-
Notifications
You must be signed in to change notification settings - Fork 1
/
dual_enc_engine.py
180 lines (148 loc) · 7.55 KB
/
dual_enc_engine.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
# ----------------------------------------------------------------------------------------------
# MGSRTR Official Code
# Copyright (c) Rajesh Baidya. All Rights Reserved
# ----------------------------------------------------------------------------------------------
# Modified from GSRTR (https://github.com/jhcho99/gsrtr)
# Licensed under the Apache License 2.0 [see LICENSE for details]
# ----------------------------------------------------------------------------------------------
"""
Train and eval functions used in main.py
"""
import math
import os
import sys
import torch
import util.misc as utils
from typing import Iterable
from tqdm import tqdm
from torch.utils.tensorboard import SummaryWriter
from transformers import T5Tokenizer
from frame_semantic_transformer.data.tasks import FrameClassificationTask
def train_one_epoch_dual_enc(model: torch.nn.Module, tokenizer:T5Tokenizer, criterion: torch.nn.Module,
data_loader: Iterable, optimizer: torch.optim.Optimizer,
device: torch.device, epoch: int, max_norm: float = 0, writer:SummaryWriter = None):
model.train()
criterion.train()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
print_freq = 100
n_batches = len(data_loader)
loader_desc = 'Epoch [{:d}]: lr = {:.4f}, loss = {:.4f}, accuracy (verb = {:.4f}, noun = {:.4f}, bounding box = {:.4f})'
train_iterator = tqdm(data_loader, desc=loader_desc.format(epoch, 0.0, 0.0, 0.0, 0.0, 0.0))
for idx, (samples, captions, targets) in enumerate(train_iterator, 1):
tasks = []
for c in captions:
tasks.append(FrameClassificationTask(text=c[0], trigger_loc=c[1]))
text_captions = [task.get_input() for task in tasks]
inputs = tokenizer(
text_captions,
padding="max_length",
truncation=True,
return_token_type_ids=True,
return_attention_mask=True,
add_special_tokens=True,
return_tensors="pt",
)
# data & target
samples = samples.to(device)
inputs = inputs.to(device)
targets = [{k: v.to(device) if type(v) is not str else v for k, v in t.items()} for t in targets]
# model output & calculate loss
outputs = model(samples, inputs, targets)
loss_dict = criterion(outputs, targets)
weight_dict = criterion.weight_dict
losses = sum(loss_dict[k] * weight_dict[k] for k in loss_dict.keys() if k in weight_dict)
# reduce losses over all GPUs for logging purposes
# scaled with different loss coefficients
loss_dict_reduced = utils.reduce_dict(loss_dict)
loss_dict_reduced_unscaled = {f'{k}_unscaled': v
for k, v in loss_dict_reduced.items()}
loss_dict_reduced_scaled = {k: v * weight_dict[k]
for k, v in loss_dict_reduced.items() if k in weight_dict}
losses_reduced_scaled = sum(loss_dict_reduced_scaled.values())
loss_value = losses_reduced_scaled.item()
# stop when loss is nan or inf
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
print(loss_dict_reduced)
sys.exit(1)
# loss backward & optimzer step
optimizer.zero_grad()
losses.backward()
if max_norm > 0:
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm)
optimizer.step()
metric_logger.update(loss=loss_value, **loss_dict_reduced_scaled, **loss_dict_reduced_unscaled)
metric_logger.update(lr=optimizer.param_groups[0]["lr"])
if idx%print_freq == 0:
items = {k: meter.global_avg for k, meter in metric_logger.meters.items()}
train_iterator.set_description(loader_desc.format(epoch, items['lr'], items['loss'], items['verb_acc_top1_unscaled'],
items['noun_acc_all_top1_unscaled'], items['bbox_acc_top5_unscaled']))
if writer:
writer.add_scalar("training loss", items['loss'], epoch * n_batches + idx)
writer.add_scalars('noun_accuracy', {
"top-1": items['noun_acc_top1_unscaled'],
"top-5": items['noun_acc_top5_unscaled'],
}, epoch * n_batches + idx)
writer.add_scalars('verb_accuracy', {
"top-1": items['verb_acc_top1_unscaled'],
"top-5": items['verb_acc_top5_unscaled'],
}, epoch * n_batches + idx)
writer.add_scalars('bounding_box_accuracy', {
"top-1": items['bbox_acc_top1_unscaled'],
"top-5": items['bbox_acc_top5_unscaled'],
}, epoch * n_batches + idx)
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
@torch.no_grad()
def evaluate_flicker(model, tokenizer:T5Tokenizer, criterion, data_loader, device):
model.eval()
criterion.eval()
metric_logger = utils.MetricLogger(delimiter=" ")
print_freq = 10
loader_desc = 'Test: loss = {:.4f}, accuracy (verb = {:.4f}, noun = {:.4f}, bounding box = {:.4f})'
test_iterator = tqdm(data_loader, desc=loader_desc.format(0.0, 0.0, 0.0, 0.0))
for idx, (samples, captions, targets) in enumerate(test_iterator, 1):
tasks = []
for c in captions:
tasks.append(FrameClassificationTask(text=c[0], trigger_loc=c[1]))
captions = [task.get_input() for task in tasks]
inputs = tokenizer(
captions,
padding="max_length",
truncation=True,
return_token_type_ids=True,
return_attention_mask=True,
add_special_tokens=True,
return_tensors="pt",
)
# data & target
samples = samples.to(device)
inputs = inputs.to(device)
targets = [{k: v.to(device) if type(v) is not str else v for k, v in t.items()} for t in targets]
# model output & calculate loss
outputs = model(samples, inputs, targets)
loss_dict = criterion(outputs, targets, eval=True)
weight_dict = criterion.weight_dict
# reduce losses over all GPUs for logging purposes
# scaled with different loss coefficients
loss_dict_reduced = utils.reduce_dict(loss_dict)
loss_dict_reduced_unscaled = {f'{k}_unscaled': v
for k, v in loss_dict_reduced.items()}
loss_dict_reduced_scaled = {k: v * weight_dict[k]
for k, v in loss_dict_reduced.items() if k in weight_dict}
losses_reduced_scaled = sum(loss_dict_reduced_scaled.values())
loss_value = losses_reduced_scaled.item()
metric_logger.update(loss=loss_value, **loss_dict_reduced_scaled, **loss_dict_reduced_unscaled)
if idx % print_freq == 0:
items = {k: meter.global_avg for k, meter in metric_logger.meters.items()}
test_iterator.set_description(
loader_desc.format(items['loss'], items['verb_acc_top1_unscaled'],
items['noun_acc_all_top1_unscaled'], items['bbox_acc_top5_unscaled']))
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
stats = {k: meter.global_avg for k, meter in metric_logger.meters.items()}
return stats