-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathflows_test.py
More file actions
214 lines (161 loc) · 9.5 KB
/
Copy pathflows_test.py
File metadata and controls
214 lines (161 loc) · 9.5 KB
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
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
# -*- coding: utf-8 -*-
import unittest
import flows
import numpy as np
import tensorflow as tf
import tensorflow_probability as tfp
import logging
logger = logging.getLogger(__name__)
import copy
def forward(flow, samples):
return flow(samples)
def backward(flow, samples):
return flow.reverse(samples)
class TestFlows(unittest.TestCase):
def __init__(self, *args):
super(TestFlows, self).__init__(*args)
#tf.keras.backend.set_floatx('float64')
np.random.seed(13)
tf.random.set_seed(13)
N = 17
K = 23 # must be prime to work with scale!=1
base = tfp.distributions.OneHotCategorical(logits=np.random.randn(N,K))
logger.debug("[init] base=%s" % (base,) )
self.base_samples = tf.cast(base.sample(256), dtype='float32')
logger.debug("[init] samples=%s (%s)" % (self.base_samples.shape, self.base_samples.dtype))
self.empty_samples = tf.zeros((256,N,K), dtype=self.base_samples.dtype)
logger.debug("[init] empty samples=%s (%s)" % (self.empty_samples.shape, self.empty_samples.dtype))
s = np.zeros((100,N,K))
s[:,:,13] = 1.
self.equal_samples = tf.constant(s, dtype=self.base_samples.dtype)
logger.debug("[init] equal samples=%s (%s)" % (self.equal_samples.shape, self.equal_samples.dtype))
s = np.zeros((100,N,K))
s[:,:,22] = 1.
self.equal_samples2 = tf.constant(s, dtype=self.base_samples.dtype)
logger.debug("[init] equal samples2=%s (%s)" % (self.equal_samples2.shape, self.equal_samples2.dtype))
self.N = N
self.K = K
self.layers_specs = [
[("M",[12, 16])],
[("M",[N])],
[("M",[12, 16]), ("M",[16, 12])],
[("MEd", N)],
[("M2", N)],
[("F", None)],
[("FU", None)],
[("FU", None), ("F", None)],
]
def setUp(self):
np.random.seed(13)
tf.random.set_seed(13)
def test_discrete_flows_invertability(self):
"""Tests various flows by passing samples both directions and checking if input is recovered."""
for i, layers in enumerate(self.layers_specs):
logger.debug("#"*30+" [%i/%i] testing invertability of %s" % (i+1, len(self.layers_specs), layers))
flow = flows.DiscreteFlow(self.N, self.K, temperature=5.0, layers=copy.deepcopy(layers))
self.assert_forward_reverse(flow, str(layers))
def test_discrete_flows_deterministic_forward(self):
"""Tests various flows by passing set of exactly the same samples and checking if result is the same."""
N = self.N
for i, layers in enumerate(self.layers_specs):
logger.debug("#"*30+" [%i/%i] testing repeated samples (forward) %s" % \
(i+1, len(self.layers_specs), layers))
flow = flows.DiscreteFlow(self.N, self.K, temperature=5.0, layers=copy.deepcopy(layers))
out = flow(self.equal_samples)
self.assertTrue( sum([bool(tf.reduce_all(out[n] == out[n+1])) for n in range(N-1)])==N-1,
"all samples the same")
out = flow(self.equal_samples2)
self.assertTrue( sum([bool(tf.reduce_all(out[n] == out[n+1])) for n in range(N-1)])==N-1,
"all samples the same")
def test_discrete_flows_deterministic_reverse(self):
"""Tests various flows by passing set of exactly the same samples and checking if result is the same."""
N = self.N
for i, layers in enumerate(self.layers_specs):
logger.debug("#"*30+" [%i/%i] testing repeated samples (reverse) %s" % \
(i+1, len(self.layers_specs), layers))
flow = flows.DiscreteFlow(self.N, self.K, temperature=5.0, layers=copy.deepcopy(layers))
out = flow.reverse(self.equal_samples)
self.assertTrue( sum([bool(tf.reduce_all(out[n] == out[n+1])) for n in range(N-1)])==N-1,
"all samples the same")
out = flow.reverse(self.equal_samples2)
self.assertTrue( sum([bool(tf.reduce_all(out[n] == out[n+1])) for n in range(N-1)])==N-1,
"all samples the same")
def test_discrete_flows_different_inputs(self):
"""Tests various flows by passing two different inputs and checking if outputs differ."""
