-
Notifications
You must be signed in to change notification settings - Fork 8
/
learner.py
254 lines (222 loc) · 11.2 KB
/
learner.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
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
import collections
import math
import os
import sys
import time
from pathlib import Path
import torch
from torch import nn
from torch.optim import AdamW
from torch.optim.lr_scheduler import CyclicLR
from tqdm import tqdm
from models.lxmert_utils import load_lxmert_qa
from optimizers.lamb import Lamb
from optimizers.lookahead import Lookahead
from pretrain.qa_answer_table import load_lxmert_qa
home = str(Path.home())
DataTuple = collections.namedtuple("DataTuple", "dataset loader evaluator")
load_lxmert_qa_path = home + "/snap/pretrained/model"
class Learner:
def __init__(self, model, data_tuple_dict, config):
self.model = model
self.criterion = nn.BCEWithLogitsLoss()
base_optim = Lamb(
params=self.model.parameters(), lr=1e-5, weight_decay=1.2e-6, min_trust=0.25
)
self.optim = Lookahead(base_optimizer=base_optim, k=5, alpha=0.8)
self.lr_scheduler = CyclicLR(
self.optim, base_lr=1e-5, max_lr=5e-5, cycle_momentum=False
)
self.train_tuple = data_tuple_dict["train_tuple"]
self.valid_tuple = data_tuple_dict["valid_tuple"]
self.test_tuple = data_tuple_dict["test_tuple"]
self.device = (
torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
)
self.output = home + "/snap/"
os.makedirs(self.output, exist_ok=True)
self.model.to(self.device)
self.adaptive = config["adaptive_enable"]
self.measure_flops = config["measure_flops"]
if self.measure_flops:
from thop import clever_format, profile
self.sparse = sparse = config["sparse_enable"]
if config["load_model"] == None:
load_lxmert_qa(
load_lxmert_qa_path, self.model, label2ans=self.train_tuple[0].label2ans
)
def train(self, num_epochs):
dset, loader, evaluator = self.train_tuple
best_valid = 0.0
iter_wrapper = lambda x: tqdm(x, total=len(loader))
for epoch in range(num_epochs):
t0 = time.time()
quesid2ans = {}
for i, (ques_id, feats, boxes, sent, target) in iter_wrapper(
enumerate(loader)
):
self.model.train()
self.optim.zero_grad()
feats, boxes, target = (
feats.to(self.device),
boxes.to(self.device),
target.to(self.device),
)
logit = self.model(feats, boxes, sent)
assert logit.dim() == target.dim() == 2
loss = self.criterion(logit, target) * logit.size(1)
if self.adaptive:
adapt_span_loss = 0.0
for l in self.model.lxrt_encoder.model.bert.encoder.layer:
adapt_span_loss += l.attention.self.adaptive_span.get_loss()
for l in self.model.lxrt_encoder.model.bert.encoder.x_layers:
adapt_span_loss += (
l.visual_attention.att.adaptive_span.get_loss()
)
for l in self.model.lxrt_encoder.model.bert.encoder.x_layers:
adapt_span_loss += l.lang_self_att.self.adaptive_span.get_loss()
for l in self.model.lxrt_encoder.model.bert.encoder.x_layers:
adapt_span_loss += l.visn_self_att.self.adaptive_span.get_loss()
for l in self.model.lxrt_encoder.model.bert.encoder.r_layers:
adapt_span_loss += l.attention.self.adaptive_span.get_loss()
loss += adapt_span_loss
#####################################################
loss.backward()
nn.utils.clip_grad_norm_(self.model.parameters(), 5.0)
self.optim.step()
self.lr_scheduler.step()
score, label = logit.max(1)
for qid, l in zip(ques_id, label.cpu().numpy()):
ans = dset.label2ans[l]
quesid2ans[qid.item()] = ans
#####################################################
if self.adaptive:
for l in self.model.lxrt_encoder.model.bert.encoder.layer:
l.attention.self.adaptive_span.clamp_param()
for l in self.model.lxrt_encoder.model.bert.encoder.x_layers:
l.visual_attention.att.adaptive_span.clamp_param()
for l in self.model.lxrt_encoder.model.bert.encoder.x_layers:
l.lang_self_att.self.adaptive_span.clamp_param()
for l in self.model.lxrt_encoder.model.bert.encoder.x_layers:
l.visn_self_att.self.adaptive_span.clamp_param()
for l in self.model.lxrt_encoder.model.bert.encoder.r_layers:
l.attention.self.adaptive_span.clamp_param()
#####################################################
log_str = "\nEpoch %d: Train %0.2f\n" % (
epoch,
evaluator.evaluate(quesid2ans) * 100.0,
)
log_str += "Loss: " + str(loss.item()) + "\t"
if self.adaptive:
log_str += "\tAdapt Span Loss: " + str(adapt_span_loss.item()) + "\n"
if self.measure_flops:
