Image2Text/train.py at master · koren-v/Image2Text · 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
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
import argparse

import nltk
from data_loader import get_loader
from torchvision import transforms


import torch
import torch.nn as nn

from model import EncoderCNN, DecoderRNN
from learner import Learner

# Define a transform to pre-process the training images.
transform_train = transforms.Compose([
    transforms.Resize(256),                          # smaller edge of image resized to 256
    transforms.RandomCrop(224),                      # get 224x224 crop from random location
    transforms.RandomHorizontalFlip(),               # horizontally flip image with probability=0.5
    transforms.ToTensor(),                           # convert the PIL Image to a tensor
    transforms.Normalize((0.485, 0.456, 0.406),      # normalize image for pre-trained model
                         (0.229, 0.224, 0.225))])


device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

if __name__ == "__main__":

    nltk.download('punkt')

    parser = argparse.ArgumentParser()
    parser.add_argument("num_epochs", type=int,
                        help="num of epoches")
    parser.add_argument("--dataset", default='coco')
    parser.add_argument("--encoder_lr", type=float)
    parser.add_argument("--decoder_lr", type=float)
    parser.add_argument("--accum_step", type=int, default=1)
    parser.add_argument("--stage", default='default_stage')
    parser.add_argument("--cnn", default='resnet101')
    parser.add_argument("--vocab_from_file", action="store_true")
    parser.add_argument("--hidden_size", type=int, default=768)
    parser.add_argument("--num_layers", type=int, default=2)
    parser.add_argument("--dropout", type=float, default=0.0)
    parser.add_argument("--weight_decay", type=float, default=0.0)
    parser.add_argument("--freeze_glove", action="store_true")
    parser.add_argument("--adam", action="store_true")
    parser.add_argument("--scheduler", action="store_true")
    parser.add_argument("--load_model", action="store_true")
    parser.add_argument("--unfreeze_encoder", type=int)
    parser.add_argument("--batch_size", type=int, default=512)
    args = parser.parse_args()

    batch_size = args.batch_size
    num_layers = args.num_layers
    hidden_size = args.hidden_size
    dropout = args.dropout
    vocab_threshold = 3
    vocab_from_file = args.vocab_from_file

    train_data_loader = get_loader(transform=transform_train,
                                   mode='train',
                                   batch_size=batch_size,
                                   vocab_threshold=vocab_threshold,
                                   vocab_from_file=vocab_from_file,
                                   dataset=args.dataset)

    val_data_loader = get_loader(transform=transform_train,
                                 mode='val',
                                 batch_size=batch_size,
                                 vocab_from_file=True,
                                 dataset=args.dataset)

    dataloader_dict = {'train':train_data_loader, 'val': val_data_loader}
    vocab_size = len(train_data_loader.dataset.vocab)
    weight_matrix = train_data_loader.dataset.vocab.weight_matrix
    _, embed_size = weight_matrix.shape

    cnn = args.cnn

    encoder = EncoderCNN(embed_size, cnn)
    encoder = encoder.to(device)

    decoder = DecoderRNN(weight_matrix,
                         hidden_size=hidden_size,
                         vocab_size=vocab_size,
                         num_layers=num_layers,
                         dropout=dropout,
                         non_trainable=args.freeze_glove)

    decoder = decoder.to(device)

    if args.load_model:
        name = input('Type name of encoder/decoder')
        last_epoch = int(name[1])
        if torch.cuda.is_available():
            encoder.load_state_dict(torch.load('./models/encoder'+name+'.pth'))
            decoder.load_state_dict(torch.load('./models/decoder'+name+'.pth'))
        else:
            encoder.load_state_dict(torch.load('./models/encoder'+name+'.pth', map_location=torch.device('cpu')))
            decoder.load_state_dict(torch.load('./models/encoder'+name+'.pth', map_location=torch.device('cpu')))

    num_epochs = args.num_epochs
    grad_accumulation_step = args.accum_step
    decoder_lr = args.decoder_lr
    encoder_lr = args.encoder_lr
    stage = args.stage
    last_epoch = None

    if args.unfreeze_encoder:
        encoder.unfreeze_encoder(args.unfreeze_encoder)
        encoder_params = []
        for name, param in encoder.named_parameters():
            if param.requires_grad:
                encoder_params.append(param)
                print("\t", name)
    else:
        encoder_params = list(encoder.linear.parameters()) + list(encoder.bn1.parameters())

    criterion = nn.CrossEntropyLoss()

    if args.adam:
        optimizer_name = 'Adam'
        optimizer = torch.optim.Adam(
            [
                {"params": decoder.parameters(), "lr": decoder_lr},
                {"params": encoder_params, "lr": encoder_lr},
            ], weight_decay=args.weight_decay)
    else:
        optimizer_name = 'ASGD'
        optimizer = torch.optim.ASGD(
            [
                {"params":decoder.parameters(),"lr": decoder_lr},
                {"params":encoder_params, "lr": encoder_lr},
            ], weight_decay = args.weight_decay)

    if args.scheduler:
        scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
                                                                patience=1,
                                                                verbose=True)
    else:
        scheduler = None

    model = {'encoder': encoder, 'decoder': decoder}
    hyper_params = {'embed_size': embed_size,
                    'batch_size': batch_size,
                    'num_layers': num_layers,
                    'dropout': dropout,
                    'weight_decay': args.weight_decay,
                    'optimizer': optimizer_name,
                    'hidden_size': hidden_size,
                    'cnn': cnn,
                    'decoder_lr': decoder_lr,
                    'encoder_lr': encoder_lr,
                    'grad_accumulation_step': grad_accumulation_step}

    learner = Learner(model, criterion, optimizer, dataloader_dict,
                      num_epochs, device, stage, hyper_params,
                      scheduler=scheduler, last_epoch=last_epoch,
                      grad_accumulation_step=grad_accumulation_step)

    print('Start Training!')
    learner.fit()