300 lines
12 KiB
Python
300 lines
12 KiB
Python
import time
|
|
import os
|
|
import glog as log
|
|
import numpy as np
|
|
import json
|
|
import torch
|
|
import torch.nn.functional as F
|
|
from runners.runner import Runner
|
|
from copy import deepcopy
|
|
from optim_utils import init_optim
|
|
from transformers.models.bart.configuration_bart import BartConfig
|
|
from models.nextqa_bart import AVSDBart
|
|
from time import time
|
|
|
|
|
|
def tokenize(obj, tokenizer):
|
|
if isinstance(obj, str):
|
|
return tokenizer.convert_tokens_to_ids(tokenizer.tokenize(obj))
|
|
if isinstance(obj, dict):
|
|
return dict((n, tokenize(o)) for n, o in obj.items())
|
|
return list(tokenize(o) for o in obj)
|
|
|
|
|
|
class NEXTQARunner(Runner):
|
|
def __init__(self, config, tokenizer, vocab_size):
|
|
super(NEXTQARunner, self).__init__(config)
|
|
bart_config = BartConfig.from_json_file(self.config['bart_config'])
|
|
|
|
self.model = AVSDBart.from_pretrained(
|
|
'facebook/bart-{}'.format(self.config['bart_size']), config=bart_config)
|
|
|
|
# Resize the embedding to match the vocab with additional special toks
|
|
# This takes care of resizing weights of related parts of the network
|
|
|
|
if vocab_size != bart_config.vocab_size:
|
|
self.model.resize_token_embeddings(vocab_size)
|
|
|
|
self.model.to(self.config['device'])
|
|
if not self.config['generating']:
|
|
self.optimizer, self.scheduler = init_optim(self.model, self.config)
|
|
self.tokenizer = tokenizer
|
|
|
|
def forward(self, batch):
|
|
|
|
for key in batch:
|
|
if isinstance(batch[key], torch.Tensor):
|
|
batch[key] = batch[key].cuda()
|
|
|
|
########################################################
|
|
input_ids = batch['input_ids']
|
|
video_place_holder_ids = batch['video_place_holder_ids']
|
|
app_feats = batch['app_feats']
|
|
mot_feats = batch['mot_feats']
|
|
lm_labels = batch['lm_labels']
|
|
input_mask = batch['input_mask']
|
|
|
|
app_interval = batch['app_interval']
|
|
mot_interval = batch['mot_interval']
|
|
question_intervals = batch['question_intervals']
|
|
vis_state_vector_idx = batch['vis_state_vector_idx']
|
|
question_state_vector_idx = batch['question_state_vector_idx']
|
|
########################################################
|
|
|
|
bart_output = self.model(
|
|
input_ids=input_ids,
|
|
video_place_holder_ids=video_place_holder_ids,
|
|
i3d_rgb=app_feats,
|
|
i3d_flow=mot_feats,
|
|
attention_mask=input_mask,
|
|
labels=lm_labels,
|
|
i3d_rgb_interval=app_interval,
|
|
i3d_flow_interval=mot_interval,
|
|
question_intervals=question_intervals,
|
|
vis_state_vector_idx=vis_state_vector_idx,
|
|
question_state_vector_idx=question_state_vector_idx,
|
|
output_attentions=True,
|
|
return_dict=True
|
|
)
|
|
|
|
output = {}
|
|
|
|
if self.config['print_output']:
|
|
logits = bart_output['logits']
|
|
probs = F.softmax(logits, dim=-1)
|
|
preds = torch.topk(probs, 1)[1].squeeze(-1)
|
|
preds = preds.tolist()
|
|
lm_labels_list = lm_labels[:, 1:].tolist()
|
|
lm_labels_list = [[s for s in label if s != -1] for label in lm_labels_list]
|
|
reponses = ''
|
|
labels = ''
|
|
for pred, label in zip(preds, lm_labels_list):
|
|
reponses += self.tokenizer.decode(pred) + '\n'
|
|
labels += self.tokenizer.decode(label) + '\n'
|
|
|
|
output['reponses'] = reponses
|
|
output['gt'] = labels
|
|
|
|
|
|
gen_key = 'gen_loss (x{})'.format(self.config['gen_coeff'])
|
|
gen_loss = bart_output['gen_loss']
|
|
gen_loss = self.config['gen_coeff'] * gen_loss