N = self.N
for i, layers in enumerate(self.layers_specs):
logger.debug("#"*30+" [%i/%i] testing repeated samples (reverse) %s" % \
(i+1, len(self.layers_specs), layers))
flow = flows.DiscreteFlow(self.N, self.K, temperature=5.0, layers=copy.deepcopy(layers))
out = flow(self.equal_samples[0:1])
out2 = flow(self.equal_samples2[0:1])
self.assertTrue(out.shape==out2.shape)
self.assertTrue( sum(bool(tf.reduce_any(out[0,n]!=out2[0,n])) for n in range(self.N))>0, \
"at least one dim in outputs must differ" )
def test_temperature_update(self):
"""Test if temperature updates affect gradients."""
flow = flows.DiscreteFlow(self.N, self.K,
temperature=10., layers=[("M",[256, 256]), ("F", None), ("M",[])])
with tf.GradientTape() as tape:
loss = -tf.reduce_sum(forward(flow, self.base_samples))
g1 = tape.gradient(loss, flow.trainable_variables)
with tf.GradientTape() as tape:
flow.temperature = 1.0
loss = -tf.reduce_sum(forward(flow, self.base_samples))
g2 = tape.gradient(loss, flow.trainable_variables)
with tf.GradientTape() as tape:
flow.temperature = 0.1
loss = -tf.reduce_sum(forward(flow, self.base_samples))
g3 = tape.gradient(loss, flow.trainable_variables)
difference_g1g2 = sum( np.sum(abs(e1-e2)) for e1, e2 in zip(g1,g2) )
difference_g1g3 = sum( np.sum(abs(e1-e2)) for e1, e2 in zip(g1,g3) )
logger.debug("Total gradient difference for temperature 10 vs 1: %s" % difference_g1g2)
logger.debug("Total gradient difference for temperature 10 vs 0.1: %s" % difference_g1g3)
self.assertNotEqual(difference_g1g2, 0.0)
self.assertNotEqual(difference_g1g3, 0.0)
self.assertNotEqual(difference_g1g3, difference_g1g2)
def assert_forward_reverse(self, flow, msg="", precision=4):
logger.debug("[assert_forward_reverse] flow = %s" % flow)
z = forward(flow, self.base_samples)
error = tf.reduce_max( tf.math.abs(backward(flow, z)-self.base_samples) ).numpy()
num_errors = len( np.nonzero( tf.math.abs(backward(flow, z)-self.base_samples).numpy() > 0.0001 )[0] )
logger.debug(" worst case error for forward-reverse %s: %s (%s/%s errors)" % \
(msg, error, num_errors, len(self.base_samples)) )
self.assertAlmostEqual(error, 0., precision)
error1fr = error
z = backward(flow, self.base_samples)
error = tf.reduce_max( abs(forward(flow, z)-self.base_samples) ).numpy()
num_errors = len( np.nonzero( tf.math.abs(forward(flow, z)-self.base_samples).numpy() > 0.0001 )[0] )
logger.debug(" worst case error for reverse-forward %s: %s (%s/%s errors)" % \
(msg, error, num_errors, len(self.base_samples)))
self.assertAlmostEqual(error, 0., precision)
error1rf = error
z = forward(flow, self.base_samples)
error = tf.reduce_max( tf.math.abs(backward(flow, z)-self.base_samples) ).numpy()
num_errors = len( np.nonzero( tf.math.abs(backward(flow, z)-self.base_samples).numpy() > 0.0001 )[0] )
logger.debug(" worst case error for forward-reverse %s: %s (%s/%s errors)" % \
(msg, error, num_errors, len(self.base_samples)) )
self.assertAlmostEqual(error, 0., precision)
self.assertEqual(error, error1fr)
z = backward(flow, self.base_samples)
error = tf.reduce_max( abs(forward(flow, z)-self.base_samples) ).numpy()
num_errors = len( np.nonzero( tf.math.abs(forward(flow, z)-self.base_samples).numpy() > 0.0001 )[0] )
logger.debug(" worst case error for reverse-forward %s: %s (%s/%s errors)" % \
(msg, error, num_errors, len(self.base_samples)))
self.assertAlmostEqual(error, 0., precision)
self.assertEqual(error, error1rf)
z = forward(flow, self.empty_samples)
error = tf.reduce_max( tf.math.abs(backward(flow, z)-self.empty_samples) ).numpy()
num_errors = len( np.nonzero( tf.math.abs(backward(flow, z)-self.empty_samples).numpy() > 0.0001 )[0] )
logger.debug(" worst case error for empty forward-reverse %s: %s (%s/%s errors)" % \
(msg, error, num_errors, len(self.empty_samples)))
self.assertAlmostEqual(error, 0., precision)
z = backward(flow, self.empty_samples)
error = tf.reduce_max( abs(forward(flow, z)-self.empty_samples) ).numpy()
num_errors = len( np.nonzero( tf.math.abs(forward(flow, z)-self.empty_samples).numpy() > 0.0001 )[0] )
logger.debug(" worst case error for empty reverse-forward %s: %s (%s/%s errors)" % \
(msg, error, num_errors, len(self.empty_samples)))
self.assertAlmostEqual(error, 0., precision)
if __name__ == '__main__':
logging.basicConfig(level="DEBUG")
unittest.main()