-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathlosses-experiments.py
160 lines (134 loc) · 6.17 KB
/
losses-experiments.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
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import dataset
import tensorflow.contrib.layers as layers
import tensorflow.contrib.slim as slim
from tensorflow.contrib.slim import arg_scope
# def playground_loss_function(labels, logits):
# # in rank 2, [elements, classes]
#
# # tf.nn.weighted_cross_entropy_with_logits(labels, logits, weights)
# losses = tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits)
# return losses
#
#
# def prob_to_logit(probs):
# return np.log(probs / (1 - probs))
#
#
# def softmax(x):
# """Same behaviour as tf.nn.softmax in tensorflow"""
# e_x = np.exp(x)
# sum_per_row = np.tile(e_x.sum(axis=1), (x.shape[1], 1)).T
# print('e_x', '\n', e_x)
# print('sum_per_row', '\n', sum_per_row)
# return e_x / sum_per_row
#
#
# def softmax_cross_entropy_loss(labels, logits):
# """Same behaviour as tf.nn.softmax_cross_entropy_with_logits in tensorflow"""
# loss_per_row = - np.sum(labels * np.log(softmax(logits)), axis=1)
# return loss_per_row
def labels_to_info_gain(labels, logits, alpha=0.2):
last_axis = len(logits.shape) - 1
label_idx = np.tile(np.argmax(labels, axis=last_axis), (labels.shape[last_axis], 1)).T
prob_bin_idx = np.tile(range(logits.shape[last_axis]), (labels.shape[0], 1))
# print('label_idx', '\n', label_idx)
# print('probs_idx', '\n', prob_bin_idx)
info_gain = np.exp(-alpha * (label_idx - prob_bin_idx) ** 2)
print('info gain', '\n', info_gain)
return info_gain
def tf_labels_to_info_gain(labels, logits, alpha=0.2):
# int 16 stačí, protože je to index binu pro hloubku
last_axis = len(logits.shape) - 1
label_idx = tf.expand_dims(tf.argmax(labels, axis=last_axis), last_axis)
label_idx = tf.cast(label_idx, dtype=tf.int32)
# expanding back to have size in dim 4 (reduced by argmax)
tiling_shape = list(labels.shape)
tiling_shape[0:last_axis] = [tf.Dimension(1) for i in range(last_axis)]
label_idx = tf.tile(label_idx, tiling_shape)
prob_bin_idx = tf.range(logits.shape[last_axis], dtype=tf.int32)
for i in range(last_axis):
prob_bin_idx = tf.expand_dims(prob_bin_idx, 0)
# prob_bin_idx = tf.transpose(prob_bin_idx)
tiling_shape = list(labels.shape)
tiling_shape[0] = tf.shape(labels)[0]
tiling_shape[last_axis] = tf.Dimension(1)
prob_bin_idx = tf.tile(prob_bin_idx, tiling_shape)
difference = (label_idx - prob_bin_idx) ** 2
difference = tf.cast(difference, dtype=tf.float32)
info_gain = tf.exp(-alpha * difference)
return info_gain
def softmax_loss(labels, logits):
return tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits))
def information_gain_loss(labels, logits, alpha=0.2):
return tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=tf_labels_to_info_gain(labels=labels, logits=logits, alpha=alpha),
logits=logits))
def inference():
with arg_scope([layers.fully_connected],
weights_initializer=layers.xavier_initializer(uniform=True),
biases_initializer=tf.constant_initializer(0.1),
):
x = tf.placeholder(tf.float32, shape=[None, IMG_H, IMG_W, 5], name='x')
l1 = slim.fully_connected(x, num_outputs=8, scope='fc1', activation_fn=tf.nn.relu)
l2 = slim.fully_connected(l1, num_outputs=8, scope='fc2', activation_fn=tf.nn.relu)
l3 = slim.fully_connected(l2, num_outputs=8, scope='fc3', activation_fn=tf.nn.relu)
l4 = slim.fully_connected(l3, num_outputs=5, scope='fc4', activation_fn=None)
return x, l4
if __name__ == '__main__':
IMG_H = 3
IMG_W = 4
with tf.Graph().as_default():
with tf.Session() as sess:
x, logits = inference()
probs = slim.softmax(logits)
labels = tf.placeholder(tf.float32, shape=[None, IMG_H, IMG_W, 5], name='labels')
# loss = softmax_loss(labels=labels, logits=logits)
# log_dir = 'playground/simple_lr_1e-3'
loss = information_gain_loss(labels=labels, logits=logits, alpha=1)
log_dir = 'playground/info_lr_1e-4_a_1'
optimizer = tf.train.AdamOptimizer(learning_rate=1e-4)
train_op = optimizer.minimize(loss)
tf.summary.scalar("loss", loss)
tf.summary.histogram('probs', probs)
tf.summary.histogram('logits', logits)
# tf.summary.scalar("prob0", probs[0, 0])
# tf.summary.scalar("prob1", probs[0, 1])
# tf.summary.scalar("prob2", probs[0, 2])
# tf.summary.scalar("prob3", probs[0, 3])
# tf.summary.scalar("prob4", probs[0, 4])
for var in tf.trainable_variables():
tf.summary.histogram(var.op.name, var)
summary = tf.summary.merge_all() # merge all summaries to dump them for tensorboard
writer = tf.summary.FileWriter(log_dir, sess.graph)
sess.run(tf.global_variables_initializer())
for i in range(5000):
x_val = np.array([
[1, 1, 1, 1, 1],
[0, 0, 0, 1, 1],
[1, 1, 1, 0, 0],
])
x_val = np.expand_dims(x_val, 1)
x_val = np.expand_dims(x_val, 1)
x_val = np.tile(x_val, [1, 3, 4, 1])
labels_val = np.array([
[0, 1, 0, 0, 0],
[0, 0, 1, 0, 0],
[1, 0, 0, 0, 0],
])
labels_val = np.expand_dims(labels_val, 1)
labels_val = np.expand_dims(labels_val, 1)
labels_val = np.tile(labels_val, [1, IMG_H, IMG_W, 1])
_, cost, predicted = sess.run([train_op, loss, probs], feed_dict={
x: x_val,
labels: labels_val,
})
if i % 10 == 0:
summary_str = sess.run(summary, feed_dict={
x: x_val,
labels: labels_val,
})
writer.add_summary(summary_str, i)
print('iteration i:', i)