-
Notifications
You must be signed in to change notification settings - Fork 3
/
intrinsic_dimensions_measurement.py
175 lines (133 loc) · 6.71 KB
/
intrinsic_dimensions_measurement.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
168
169
170
171
172
173
174
175
import logging
import time
import numpy as np
import tensorflow as tf
import sys
from utils import prepare_mnist_dataset, make_session, configure_logging
configure_logging()
MAX_EPOCHS = 100
class IntrinsicDimensionExperiment:
"""Measure intrinsic dimension in model for MNIST.
"""
def __init__(self, d, layer_sizes, clip_norm=None):
self.sess = make_session()
self.clip_norm = clip_norm
dataset, (_, _), (self.x_test, self.y_test) = prepare_mnist_dataset(
batch_size=128, train_sample_size=6000, test_sample_size=1000)
self.iterator = dataset.make_initializable_iterator()
self.batch_input, self.batch_label = self.iterator.get_next()
self.hidden_layers = layer_sizes
self.n_layers = len(self.hidden_layers) + 1 # plus the output layer
self.shape_per_layer = [(28 * 28, self.hidden_layers[0])] + [
(self.hidden_layers[i], self.hidden_layers[i + 1])
for i in range(len(self.hidden_layers) - 1)
] + [(self.hidden_layers[-1], 10)]
logging.info(f"shape per layer: {self.shape_per_layer}")
self.d = d
self.D = np.sum((h + 1) * w for h, w in self.shape_per_layer)
logging.info(f"Experiment config: d={d} D={self.D}")
self.loss, self.accuracy = self.build_network(self.batch_input, self.batch_label)
for t in tf.trainable_variables():
print(t)
np.random.seed(int(time.time()))
def build_network(self, input_ph, label_ph):
self.transforms = self.sample_transform_matrices()
self.subspace = tf.get_variable("subspace", shape=(self.d, 1), trainable=True,
initializer=tf.zeros_initializer())
with tf.variable_scope("mnist_dense_nn", reuse=False):
label_ohe = tf.one_hot(label_ph, 10, dtype=tf.float32)
out = tf.reshape(tf.cast(input_ph, tf.float32), (-1, 28 * 28))
for i, (h, w) in enumerate(self.shape_per_layer):
if i > 0:
# No dropout on the input layer.
out = tf.nn.dropout(out, keep_prob=0.9)
weights = tf.get_variable(f'w_{i}', shape=(h, w), trainable=False,
initializer=tf.glorot_uniform_initializer())
new_weights = tf.stop_gradient(weights) + tf.reshape(tf.matmul(
tf.stop_gradient(self.transforms[i][0]), self.subspace), weights.shape)
biases = tf.get_variable(f'b_{i}', shape=(w,), trainable=False,
initializer=tf.zeros_initializer())
new_biases = tf.stop_gradient(biases) + tf.reshape(tf.matmul(
tf.stop_gradient(self.transforms[i][1]), self.subspace), biases.shape)
out = tf.matmul(out, new_weights) + new_biases
if i < self.n_layers - 1:
out = tf.nn.relu(out)
logits = out
pred_probas = tf.nn.softmax(logits)
pred_labels = tf.cast(tf.argmax(pred_probas, 1), tf.uint8)
loss = tf.losses.softmax_cross_entropy(label_ohe, logits)
# loss = tf.losses.mean_squared_error(label_ohe, preds)
accuracy = tf.reduce_sum(
tf.cast(tf.equal(pred_labels, label_ph), tf.float32)
) / tf.cast(tf.shape(label_ph)[0], tf.float32)
return loss, accuracy
def sample_transform_matrices(self):
"""Matrix P in the paper of size (D, d)
Columns of P are normalized to unit length.
"""
matrix = np.random.randn(self.D, self.d).astype(np.float32)
for i in range(self.d):
# each column is normalized to have unit 1.
matrix[:, i] /= np.linalg.norm(matrix[:, i])
# split P according to num. params per layer.
w_matrices = []
b_matrices = []
offset = 0
for i, (h, w) in enumerate(self.shape_per_layer):
w_matrix_values = matrix[offset:offset + h * w]
w_matrix = tf.Variable(w_matrix_values, dtype=tf.float32, name=f'w_matrix_{i}',
trainable=False)
w_matrices.append(w_matrix) # for weights
offset += h * w
b_matrix_values = matrix[offset:offset + w]
b_matrix = tf.Variable(b_matrix_values, dtype=tf.float32, name=f'b_matrix_{i}',
trainable=False)
b_matrices.append(b_matrix) # for weights
offset += w
logging.info(f"shape of transform matrices for weights: {[m.shape for m in w_matrices]}")
logging.info(f"shape of transform matrices for biases: {[m.shape for m in b_matrices]}")
return list(zip(w_matrices, b_matrices))
def _initialize(self):
self.sess.run(self.iterator.initializer)
self.sess.run(tf.global_variables_initializer())
def _get_eval_results(self):
feed_dict = {self.batch_input: self.x_test, self.batch_label: self.y_test}
return self.sess.run([self.loss, self.accuracy], feed_dict=feed_dict)
def _get_train_op(self, lr):
optimizer = tf.train.AdamOptimizer(lr)
grads_tvars = optimizer.compute_gradients(self.loss, var_list=[self.subspace])
if self.clip_norm:
grads_tvars = [(tf.clip_by_norm(g, self.clip_norm), v) for g, v in grads_tvars]
train_op = optimizer.apply_gradients(grads_tvars)
return train_op, grads_tvars
def train(self, lr):
lr_ph = tf.placeholder(tf.float32, shape=(), name='learning_rate')
train_op, grads_tvars = self._get_train_op(lr_ph)
self._initialize()
step = 0
epoch = 0
while epoch <= MAX_EPOCHS:
while True:
try:
_, loss, acc = self.sess.run([train_op, self.loss, self.accuracy],
feed_dict={lr_ph: lr})
step += 1
if step % 100 == 0:
logging.info(f"[step:{step}|epoch:{epoch}] loss={loss} acc={acc}")
except tf.errors.OutOfRangeError:
# complete one training epoch.
self.sess.run(self.iterator.initializer)
break
epoch += 1
if epoch % 50 == 0:
lr *= 0.5
_, final_eval_acc = self._get_eval_results()
logging.info(f"[final] d={self.d} eval_acc={final_eval_acc}")
return epoch, final_eval_acc
if __name__ == '__main__':
net_layers = list(map(int, sys.argv[1].split(',')))
d = int(sys.argv[2])
logging.info(f"Testing net_layers={net_layers} d={d} ")
exp = IntrinsicDimensionExperiment(d, net_layers)
_, final_eval_acc = exp.train(0.001)
print(final_eval_acc)