image_caption/train.py at master · MaticsL/image_caption · 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
162
163
164
165
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import sys
import pickle

import csv
import numpy as np

from model import image_caption_model
from joblib import Parallel, delayed
import time
from keras.utils.layer_utils import print_summary


def gene_word_caption(captions):
    word2idx = dict()
    idx2word = dict()
    word_count = 0

    ret = []
    for caption in captions:
        caption = caption.replace(',', ' ,')
        caption = caption.replace(';', ' ;')
        caption = caption.replace(':', ' :')
        caption = caption.replace('.', ' .')
        tokens = caption.split(' ')
        res = []
        for word in tokens:
            if word == '':
                continue
            if word not in word2idx:
                word2idx[word] = word_count
                idx2word[word_count] = word
                word_count += 1
            res.append(word2idx[word])
        ret.append(res)

    # add start sign
    word2idx['+'] = word_count
    idx2word[word_count] = '+'
    word_count += 1

    return ret, word2idx, idx2word, word_count


def gen_batch_in_thread(img_map, df_cap, vocab_size, n_jobs=4,
                        size_per_thread=32):
    imgs, curs, nxts, seqs, vhists = [], [], [], [], []
    returns = Parallel(n_jobs=4, backend='threading')(
                            delayed(generate_batch)
                            (img_train, df_cap, vocab_size, size=size_per_thread)
                            for i in range(0, n_jobs))

    for triple in returns:
        imgs.extend(triple[0])
        curs.extend(triple[1])
        nxts.extend(triple[2])
        seqs.extend(triple[3])
        vhists.extend(triple[4])

    return np.array(imgs), np.array(curs).reshape((-1, 1)), np.array(nxts), \
        np.array(seqs), np.array(vhists)


def generate_batch(img_map, df_cap, vocab_size, size=32, max_caplen=28):
    imgs, curs, nxts, seqs, vhists = [], [], [], [], []

    for idx in np.random.randint(len(df_cap), size=size):
        row = df_cap[idx]
        cap = row['caption']

        img = img_map[row['img_id']]

        vhist = np.zeros((len(cap)-1, vocab_size))

        for i in range(1, len(cap)):
            seq = np.zeros((max_caplen))
            nxt = np.zeros((vocab_size))
            nxt[cap[i]] = 1
            curs.append(cap[i-1])
            seq[i-1] = 1

            if i < len(cap)-1:
                vhist[i, :] = np.logical_or(vhist[i, :], vhist[i-1, :])
                vhist[i, cap[i-1]] = 1

            nxts.append(nxt)
            imgs.append(img)
            seqs.append(seq)

        vhists.extend(vhist)

    return imgs, curs, nxts, seqs, vhists


if __name__ == '__main__':
    # initialization
    n_jobs = 64
    size_per_thread = 64
    mdl_path = 'weights/'

    pkl_count = 0
    img_train = {}
    while True:
        pkl_count += 1
        try:
            with open('train/train_img2048_%d.pkl' % pkl_count, 'rb') as f:
                img_train.update(pickle.load(f))
        except FileNotFoundError:
            break
    print('got %d image context vectors' % len(img_train))

    anns = csv.reader(open("train/anns.csv"))
    anns = [row for row in anns if row[1] != 'caption']
    print('got %d training captions' % len(anns))

    captions = ['+ ' + row[1] for row in anns]
    captions, word2idx, idx2word, vocab_size = gene_word_caption(captions)
    print('got %d words, saving word translate dict...' % vocab_size)
    with open('./train/word2idx.pkl', 'wb') as f:
        pickle.dump(word2idx, f)
    with open('./train/idx2word.pkl', 'wb') as f:
        pickle.dump(idx2word, f)

    img_ids = [int(row[3]) for row in anns]

    df_cap = []
    for i in range(0, len(img_ids)):
        df_cap.append(
            dict(img_id=img_ids[i], caption=captions[i])
        )

    model = image_caption_model(vocab_size=vocab_size)

    if len(sys.argv) >= 2:
        print('load weights from : {}'.format(sys.argv[1]))
        model.load_weights(sys.argv[1])

    # insert ur version name here
    version = 'v1.0.0'
    batch_num = 70
    hist_loss = []

    for i in range(0, 100):
        for j in range(0, batch_num):
            s = time.time()
            img1, cur1, nxt1, seq1, vhists1 = gen_batch_in_thread(
                img_train, df_cap, vocab_size, n_jobs=n_jobs,
                size_per_thread=size_per_thread)
            hist = model.fit([img1, cur1, seq1, vhists1], nxt1,
                             batch_size=n_jobs * size_per_thread,
                             nb_epoch=1, verbose=0, shuffle=True)

            print("epoch {0}, batch {1} - training loss : {2}".format(
                i, j, hist.history['loss'][-1]))
            # record the training history
            hist_loss.extend(hist.history['loss'])

            if j % int(batch_num / 2) == 0:
                print('check point arrived, saving...')
                m_name = "{0}{1}_{2}_{3}_{4}_{5}.h5".format(
                    mdl_path, version, i, j, time.time(), str(vocab_size))
                model.save_weights(m_name)