|
|
|
|
|
|
elbo_global_key = 'elbo_loss_global (x{})'.format(self.config['elbo_global_coeff'])
|
|
if bart_output['elbo_loss_global'] is not None:
|
|
elbo_global_loss = bart_output['elbo_loss_global']
|
|
elbo_global_loss = self.config['elbo_global_coeff'] * elbo_global_loss
|
|
else:
|
|
elbo_global_loss = torch.tensor(0.0)
|
|
|
|
elbo_local_key = 'elbo_loss_local (x{})'.format(self.config['elbo_local_coeff'])
|
|
if bart_output['elbo_loss_local'] is not None:
|
|
elbo_local_loss = bart_output['elbo_loss_local']
|
|
elbo_local_loss = self.config['elbo_local_coeff'] * elbo_local_loss
|
|
else:
|
|
elbo_local_loss = torch.tensor(0.0)
|
|
|
|
total_loss = gen_loss + elbo_global_loss + elbo_local_loss
|
|
|
|
output['losses'] = {
|
|
gen_key: gen_loss,
|
|
elbo_local_key: elbo_local_loss,
|
|
elbo_global_key: elbo_global_loss,
|
|
'tot_loss': total_loss
|
|
}
|
|
del bart_output
|
|
return output
|
|
|
|
|
|
def generate(self, dataset, app_feats, mot_feats, tag, tokenizer, start_idx_gen, end_idx_gen, gen_subset_num=None):
|
|
|
|
self.model.eval()
|
|
results = {}
|
|
app_sep, mot_sep, ph_token = tokenizer.convert_tokens_to_ids(
|
|
['<s0>', '<s1>', '<place_holder>'])
|
|
|
|
# Generate the repsonse for each round
|
|
log.info('[INFO] Generating responses for {} samples'.format(len(dataset)))
|
|
with torch.no_grad():
|
|
counter = 0
|
|
for idx in range(start_idx_gen, end_idx_gen):
|
|
start_time = time()
|
|
cur_sample = dataset.loc[idx]
|
|
video_name, ques, ans, qid = str(cur_sample['video']), str(cur_sample['question']),\
|
|
str(cur_sample['answer']), str(cur_sample['qid'])
|
|
if video_name not in results:
|
|
results[video_name] = {}
|
|
|
|
input_ids = tokenize(ques, tokenizer)
|
|
|
|
app_feat = app_feats[video_name]
|
|
app_feat = torch.from_numpy(app_feat).type(torch.float32)
|
|
|
|
mot_feat = mot_feats[video_name]
|
|
mot_feat = torch.from_numpy(mot_feat).type(torch.float32)
|
|
|
|
bos, eos, ques_state = self.tokenizer.convert_tokens_to_ids(['<s>', '</s>', '<s2>'])
|
|
|
|
# Add state tokens
|
|
input_ids.insert(0, ques_state)
|
|
|
|
input_ids = torch.Tensor(input_ids).long()
|
|
|
|
dummy = torch.ones((1, 16)) * ph_token
|
|
video_place_holder_ids = torch.cat(
|
|
[torch.ones((1, 1)) * app_sep, dummy,
|
|
torch.ones((1, 1)) * mot_sep, dummy,
|
|
], dim=-1).long()
|
|
|
|
# Now we get the intervals of the visual input tokens
|
|
# Here the interval do not change across the batch dimension
|
|
app_interval = [0, 16 + 1] # the last token is not part of this modality
|
|
mot_interval = [16 + 1, 2 * 16 + 2]
|
|
vis_state_vector_idx = [app_interval[0], mot_interval[0]]
|
|
|
|
response = self.beam_search_generation(
|
|
input_ids,
|
|
app_feat, mot_feat,
|
|
app_interval, mot_interval,
|
|
vis_state_vector_idx, video_place_holder_ids, tokenizer)
|
|
|
|
# Decode the response
|
|
response = self.tokenizer.decode(response)
|
|
|
|
results[video_name][qid] = response
|
|
time_elapsed = int(time() - start_time)
|
|
print('Generating resonse {} / {} -- took {}s'.format(counter + 1, len(dataset), time_elapsed))
|
|
counter += 1
|
|
|
|
# Create a file with all responses
|
|
file_name = 'results_nextqa_beam_depth_{}'.format(self.config['beam_depth'])
|
|
if gen_subset_num is not None:
|
|
file_name += f'-part_{gen_subset_num}'
|
|
file_name = f'{tag}_' + file_name
|
|
output_path = os.path.join(self.config['output_dir_nextqa'], file_name + '.json')
|
|
with open(output_path, 'w') as f:
|
|
json.dump(results, f, indent=4)
