This might be a silly question. But I try to config a customized Bart, and use
from_pretrained method to load weights. And I expect some errors raised, as my config neither completely matches the config in config.json · facebook/bart-base at main nor config.json · facebook/bart-large at main. But it didn’t.
from transformers import BartForConditionalGeneration, BartConfig myconfig = BartConfig(d_model=1024, max_position_embeddings=256, encoder_attention_heads=8, decoder_attention_heads=8, encoder_layers=10, decoder_layers=10) model = BartForConditionalGeneration(myconfig) model = model.from_pretrained('facebook/bart-base') model = BartForConditionalGeneration(myconfig) model = model.from_pretrained('facebook/bart-large')
Just wonder why…