macs, params = profile(self.model, inputs=(feats, boxes, sent))
macs, params = clever_format([macs, params], "%.3f")
log_str += "\nMacs: " + macs + "\tParams: " + params + "\n"
if self.adaptive:
for layer_idx, i in enumerate(
self.model.lxrt_encoder.model.bert.encoder.layer
):
l = i.attention.self.adaptive_span.get_current_avg_span()
log_str += "Self Language %d %d\t" % (layer_idx, l)
log_str += "\n"
for layer_idx, i in enumerate(
self.model.lxrt_encoder.model.bert.encoder.x_layers
):
l = i.visual_attention.att.adaptive_span.get_current_avg_span()
log_str += "Cross %d %d\t" % (layer_idx, l)
log_str += "\n"
for layer_idx, i in enumerate(
self.model.lxrt_encoder.model.bert.encoder.x_layers
):
l = i.lang_self_att.self.adaptive_span.get_current_avg_span()
log_str += "Cross Self Language %d %d\t" % (layer_idx, l)
log_str += "\n"
for layer_idx, i in enumerate(
self.model.lxrt_encoder.model.bert.encoder.x_layers
):
l = i.visn_self_att.self.adaptive_span.get_current_avg_span()
log_str += "Cross Self Vision %d %d\t" % (layer_idx, l)
log_str += "\n"
for layer_idx, i in enumerate(
self.model.lxrt_encoder.model.bert.encoder.r_layers
):
l = i.attention.self.adaptive_span.get_current_avg_span()
log_str += "Self Vision %d %d\t" % (layer_idx, l)
# if self.sparse:
# alpha_val = {}
# for l in self.model.lxrt_encoder.model.bert.encoder.layer:
# alpha_val["lang_layer"] = l.attention.self.entmax_alpha.alpha_chooser
# for l in self.model.lxrt_encoder.model.bert.encoder.x_layers:
# alpha_val["cross_layer"] = l.visual_attention.att.entmax_alpha.alpha_chooser
# for l in self.model.lxrt_encoder.model.bert.encoder.x_layers:
# alpha_val["cross_lang_layer"] = l.lang_self_att.self.entmax_alpha.alpha_chooser
# for l in self.model.lxrt_encoder.model.bert.encoder.x_layers:
# alpha_val["cross_vision_layer"] = l.visn_self_att.self.entmax_alpha.alpha_chooser
# for l in self.model.lxrt_encoder.model.bert.encoder.r_layers:
# alpha_val["vision_layer"] = l.attention.self.entmax_alpha.alpha_chooser
# print("Alpha Values from Entmax have been saved at "+ home+'/snap/alpha_val_'+str(epoch)+'.pth')
# torch.save(alpha_val, home+'/snap/alpha_val_' + str(epoch)+ '.pth')
#####################################################
if self.valid_tuple is not None: # Do Validation
valid_score = self.evaluate(self.valid_tuple)
if valid_score > best_valid:
best_valid = valid_score
self.save("BEST")
log_str += "Epoch %d: Valid %0.2f\n" % (
epoch,
valid_score * 100.0,
) + "Epoch %d: Best %0.2f\n" % (epoch, best_valid * 100.0)
current_time = time.time() - t0
print(current_time)
log_str += "Time elpased for epoch %f\n" % (current_time)
print(log_str, end="")
with open(self.output + "/log.log", "a") as f:
f.write(log_str)
f.flush()
self.save("LAST")
def predict(self, eval_tuple, dump=None):
"""
Predict the answers to questions in a data split.
:param eval_tuple: The data tuple to be evaluated.
:param dump: The path of saved file to dump results.
:return: A dict of question_id to answer.
"""
self.model.eval()
dset, loader, evaluator = eval_tuple
quesid2ans = {}
iter_wrapper = lambda x: tqdm(x, total=len(loader))
print("Predict in progress")
for i, datum_tuple in iter_wrapper(enumerate(loader)):
ques_id, feats, boxes, sent = datum_tuple[:4] # Avoid seeing ground truth
with torch.no_grad():
feats, boxes = feats.to(self.device), boxes.to(self.device)
logit = self.model(feats, boxes, sent)
score, label = logit.max(1)
for qid, l in zip(ques_id, label.cpu().numpy()):
ans = dset.label2ans[l]
quesid2ans[qid.item()] = ans
if dump is not None:
evaluator.dump_result(quesid2ans, dump)
return quesid2ans
def evaluate(self, eval_tuple: DataTuple, dump=None):
"""Evaluate all data in data_tuple."""
quesid2ans = self.predict(eval_tuple, dump)
return eval_tuple.evaluator.evaluate(quesid2ans)
@staticmethod
def oracle_score(data_loader):
quesid2ans = {}
for i, (ques_id, feats, boxes, sent, target) in enumerate(data_loader):
_, label = target.max(1)
for qid, l in zip(ques_id, label.cpu().numpy()):
ans = dset.label2ans[l]
quesid2ans[qid.item()] = ans
return evaluator.evaluate(quesid2ans)
def save(self, name):
torch.save(self.model.state_dict(), os.path.join(self.output, "%s.pth" % name))
def load(self, path):
state_dict = torch.load(
"%s.pth" % path,
map_location=self.device
)
self.model.load_state_dict(state_dict)