|
|
log.info('Results logged to {}'.format(output_path))
|
|
print(os.getcwd())
|
|
# Switch back to training mode
|
|
self.model.train()
|
|
|
|
|
|
def beam_search_generation(
|
|
self, input_ids,
|
|
app_feat, mot_feat,
|
|
app_interval, mot_interval,
|
|
vis_state_vector_idx, video_place_holder_ids, tokenizer):
|
|
|
|
eos_token = tokenizer.eos_token_id
|
|
unk_token = tokenizer.unk_token_id
|
|
question_sep = tokenizer.convert_tokens_to_ids('<s2>')
|
|
|
|
gen_ans = [eos_token]
|
|
hyplist = [([], 0.0, [eos_token])]
|
|
best_state = None
|
|
comp_hyplist = []
|
|
|
|
app_feat = app_feat.unsqueeze(0).cuda()
|
|
mot_feat = mot_feat.unsqueeze(0).cuda()
|
|
video_place_holder_ids = video_place_holder_ids.cuda()
|
|
text_shift_len = video_place_holder_ids.size(-1)
|
|
|
|
question_intervals = [[0 + text_shift_len, input_ids.size(0) + text_shift_len]] # The last token is the question state token (not part of the history)
|
|
|
|
question_state_vector_idx = [x[0] for x in question_intervals]
|
|
|
|
input_ids = input_ids.long().cuda().unsqueeze(0)
|
|
encoder_outputs = None
|
|
|
|
for i in range(self.config['max_generation_length']):
|
|
new_hyplist = []
|
|
argmin = 0
|
|
for out, lp, st in hyplist:
|
|
decoder_input_ids = torch.tensor(st).long().cuda().unsqueeze(0)
|
|
|
|
bart_output = self.model(
|
|
input_ids=input_ids,
|
|
video_place_holder_ids=video_place_holder_ids,
|
|
i3d_rgb=app_feat,
|
|
i3d_flow=mot_feat,
|
|
encoder_outputs=encoder_outputs,
|
|
decoder_input_ids=decoder_input_ids,
|
|
i3d_rgb_interval=app_interval,
|
|
i3d_flow_interval=mot_interval,
|
|
question_intervals=question_intervals,
|
|
vis_state_vector_idx=vis_state_vector_idx,
|
|
question_state_vector_idx=question_state_vector_idx,
|
|
output_attentions=True,
|
|
generate=True,
|
|
return_dict=True
|
|
)
|
|
|
|
if encoder_outputs is None:
|
|
encoder_outputs = [
|
|
bart_output['encoder_last_hidden_state'],
|
|
bart_output['encoder_hidden_states'],
|
|
bart_output['encoder_attentions'],
|
|
bart_output['encoder_QAs_local'],
|
|
bart_output['encoder_PAs_local'],
|
|
bart_output['encoder_QA_global'],
|
|
bart_output['encoder_PA_global'],
|
|
bart_output['encoder_state_vectors']
|
|
]
|
|
|
|
logits = bart_output['logits'][:,-1,:].squeeze() # get the logits of the last token
|
|
logp = F.log_softmax(logits, dim=0)
|
|
lp_vec = logp.cpu().data.numpy() + lp
|
|
if i >= self.config['min_generation_length']:
|
|
new_lp = lp_vec[eos_token] + self.config['length_penalty'] * (len(out) + 1)
|
|
comp_hyplist.append((out, new_lp))
|
|
if best_state is None or best_state < new_lp:
|
|
best_state = new_lp
|
|
count = 1
|
|
for o in np.argsort(lp_vec)[::-1]: # reverse the order
|
|
if o in [eos_token, unk_token]:
|
|
continue
|
|
new_lp = lp_vec[o]
|
|
if len(new_hyplist) == self.config['beam_depth']:
|
|
if new_hyplist[argmin][1] < new_lp:
|
|
new_st = deepcopy(st)
|
|
new_st.append(int(o))
|
|
new_hyplist[argmin] = (out + [o], new_lp, new_st)
|
|
argmin = min(enumerate(new_hyplist), key=lambda h: h[1][1])[0]
|
|
else:
|
|
break
|
|
else:
|
|
new_st = deepcopy(st)
|
|
new_st.append(int(o))
|
|
new_hyplist.append((out + [o], new_lp, new_st))
|
|
if len(new_hyplist) == self.config['beam_depth']:
|
|
argmin = min(enumerate(new_hyplist), key=lambda h: h[1][1])[0]
|
|
count += 1
|
|
hyplist = new_hyplist
|
|
|
|
if len(comp_hyplist) > 0:
|
|
maxhyps = sorted(comp_hyplist, key=lambda h: -h[1])[:1]
|
|
return maxhyps[0][0]
|
|
else:
|
|
return []
|