Thinking-while-Observing / code /model_LXM2T5.py
QingyiSi's picture
Upload 14 files
44ddffd
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import init
import copy
from config4LXMT5_DDP import args
import collections
from transformers import LxmertConfig, LxmertTokenizer, LxmertModel,BertTokenizer#,BaseModelOutputWithPastAndCrossAttentions
from transformers import T5Tokenizer, T5Model, T5Config, T5ForConditionalGeneration
from transformers.modeling_outputs import BaseModelOutputWithPastAndCrossAttentions
T5tokenizer = T5Tokenizer.from_pretrained("../model/t5-large")#"t5-large")
LXMtokenizer = BertTokenizer.from_pretrained('../model/bert-base-uncased/vocab.txt')
T5config = T5Config.from_pretrained('../model/t5-large')
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
LXM_source_masks=None,token_type_ids=None, visual_features=None, spatial_features=None,T5_target_ids=None,T5_target_masks=None):
attention_mask=LXM_source_masks, token_type_ids=token_type_ids, visual_feats=visual_features, visual_pos=spatial_features)
class LXMT52T5(nn.Module):
def __init__(self):
super(LXMT52T5, self).__init__()
self.T5model = T5ForConditionalGeneration.from_pretrained("../model/t5-large").to(device)
self.LXMmodel = LxmertModel.from_pretrained('../model/lxmert-base-uncased').to(device)
self.mapping = torch.nn.Sequential(
torch.nn.Linear(768, 1024),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(1024, 1024)
)
def LXMT5end2T5dec(self, train=None, LXM_source_ids=None, LXM_source_masks=None,T5_source_ids=None, T5_source_masks=None,token_type_ids=None, visual_features=None, spatial_features=None,T5_target_ids=None,T5_target_masks=None):
if 1:
LXM_encoder_output_seq = self.LXMmodel(input_ids=LXM_source_ids, attention_mask=LXM_source_masks, token_type_ids=token_type_ids, visual_feats=visual_features, visual_pos=spatial_features)
LXM_lang_enc_out = LXM_encoder_output_seq.language_output
LXM_visual_enc_out = LXM_encoder_output_seq.vision_output
LXM_VL_encoder_output_seq = torch.cat((LXM_lang_enc_out, LXM_visual_enc_out),1)
#if 1: # (w/o wiki passages)
# T5_encoder_output_seq = self.T5model.encoder(input_ids=T5_source_ids, attention_mask=T5_source_masks)
# final_encoder_output_seq = torch.cat((final_LXM_encoder_output_seq, T5_encoder_output_seq["last_hidden_state"]),1)
if 1: # (w/ wiki passages)
final_encoder_output_seq_list = []
final_T5_encoder_output_seq_list = []
for ind in range(args.num_wiki):
T5_encoder_output_seq = self.T5model.encoder(input_ids=T5_source_ids[:,ind,:], attention_mask=T5_source_masks[:,ind,:])
#if 1: #(T5 encoder only)
# final_T5_encoder_output_seq_list.append(T5_encoder_output_seq["last_hidden_state"])
tmp_encoder_output_seq = torch.cat((final_LXM_encoder_output_seq, T5_encoder_output_seq["last_hidden_state"]),1)
final_encoder_output_seq_list.append(tmp_encoder_output_seq)
final_encoder_output_seq = torch.cat(final_encoder_output_seq_list,1)
# ablation study on two encoders
# LXMERTenc-T5dec
final_encoder_output_seq = final_LXM_encoder_output_seq
# T5enc-T5dec
final_encoder_output_seq = torch.cat(final_T5_encoder_output_seq_list,1)
my_order_dict=T5_encoder_output_seq
# replace the origin order_dict with our designed final_encoder_output_seq
my_order_dict.last_hidden_state=final_encoder_output_seq
if train:
if args.allAns:
outputs = self.T5model(encoder_outputs=my_order_dict, labels=T5_target_ids, decoder_attention_mask=T5_target_masks)
else:
outputs = self.T5model(encoder_outputs=my_order_dict, labels=T5_target_ids, decoder_attention_mask=T5_target_masks)
return outputs
else:
if torch.cuda.device_count() > 1:
pred = self.T5model.generate(encoder_outputs=my_order_dict)
else:
pred = self.T5model.generate(encoder_outputs=my_order_dict)
return pred
def forward(self, train=None, LXM_source_ids=None, LXM_source_masks=None,T5_source_ids=None, T5_source_masks=None,token_type_ids=None, visual_features=None, spatial_features=None,T5_target_ids=None,T5_target_masks=None):
return self.LXMT5end2T5dec(train, LXM_source_ids, LXM_source_masks, T5_source_ids, T5_source_masks, token_type_ids, visual_features, spatial_features, T5_target_ids, T5_target_masks)