-
Notifications
You must be signed in to change notification settings - Fork 12
/
Copy pathtrain.py
160 lines (124 loc) · 6.3 KB
/
train.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
from model import create_model
from utility.utility import load_training_data, load_validation_data
from utility.language_encoder import LanguageEncoder
import numpy as np
from matplotlib import pyplot as plt
from keras.callbacks import TensorBoard
import keras.backend as K
from scipy.misc import imresize
import skimage.transform
import matplotlib.pyplot as plt
from keras.models import load_model
from kulc.attention import ExternalAttentionRNNWrapper
import pathlib
from keras.optimizers import adam
LOAD_MODEL = False
def generator(batch_size, captions, get_image):
while True:
batch_indices = np.random.randint(0, len(captions), size=batch_size, dtype=np.int)
batch_image_features = np.empty((len(batch_indices), 14*14, 512))
for i, j in enumerate(batch_indices):
batch_image_features[i] = get_image(str(j)).value.reshape((14*14, 512))
batch_captions = [captions[item] for item in batch_indices]
batch_captions = [x[np.random.randint(0, len(x))][:MAXIMUM_CAPTION_LENGTH-1] for x in batch_captions]
input_captions = [[le.transform_word("<START>")] + x for x in batch_captions]
output_captions = [x + [le.transform_word("<STOP>")] for x in batch_captions]
input_captions = np.array([x+[le.transform_word("<NULL>")]*(MAXIMUM_CAPTION_LENGTH-len(x)) for x in input_captions]).astype(np.float32)
output_captions = one_hot_encode(output_captions, MAXIMUM_CAPTION_LENGTH, MAXIMUM_VOCABULARY_SIZE)
batch_image_features = np.array(batch_image_features, dtype=np.float32)
x_data = {
"image_features_input": batch_image_features,
"captions_input": input_captions
}
y_data = {
"output": output_captions
}
yield (x_data, y_data)
def one_hot_encode(data, MAXIMUM_CAPTION_LENGTH, n_classes):
result = np.zeros((len(data), MAXIMUM_CAPTION_LENGTH, n_classes))
for i, item in enumerate(data):
for j, word in enumerate(item):
result[i, j, word] = 1.0
for k in range(j+1, MAXIMUM_CAPTION_LENGTH):
result[i, k, le.transform_word("<NULL>")] = 1.0
return result
def inference(image_features, plot_attention):
image_features = np.array([image_features])
state_h, state_c = initial_state_inference_model.predict(image_features)
caption = [word_index_map["<START>"]]
attentions = []
current_word = None
for t in range(MAXIMUM_CAPTION_LENGTH):
caption_array = np.array(caption).reshape(1, -1)
output, state_h, state_c, attention = inference_model.predict([image_features, caption_array, state_h, state_c])
attentions.append(attention[0, -1].reshape((14, 14)))
current_word = np.argmax(output[0, -1])
caption.append(current_word)
if current_word == word_index_map["<STOP>"]:
break
sentence = [index_word_map[i] for i in caption[1:]]
if plot_attention:
print(len(attentions))
x = int(np.sqrt(len(attentions)))
y = int(np.ceil(len(attentions) / x))
_, axes = plt.subplots(y, x, sharex="col", sharey="row")
axes = axes.flatten()
for i in range(len(attentions)):
atn = skimage.transform.pyramid_expand(attentions[i], upscale=16, sigma=20)
axes[i].set_title(sentence[i])
axes[i].imshow(atn, cmap="gray")
plt.show()
return " ".join(sentence) + " ({0})".format(len(caption)-1)
pathlib.Path('./models').mkdir(exist_ok=True)
MAXIMUM_VOCABULARY_SIZE = 10000
EMBEDDING_SIZE = 512 # 1024
MAXIMUM_CAPTION_LENGTH = 16
"""
filters='!"#$%&()*+,-./:;<=>?@[\\]^_`{|}~\t\n'
translate_dict = dict((c, " ") for c in filters)
translate_map = str.maketrans(translate_dict)
captions_train_raw, get_image_features_train = load_training_data(MAXIMUM_CAPTION_LENGTH)
captions_val_raw, get_image_features_val = load_validation_data(MAXIMUM_CAPTION_LENGTH)
for i, entry in enumerate(captions_val_raw):
for j, item in enumerate(entry):
captions_val_raw[i][j] = item.translate(translate_map).lower()
for i, entry in enumerate(captions_train_raw):
for j, item in enumerate(entry):
captions_train_raw[i][j] = item.translate(translate_map).lower()
word_index_map, index_word_map = create_vocabulary(MAXIMUM_VOCABULARY_SIZE, captions_train_raw)
MAXIMUM_VOCABULARY_SIZE = len(word_index_map)
captions_train = encode_annotations(captions_train_raw, word_index_map, MAXIMUM_CAPTION_LENGTH)
captions_val = encode_annotations(captions_val_raw, word_index_map, MAXIMUM_CAPTION_LENGTH)
"""
captions_train_raw, get_image_features_train = load_training_data(MAXIMUM_CAPTION_LENGTH)
captions_val_raw, get_image_features_val = load_validation_data(MAXIMUM_CAPTION_LENGTH)
le = LanguageEncoder(MAXIMUM_VOCABULARY_SIZE)
captions_train = le.fit_transform(captions_train_raw)
captions_val = le.transform(captions_val_raw)
le.save("./models/language.pkl")
def masked_categorical_crossentropy(y_true, y_pred):
mask_value = le._word_index_map["<NULL>"]
y_true_id = K.argmax(y_true)
mask = K.cast(K.equal(y_true_id, mask_value), K.floatx())
mask = 1.0 - mask
loss = K.categorical_crossentropy(y_true, y_pred) * mask
# take average w.r.t. the number of unmasked entries
return K.sum(loss) / K.sum(mask)
training_model, inference_model, initial_state_inference_model = create_model(le._vocabulary_size, EMBEDDING_SIZE, None, 14*14, 512)
training_model.compile(adam(0.001), loss=masked_categorical_crossentropy, metrics=["accuracy"])
batch_size = 64
def train(epochs=100):
tbCallback = TensorBoard(log_dir='./logs', histogram_freq=0, batch_size=batch_size, write_graph=True, write_grads=False, write_images=False)
history = training_model.fit_generator(generator(batch_size=batch_size, captions=captions_train, get_image=get_image_features_train), steps_per_epoch=len(captions_train)//batch_size, epochs=epochs, verbose=1, callbacks=[tbCallback])
training_model.save("./models/sat_train_{0}.h5".format(epochs))
inference_model.save("./models/sat_inf_{0}.h5".format(epochs))
initial_state_inference_model.save("./models/sat_inf_init_{0}.h5".format(epochs))
for key in history.history.keys():
f = plt.figure()
data = history.history[key]
plt.plot(data)
plt.show()
epochs = input("Number of epochs: ")
epochs = int(epochs)
train(epochs=epochs)
input("done. <read key>")