-
Notifications
You must be signed in to change notification settings - Fork 11
/
main_retrieval_mlm.py
208 lines (182 loc) · 7.41 KB
/
main_retrieval_mlm.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
from utils.lib import *
from utils.args import get_args
from utils.logger import LOGGER, add_log_to_file
from utils.dist import (
is_main_process,
get_rank, get_world_size, iter_tqdm, all_gather,
NoOp)
from dataset import get_tsv_dls
from main_retrieval_task_specific import (
Dataset_Retrieval_TS)
from agent import Agent_Base
from model import LAVENDER_Base
class Dataset_Retrieval_MLM(Dataset_Retrieval_TS):
def __init__(self, args, img_tsv_path, txt, id2lineidx, split, tokzr=None):
super().__init__(
args, img_tsv_path, txt, id2lineidx, split, tokzr=tokzr)
def str2txt(self, s):
txt, mask = super().str2txt(s)
txt, mask = self.append_mask_tok2txt(txt, mask)
return txt, mask
@property
def prompt_text(self):
return "is the video-text paired, true or false?"
class LAVENDER_Retrieval_MLM(LAVENDER_Base):
def __init__(self, args, tokzr=None):
super().__init__(args, tokzr)
# bert = transformers.BertForMaskedLM.from_pretrained(
# './_models/huggingface_transformers/bert-base-uncased')
# config = bert.config
# self.fc_mtm = BertOnlyMLMHead(config)
# del bert
bert = transformers.AutoModelForMaskedLM.from_pretrained(
self.args.tokenizer)
if isinstance(bert, transformers.RobertaForMaskedLM):
self.fc_mtm = bert.lm_head
else:
self.fc_mtm = bert.cls
del bert
self.task_tok2id = {"vtm": 0, "mc": 1, "oe": 2, "cap": 3}
self.emb_task = T.nn.Parameter(
0.02*T.randn(10, self.hidden_size))
def forward(self, batch):
batch = defaultdict(lambda: None, batch)
img, txt, mask, vid = [
batch[key] for key in [
"img", "txt", "mask", "vid"]]
(_B, _T, _, _H, _W) = img.shape
_h, _w = _H//32, _W//32
feat_img, mask_img, feat_txt, mask_txt = self.go_feat(img, txt, mask)
pdt_feat_img, pdt_mask_img, pdt_feat_txt, pdt_mask_txt = [], [], [], []
mtm_ans = []
for i in range(_B):
for j in range(_B):
mt = mask_txt[j]
t = txt[j]
ft = feat_txt[j]
t, mt, ft = self.prepro_txt_inputs(
t, mt, ft, task_name=batch["task_name"],
prompt=batch["prompt"])
pdt_feat_img.append(feat_img[i].unsqueeze(0))
pdt_mask_img.append(mask_img[i].unsqueeze(0))
pdt_feat_txt.append(ft.unsqueeze(0))
pdt_mask_txt.append(mt.unsqueeze(0))
gt_txt = T.ones_like(t)*-1
if vid[i] == vid[j]:
gt_txt[-1] = self.true_token_id
else:
gt_txt[-1] = self.false_token_id
mtm_ans.append(gt_txt.unsqueeze(0))
pdt_feat_img, pdt_mask_img, pdt_feat_txt, pdt_mask_txt, mtm_ans = [
T.cat(x, dim=0)
for x in [pdt_feat_img, pdt_mask_img,
pdt_feat_txt, pdt_mask_txt, mtm_ans]
]
out, _ = self.go_cross(
pdt_feat_img, pdt_mask_img, pdt_feat_txt, pdt_mask_txt)
out = self.fc_mtm(out[:, (1+_h*_w)*_T:])
return out, mtm_ans
class Agent_Retrieval_MLM(Agent_Base):
def __init__(self, args, model):
super().__init__(args, model)
self.log = {'ls_tr': [], 'ac_vl': [], 'ac_ts': []}
def step(self, batch, is_train):
with T.cuda.amp.autocast(enabled=not self.args.deepspeed):
out = self.forward_step(batch)
out, ans = out
