-
Notifications
You must be signed in to change notification settings - Fork 9
/
main_fmnist_capsnet.py
167 lines (148 loc) · 6.91 KB
/
main_fmnist_capsnet.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
import tensorflow as tf
import numpy as np
import capsule as caps
from matplotlib import pyplot as plt
from tensorflow.python.keras._impl.keras.preprocessing.image import ImageDataGenerator
from load_data import load_data
from utils import reconMashup, reconstruction_loss, margin_loss, decoder_nn, mask_one
dataset_size = 60000
epsilon = 1e-9
regularization = True
lambda_reg = 0.4 # ~28*28*0.0005
iter_routing = 2 # routing 2 in this implementation corresponds to routing 3 in the paper
num_epochs = 100
batch_size = 128
steps_per_epoch = dataset_size/batch_size
steps_train = steps_per_epoch*num_epochs
start_lr = 0.001
decay_steps = steps_per_epoch
decay_rate = 0.9
plot_num = 100
config = tf.estimator.RunConfig(save_summary_steps=100, log_step_count_steps=100)
model_dir = "/tmp/fmnist/r2_reg1"
mapfn_parallel_iterations = batch_size
def caps_model_fn(features, labels, mode):
"""Model function for CNN."""
# Input Layer
# Reshape X to 4-D tensor: [batch_size, width, height, channels]
# Fashion MNIST images are 28x28 pixels, and have one color channel
input_layer = tf.reshape(features["x"], [-1, 28, 28, 1])
# A little bit cheaper version of the capsule network in: Dynamic Routing Between Capsules
# Std. convolutional layer
conv1 = tf.layers.conv2d(
inputs=input_layer,
filters=256,
kernel_size=[9, 9],
padding="valid",
activation=tf.nn.relu,
name="ReLU_Conv1")
conv1 = tf.expand_dims(conv1, axis=-2)
# Convolutional capsules, no routing as the dimension of the units of previous layer is one
primarycaps = caps.conv2d(conv1, 32, 8, [9,9], strides=(2,2), name="PrimaryCaps")
primarycaps = tf.reshape(primarycaps, [-1, primarycaps.shape[1].value*primarycaps.shape[2].value*32, 8])
# Fully connected capsules with routing by agreement
digitcaps = caps.dense(primarycaps, 10, 16, iter_routing=iter_routing, mapfn_parallel_iterations=mapfn_parallel_iterations, name="DigitCaps")
# The length of the capsule activation vectors encodes the probability of an entity being present
lengths = tf.sqrt(tf.reduce_sum(tf.square(digitcaps),axis=2) + epsilon, name="Lengths")
# Predictions for (PREDICTION mode)
predictions = {
# Generate predictions (for PREDICT and EVAL mode)
"classes": tf.argmax(lengths, axis=1),
# Add `softmax_tensor` to the graph. It is used for PREDICT and by the
# `logging_hook`.
"probabilities": tf.nn.softmax(lengths, name="Softmax")
}
if regularization:
masked_digitcaps_pred = mask_one(digitcaps, lengths, is_predicting=True)
with tf.variable_scope(tf.get_variable_scope()):
reconstruction_pred = decoder_nn(masked_digitcaps_pred)
predictions["reconstruction"] = reconstruction_pred
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)
# Calculate Loss (for both TRAIN and EVAL modes)
onehot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32), depth=10)
loss = margin_loss(onehot_labels, lengths)
tf.summary.scalar("margin_loss", loss)
if regularization:
masked_digitcaps = mask_one(digitcaps, onehot_labels)
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
reconstruction = decoder_nn(masked_digitcaps)
rec_loss = reconstruction_loss(input_layer, reconstruction)
tf.summary.scalar("reconstruction_loss", rec_loss)
loss += lambda_reg * rec_loss
# Configure the Training Op (for TRAIN mode)
if mode == tf.estimator.ModeKeys.TRAIN:
# Summary hook
summary_hook = tf.train.SummarySaverHook(
save_steps=config.save_summary_steps,
output_dir=model_dir,
summary_op=tf.summary.merge_all())
global_step = tf.Variable(0, trainable=False)
learning_rate = tf.train.exponential_decay(start_lr, global_step, decay_steps, decay_rate)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
train_op = optimizer.minimize(
loss=loss, global_step=tf.train.get_global_step())
return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op, training_hooks=[summary_hook])
# Add evaluation metrics (for EVAL mode)
eval_metric_ops = {
"accuracy": tf.metrics.accuracy(labels=labels, predictions=predictions["classes"])
}
return tf.estimator.EstimatorSpec(mode=mode, loss=loss, eval_metric_ops=eval_metric_ops)
def main(unused_argv):
tf.logging.set_verbosity(tf.logging.INFO)
(x_train, y_train), (x_test, y_test) = load_data()
labels = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat', 'Sandal',
'Shirt', 'Sneaker', 'Bag', 'Ankle boot']
img_rows, img_cols = x_train[0].shape
x_train = x_train.astype(np.float32)
x_test = x_test.astype(np.float32)
x_train /= 255
x_test /= 255
x_train = np.reshape(x_train, (-1, img_rows*img_cols))
x_test = np.reshape(x_test, (-1, img_rows*img_cols))
# Load training and eval data
train_data = x_train # Returns np.array
train_labels = y_train
eval_data = x_test # Returns np.array
eval_labels = y_test
# Create the Estimator
mnist_classifier = tf.estimator.Estimator(
model_fn=caps_model_fn, config=config,
model_dir=model_dir)
# Train the model #
# Data Augumentation
generator = ImageDataGenerator(width_shift_range=0.1, height_shift_range=0.1)
train_data_im = np.reshape(train_data, [-1, 28, 28, 1])
flow = generator.flow(train_data_im, train_labels, batch_size=num_epochs*dataset_size)
train_data, train_labels = flow.next()
train_data = np.reshape(train_data, [-1, 28*28])
train_input_fn = tf.estimator.inputs.numpy_input_fn(
x={"x": train_data},
y=train_labels,
batch_size=batch_size,
num_epochs=None,
shuffle=True)
mnist_classifier.train(input_fn=train_input_fn, steps=steps_train)
# Evaluate the model and print results
eval_input_fn = tf.estimator.inputs.numpy_input_fn(
x={"x": eval_data},
y=eval_labels,
batch_size=batch_size,
num_epochs=1,
shuffle=False)
eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn)
print(eval_results)
if regularization:
# do some predictions and reconstructions
pred_input_fn = tf.estimator.inputs.numpy_input_fn(
x={"x": eval_data[:plot_num]},
batch_size=batch_size,
num_epochs=1,
shuffle=False)
predictions = mnist_classifier.predict(input_fn=pred_input_fn)
prediction_pics = [np.reshape(p['reconstruction'], (28,28)) for p in predictions]
eval_pics = np.reshape(eval_data[:plot_num], (-1, 28, 28))
plt.imshow(reconMashup(eval_pics, prediction_pics), cmap='gray')
plt.axis('off')
if __name__ == "__main__":
tf.app.run()