def _build_word_embedding(self):
self.bert_config = BertConfig.from_pretrained(self.config.bert_model_name)
if self.config.pretrained_bert:
bert_model = BertForPreTraining.from_pretrained(self.config.bert_model_name)
self.word_embedding = bert_model.bert.embeddings
self.pooler = bert_model.bert.pooler
self.pooler.apply(self.init_weights)
else:
self.pooler = BertPooler(self.bert_config)
self.word_embedding = BertEmbeddings(self.bert_config)
def __init__(self, model_name: str) -> None:
super().__init__()
config = BertConfig.from_pretrained(model_name)
self.input_dim = config.hidden_size
self.output_dim = config.vocab_size
# TODO(mattg): It's possible that we could use some kind of cache like we have in
# allennlp.modules.token_embedders.bert_token_embedder.PretrainedBertModel. That way, we
# would only load the BERT weights once. Though, it's not clear how to do that here, as we
# need to load `BertForMaskedLM`, not just `BertModel`...
bert_model = BertForMaskedLM.from_pretrained(model_name)
self.bert_lm_head = bert_model.cls
params = {
'tokenizer_config': {
'type': 'bert-base-uncased',
'params': {
'do_lower_case': True
}
},
'mask_probability': 0,
'max_seq_length': 128
}
mmf_tok = MMFTokenizer(OmegaConf.create(params))
mmf_tok._tokenizer = BertTokenizer(vocab_file="vocabulary.txt")
config = BertConfig.from_pretrained('bert-large-uncased',
num_labels=2,
vocab_size=len(vocabulary),
num_hidden_layers=3)
net = VisualBertModel(config, visual_embedding_dim=2048).cuda()
out_txt = mmf_tok({'text': report})
input_ids = torch.tensor(out_txt['input_ids']).unsqueeze(0)
input_mask = torch.tensor(out_txt['input_mask']).unsqueeze(0)
img = torch.zeros(1, 14, 2048)
out = net(input_ids=input_ids.cuda(),
text_mask=input_mask.cuda(),
visual_embeddings=img.cuda())
print(net.config.add_cross_attention)
print(out.keys())