SPRM/main.py at main · drigoni/SPRM · GitHub

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
import argparse
import os
import random
import warnings

import numpy as np
import torch
from torch.utils.data import DataLoader

from model.dataset_flickr import Flickr30Dataset
from model.dataset_referit import ReferitDataset
# from model import MATnet
from model.model import ConceptNet
from model.model_MATnet import MATnet
from model.loss import SPRMLoss
from model.train_model import train, evaluate
from utils.utils import load_vocabulary, init_net

with warnings.catch_warnings():
	warnings.filterwarnings("ignore", category = FutureWarning)


def get_datasets(args):
	if args.test_set:
		test_split = "test"
	else:
		test_split = "val"

	options = {
		"train_fract": args.train_fract,
		"do_spellchecker": args.do_spellchecker,
		"do_oov": args.do_oov,
		"do_head": args.do_head,
		"do_bert": args.do_bert,
		"do_relations": args.do_relations,
		"do_locations": args.do_locations,
		"relations_strategy": args.relations_strategy,
	}

	if args.dataset == "flickr30k":
		test_dset = Flickr30Dataset(wordEmbedding, test_split, **options)
		if args.dry_run:
			train_dset = test_dset
		else:
			train_dset = Flickr30Dataset(wordEmbedding, "train", **options)
	else:
		test_dset = ReferitDataset(wordEmbedding, test_split, **options)
		if args.dry_run:
			train_dset = test_dset
		else:
			train_dset = ReferitDataset(wordEmbedding, "train", **options)
	return train_dset, test_dset


def parse_args():
	parser = argparse.ArgumentParser()
	parser.add_argument('--batch', type = int, default = 64, help = "batch size for training")
	parser.add_argument('--lr', type = float, default = 1e-5, help = "learning rate")
	parser.add_argument('--eval', action = 'store_true', help = "evaluation mode")
	parser.add_argument('--epochs', type = int, default = 25, help = "training epochs")
	parser.add_argument('--save_name', type = str, default = 'model', help = "name for saved model")
	parser.add_argument('--test_set', action = 'store_true', help = "use test set for evaluation")
	parser.add_argument('--seed', type = int, default = 0, help = "random seed")
	parser.add_argument('--device', type= str, default='cuda', choices=['cuda', 'cpu'])
	parser.add_argument('--dataset', type= str, default='flickr30k', choices=['flickr30k', 'referit'])
	parser.add_argument('--cosine_similarity_strategy', type= str, default='mean', choices=['mean', 'max'])
	parser.add_argument('--loss_strategy', type= str, default='negN', choices=['neg1', 'neg1_min', 'neg1_max', 'negN', 'ce'])
	parser.add_argument('--do_spellchecker', action="store_true", default=False)
	parser.add_argument('--do_oov', action="store_true", default=False)
	parser.add_argument('--do_negative_weighting', action="store_true", default=False)
	parser.add_argument('--do_head', action="store_true", default=False)
	parser.add_argument('--dry_run', action="store_true", default=False)
	parser.add_argument('--emb_dim', type= int, default=300)
	parser.add_argument('--word_emb_dim', type= int, default=300)
	parser.add_argument('--feature_dim', type= int, default=2048)
	parser.add_argument('--cosine_weight', type=float, default=0.5)
	parser.add_argument('--use_att_for_query', action = 'store_true', help = "Disable LSTM for query features and use attention.")
	parser.add_argument('--use_mean_in_loss', action = 'store_true', help = "Consider all the couple <query, box> in the loss calculation.")
	parser.add_argument('--MATnet', action = 'store_true', help = "True when we want to use the original model.")
	parser.add_argument('--train_fract', type=float, default=1.0, help = "Fraction of training set to load for training.")
	parser.add_argument('--similarity_strategy', type=str, default="cosine_similarity", choices=["euclidean_distance", "cosine_similarity"])
	parser.add_argument('--use_head_for_query_embedding', action="store_true", default=False)
	parser.add_argument('--image_net_dropout', type=float, default=0.0)
	parser.add_argument('--query_net_dropout', type=float, default=0.0)
	parser.add_argument('--file', type = str)
	parser.add_argument('--lstm_num_layers', type=int, default=1)
	parser.add_argument('--use_bidirectional_lstm', action="store_true", default=False)
	parser.add_argument('--do_checkpoint', action="store_true", default=False)
	parser.add_argument('--use_head_for_concept_embedding', action="store_true", default=False)
	parser.add_argument('--do_bert', action="store_true", default=False)
	parser.add_argument('--use_minilm_for_query_embedding', action="store_true", default=False)
	parser.add_argument('--loss_sigmoid_slope', type=float, default=1.0)
	parser.add_argument('--do_relations', action="store_true", default=False)
	parser.add_argument('--do_locations', action="store_true", default=False)
	parser.add_argument('--use_wv_freezed', action="store_true", default=False)
	parser.add_argument('--use_spatial_features', action="store_true", default=False)
	parser.add_argument('--use_relations_for_concept_embedding', action="store_true", default=False)
	parser.add_argument('--relations_strategy', type=str, default="none", choices=["none", "pseudo-q", "baseline"])

	# debug mode
	parser.add_argument('--debug', action = 'store_true')
	args = parser.parse_args()
	return args


if __name__ == '__main__':
	args = parse_args()
	print(args)

	# params and seeds
	torch.manual_seed(args.seed)
	random.seed(args.seed)
	np.random.seed(args.seed)

	output_dir = "output/flickr" if args.dataset == "flickr30k" else "output/referit"
	save_path = os.path.join(output_dir, args.save_name + '.pt')
	save_checkpoint = os.path.join(output_dir, args.save_name + '_{}.pt')
	os.makedirs(output_dir, exist_ok=True)

	# config
	wordEmbedding = load_vocabulary(f"data/glove/glove.6B.{args.word_emb_dim}d.txt")
	# get dataset
	train_dset, test_dset = get_datasets(args)
	train_loader = DataLoader(train_dset, batch_size = args.batch, num_workers = 4, drop_last = True, shuffle = True)
	test_loader = DataLoader(test_dset, batch_size = args.batch, num_workers = 4, drop_last = True, shuffle = True)
	# load model
	if args.MATnet:
		model = MATnet(wordEmbedding, args)
	else:
		model = ConceptNet(wordEmbedding, args)
	loss = SPRMLoss(args)

	if args.test_set:
		if args.file:
			init_net(model, args.file)
			model.to(device=args.device)
		score = evaluate(test_loader, model, loss, device_str=args.device)
		print("untrained eval score:", score)
	else:
		best_model = train(model, loss, train_loader, test_loader, args, lr = args.lr, epochs = args.epochs, device_str=args.device, save_checkpoint=save_checkpoint if args.do_checkpoint else None)
		torch.save(best_model.cpu().state_dict(), save_path)
		print("save model to", save_path)