB ?

Yes you are right. I checked the export codes and found there was bug that output duplicate weights. The pytorch_model.bin should be 274MB instead of 843MB. I uploaded the improved version of it.
Thanks for your feedback.

Thanks, Sir/Madam. That's wonderful!!!

Luigi changed discussion status to closed

Sign up or log in to comment