Saving TF model using save_pretrained()
#4
by kostaras - opened
While the initial tf_model.h5 is about 62.5 MB, when I load it using TFT5ForConditionalGeneration.from_pretrained() and then save it using save_pretrained() the saved model is about 126 MB. That is why it saves 3 times the shared tokens_embed layer (about 31.5 MB) instead of just once. Does someone know how to properly save the shared layer only one time ?