if is_train:
out = out.flatten(0, len(out.shape)-2)
ans = ans.flatten(0, len(ans.shape)-1)
ls = self.loss_func(out, ans)
self.backward_step(ls)
return ls.item()
else:
_B = len(batch["vid"])
p_true = out[:, :, self.true_token_id]
p_false = out[:, :, self.false_token_id]
out_mtm = p_true / (p_true+p_false)
ans_mtm = ans
out_mtm = out_mtm[ans_mtm != -1].view(_B, _B)
ans_mtm = ans_mtm[ans_mtm != -1].view(_B, _B)
out_mtm = T.argmax(out_mtm, dim=-1)
ans_mtm_idx = (ans_mtm == self.true_token_id).nonzero()[:, 1]
ac = (out_mtm == ans_mtm_idx).float().tolist()
return ac
def go_dl(self, ep, dl, is_train):
if is_train:
self.model.train()
else:
self.model.eval()
ret = []
idx = 0
for idx, batch in enumerate(dl):
if idx % self.args.logging_steps == 0 and is_train:
LOGGER.info(self.log_memory(ep, idx+1))
if self.args.enable_prompt:
batch["prompt"] = dl.dataset.get_prompt()
elif self.args.enable_task_token:
batch["task_name"] = "vtm"
batch = self.prepare_batch(batch)
curr_ret = self.step(batch, is_train)
if isinstance(curr_ret, list):
ret.extend(curr_ret)
else:
ret.append(curr_ret)
if idx % self.args.logging_steps != 0 and is_train:
LOGGER.info(self.log_memory(ep, idx+1))
gathered_ret = []
for ret_per_rank in all_gather(ret):
gathered_ret.extend(ret_per_rank)
ret = float(np.average(gathered_ret))
return ret
if __name__ == '__main__':
args = get_args()
tokzr = transformers.AutoTokenizer.from_pretrained(args.tokenizer)
dl_tr, dl_vl, dl_ts = get_tsv_dls(
args, Dataset_Retrieval_MLM, tokzr=tokzr)
if args.size_epoch == 0:
args.max_iter = 1
else:
args.max_iter = len(dl_tr) * args.size_epoch
model = LAVENDER_Retrieval_MLM(args, tokzr=tokzr)
model.load_ckpt(args.path_ckpt)
model.cuda()
if args.distributed:
LOGGER.info(f"n_gpu: {args.num_gpus}, rank: {get_rank()},"
f" world_size: {get_world_size()}")
args.path_output = '%s/_%s_%s' % (
args.path_output, args.task,
datetime.now().strftime('%Y%m%d%H%M%S'))
agent = Agent_Retrieval_MLM(args, model)
if args.distributed:
agent.prepare_dist_model()
agent.save_training_meta()
if is_main_process():
add_log_to_file('%s/stdout.txt' % (args.path_output))
else:
LOGGER = NoOp()
# DIST.barrier()
LOGGER.info("Saved training meta infomation, start training ...")
if os.path.exists(args.path_ckpt):
LOGGER.info("Zero-shot Evaluation")
ac_vl = agent.go_dl(0, dl_vl, False)
ac_ts = agent.go_dl(0, dl_ts, False)
LOGGER.info('ZS: %.2f %.2f' % (
ac_vl*100, ac_ts*100))
else:
LOGGER.info("No pre-trained weight, skip zero-shot Evaluation")
if args.size_epoch:
LOGGER.info("Start training....")
for e in iter_tqdm(range(args.size_epoch)):
ls_tr = agent.go_dl(e+1, dl_tr, True)
ac_vl = agent.go_dl(e+1, dl_vl, False)
ac_ts = agent.go_dl(e+1, dl_ts, False)
agent.log['ls_tr'].append(ls_tr)
agent.log['ac_vl'].append(ac_vl)
agent.log['ac_ts'].append(ac_ts)
agent.save_model(e+1)
LOGGER.info('Ep %d: %.6f %.6f %.6f' % (
e+1, ls_tr, ac_vl, ac_ts))
best_vl, best_ts = agent.best_epoch()
LOGGER.info(f'Best val @ ep {best_vl[0]+1}, {best_vl[1]:.6f}')
LOGGER.info(f'Best test @ ep {best_ts[0]+1}, {best_ts[1]:.6f}')