-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathmain.py
200 lines (161 loc) · 7.08 KB
/
main.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
import argparse
import json
# import cPickle as pickle
import pickle as cPickle # python3
from collections import defaultdict, Counter
from os.path import dirname, join
import os
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
import numpy as np
from dataset import Dictionary, VQAFeatureDataset
import base_model
from train import train
import utils
import click
from vqa_debias_loss_functions import *
def parse_args():
parser = argparse.ArgumentParser("Train the BottomUpTopDown model with a de-biasing method")
# Arguments we added
parser.add_argument(
'--cache_features', default=True, #True
help="Cache image features in RAM. Makes things much faster, "
"especially if the filesystem is slow, but requires at least 48gb of RAM")
parser.add_argument(
'--dataset', default='cpv2',
choices=["v2", "cpv2", "cpv1"],
help="Run on VQA-2.0 instead of VQA-CP 2.0"
)
parser.add_argument(
'-p', "--entropy_penalty", default=0.36, type=float,
help="Entropy regularizer weight for the learned_mixin model")
parser.add_argument(
'--mode', default="q_v_debias_inner_contras", #updn
choices=["updn", "q_debias","v_debias","q_v_debias", "v_debias_contras", "v_debias_ori_contras_eucl", "v_debias_ori_contras_cos", "q_v_debias_ori_contras_cos", "q_v_debias_inner_contras"],
help="Kind of ensemble loss to use")
parser.add_argument(
'--debias', default="learned_mixin", # learned_mixin
choices=["learned_mixin", "reweight", "bias_product", "none",'focal'],
help="Kind of ensemble loss to use")
parser.add_argument(
'--topq', type=int,default=1,
choices=[1,2,3],
help="num of words to be masked in questio")
parser.add_argument(
'--keep_qtype', default=True,
help="keep qtype or not")
parser.add_argument(
'--topv', type=int,default=1,
choices=[1,3,5,-1],
help="num of object bbox to be masked in image")
parser.add_argument(
'--top_hint',type=int, default=9,
choices=[9,18,27,36],
help="num of hint")
parser.add_argument(
'--qvp', type=int,default=5,
choices=[0,1,2,3,4,5,6,7,8,9,10],
help="ratio of q_bias and v_bias")
parser.add_argument(
'--eval_each_epoch', default=True,
help="Evaluate every epoch, instead of at the end")
# Arguments from the original model, we leave this default, except we
# set --epochs to 30 since the model maxes out its performance on VQA 2.0 well before then
parser.add_argument('--margin', type=float, default=0.3,
help="Margin of the original contrastive loss")
parser.add_argument('--contras_loss_weight', type=int, default=2)
parser.add_argument('--epochs', type=int, default=30)
parser.add_argument('--num_hid', type=int, default=1024)
parser.add_argument('--model', type=str, default='baseline0_newatt')
parser.add_argument('--output', type=str, default='logs/q_v_debias_contras')
parser.add_argument('--batch_size', type=int, default=512)
parser.add_argument('--seed', type=int, default=0, help='random seed')
args = parser.parse_args()
return args
def get_bias(train_dset,eval_dset):
# Compute the bias:
# The bias here is just the expected score for each answer/question type
answer_voc_size = train_dset.num_ans_candidates
# question_type -> answer -> total score
question_type_to_probs = defaultdict(Counter)
# question_type -> num_occurances
question_type_to_count = Counter()
for ex in train_dset.entries:
ans = ex["answer"]
q_type = ans["question_type"]
question_type_to_count[q_type] += 1
if ans["labels"] is not None:
for label, score in zip(ans["labels"], ans["scores"]):
question_type_to_probs[q_type][label] += score
question_type_to_prob_array = {}
for q_type, count in question_type_to_count.items():
prob_array = np.zeros(answer_voc_size, np.float32)
for label, total_score in question_type_to_probs[q_type].items():
prob_array[label] += total_score
prob_array /= count
question_type_to_prob_array[q_type] = prob_array
for ds in [train_dset,eval_dset]:
for ex in ds.entries:
q_type = ex["answer"]["question_type"]
ex["bias"] = question_type_to_prob_array[q_type]
def main():
args = parse_args()
dataset=args.dataset
args.output=os.path.join('logs',args.output)
if not os.path.isdir(args.output):
utils.create_dir(args.output)
else:
if click.confirm('Exp directory already exists in {}. Erase?'
.format(args.output, default=False)):
os.system('rm -r ' + args.output)
utils.create_dir(args.output)
else:
os._exit(1)
if dataset=='cpv1':
dictionary = Dictionary.load_from_file('data/dictionary_v1.pkl')
elif dataset=='cpv2' or dataset=='v2':
dictionary = Dictionary.load_from_file('data/dictionary.pkl')
print('margin of Contrasitive loss:', args.margin)
print("Building train dataset...")
train_dset = VQAFeatureDataset('train', dictionary, dataset=dataset,
cache_image_features=args.cache_features)
print("Building test dataset...")
eval_dset = VQAFeatureDataset('val', dictionary, dataset=dataset,
cache_image_features=args.cache_features)
get_bias(train_dset,eval_dset)
# Build the model using the original constructor
constructor = 'build_%s' % args.model
model = getattr(base_model, constructor)(train_dset, args.num_hid).cuda()
if dataset=='cpv1':
model.w_emb.init_embedding('data/glove6b_init_300d_v1.npy')
elif dataset=='cpv2' or dataset=='v2':
model.w_emb.init_embedding('data/glove6b_init_300d.npy')
# Add the loss_fn based our arguments
if args.debias == "bias_product":
model.debias_loss_fn = BiasProduct()
elif args.debias == "none":
model.debias_loss_fn = Plain()
elif args.debias == "reweight":
model.debias_loss_fn = ReweightByInvBias()
elif args.debias == "learned_mixin":
model.debias_loss_fn = LearnedMixin(args.entropy_penalty)
elif args.debias=='focal':
model.debias_loss_fn = Focal()
else:
raise RuntimeError(args.mode)
if args.mode =='v_debias_ori_contras_cos' or args.mode=='q_v_debias_ori_contras_cos' or args.mode=='q_v_debias_inner_contras':
model.is_contras = True
with open('util/qid2type_%s.json'%args.dataset,'r') as f:
qid2type=json.load(f)
model=model.cuda()
batch_size = args.batch_size
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
torch.backends.cudnn.benchmark = True
train_loader = DataLoader(train_dset, batch_size, shuffle=True, num_workers=0)
eval_loader = DataLoader(eval_dset, batch_size, shuffle=False, num_workers=0)
print("Starting training...")
train(model, train_loader, eval_loader, args,qid2type)
if __name__ == '__main__':
main()