Skip to content

Commit 36a6bcf

Browse files
ducoffeMducoffeM
committed
new test file (test_backward_utils_conv) was missing
1 parent 24a49b9 commit 36a6bcf

File tree

1 file changed

+179
-0
lines changed

1 file changed

+179
-0
lines changed

‎tests/test_backward_utils_conv.py

+179
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,179 @@
1+
# Test unit for decomon with Dense layers
2+
3+
4+
import numpy as np
5+
import pytest
6+
import tensorflow.python.keras.backend as K
7+
from tensorflow.keras.layers import Input, Conv2D
8+
from tensorflow.keras.models import Model
9+
10+
from decomon.backward_layers.backward_layers import get_backward
11+
from decomon.layers.decomon_layers import DecomonDense, to_monotonic
12+
13+
from decomon.backward_layers.utils_conv import get_toeplitz
14+
from numpy.testing import assert_almost_equal
15+
16+
17+
def test_toeplitz_from_Keras(channels, filter_size, strides, flatten, data_format,
18+
padding, floatx, helpers):
19+
20+
# filter_size, strides, flatten,
21+
K.set_floatx("float{}".format(floatx))
22+
eps = K.epsilon()
23+
decimal = 5
24+
if floatx == 16:
25+
K.set_epsilon(1e-2)
26+
decimal = 0
27+
28+
if data_format == "channels_first" and not len(K._get_available_gpus()):
29+
return
30+
31+
#filter_size=3
32+
#strides=1
33+
#flatten = True
34+
odd, m_0, m_1 = 0, 0, 1
35+
36+
# should be working either with convolution of conv2D
37+
layer = Conv2D(channels, (filter_size, filter_size), strides=strides, use_bias=False,
38+
padding=padding, dtype=K.floatx())
39+
40+
inputs = helpers.get_tensor_decomposition_images_box(data_format, odd)
41+
inputs_ = helpers.get_standard_values_images_box(data_format, odd, m0=m_0, m1=m_1)
42+
y = inputs[1]
43+
result_ref = layer(y)
44+
W = get_toeplitz(layer, flatten)
45+
46+
if not flatten:
47+
w_in, h_in, c_in, w_out, h_out, c_out = W.shape
48+
W = K.reshape(W, (w_in*h_in*c_in, w_out*h_out*c_out))
49+
50+
n_in, n_out = W.shape
51+
y_flat = K.reshape(y, (-1, n_in, 1))
52+
result_flat = K.reshape(result_ref, (-1, n_out))
53+
result_toeplitz = K.sum(W[None]*y_flat, 1)
54+
output_test = K.sum((result_toeplitz-result_flat)**2)
55+
f_test = K.function(y, output_test)
56+
57+
output_test_ = f_test(inputs_[1])
58+
assert_almost_equal(
59+
output_test_,
60+
np.zeros_like(output_test_),
61+
decimal=decimal,
62+
err_msg="wrong toeplitz matrix",
63+
)
64+
65+
K.set_floatx("float{}".format(32))
66+
K.set_epsilon(eps)
67+
68+
69+
def test_toeplitz_from_Decomon(channels, filter_size, strides, flatten, data_format,
70+
padding, floatx, helpers):
71+
72+
# filter_size, strides, flatten,
73+
K.set_floatx("float{}".format(floatx))
74+
eps = K.epsilon()
75+
decimal = 5
76+
if floatx == 16:
77+
K.set_epsilon(1e-2)
78+
decimal = 0
79+
80+
if data_format == "channels_first" and not len(K._get_available_gpus()):
81+
return
82+
83+
#filter_size=3
84+
#strides=1
85+
#flatten = True
86+
odd, m_0, m_1 = 0, 0, 1
87+
88+
# should be working either with convolution of conv2D
89+
layer = Conv2D(channels, (filter_size, filter_size), strides=strides, use_bias=False,
90+
padding=padding, dtype=K.floatx())
91+
92+
inputs = helpers.get_tensor_decomposition_images_box(data_format, odd)
93+
inputs_ = helpers.get_standard_values_images_box(data_format, odd, m0=m_0, m1=m_1)
94+
y = inputs[1]
95+
result_ref = layer(y)
96+
# toeplitz matrix should be compatible with a DecomonLayer
97+
input_dim = inputs[0].shape[-1]
98+
decomon_layer = to_monotonic(layer, input_dim)[0]
99+
100+
W = get_toeplitz(decomon_layer, flatten)
101+
102+
if not flatten:
103+
w_in, h_in, c_in, w_out, h_out, c_out = W.shape
104+
W = K.reshape(W, (w_in*h_in*c_in, w_out*h_out*c_out))
105+
106+
n_in, n_out = W.shape
107+
y_flat = K.reshape(y, (-1, n_in, 1))
108+
result_flat = K.reshape(result_ref, (-1, n_out))
109+
result_toeplitz = K.sum(W[None]*y_flat, 1)
110+
output_test = K.sum((result_toeplitz-result_flat)**2)
111+
f_test = K.function(y, output_test)
112+
113+
output_test_ = f_test(inputs_[1])
114+
assert_almost_equal(
115+
output_test_,
116+
np.zeros_like(output_test_),
117+
decimal=decimal,
118+
err_msg="wrong toeplitz matrix",
119+
)
120+
121+
K.set_floatx("float{}".format(32))
122+
K.set_epsilon(eps)
123+
124+
"""
125+
def test_toeplitz_from_Decomon(channels, filter_size, strides, flatten, data_format,
126+
padding, floatx, helpers):
127+
128+
# filter_size, strides, flatten,
129+
K.set_floatx("float{}".format(floatx))
130+
eps = K.epsilon()
131+
decimal = 5
132+
if floatx == 16:
133+
K.set_epsilon(1e-2)
134+
decimal = 0
135+
136+
if data_format == "channels_first" and not len(K._get_available_gpus()):
137+
return
138+
139+
#filter_size=3
140+
#strides=1
141+
#flatten = True
142+
odd, m_0, m_1 = 0, 0, 1
143+
144+
# should be working either with convolution of conv2D
145+
layer = Conv2D(channels, (filter_size, filter_size), strides=strides, use_bias=False,
146+
padding=padding, dtype=K.floatx())
147+
148+
inputs = helpers.get_tensor_decomposition_images_box(data_format, odd)
149+
inputs_ = helpers.get_standard_values_images_box(data_format, odd, m0=m_0, m1=m_1)
150+
151+
result_ref = layer(inputs[1])
152+
# toeplitz matrix should be compatible with a DecomonLayer
153+
input_dim = inputs[0].shape[-1]
154+
decomon_layer = to_monotonic(layer, input_dim)
155+
156+
W = get_toeplitz(layer, flatten)
157+
158+
if not flatten:
159+
w_in, h_in, c_in, w_out, h_out, c_out = W.shape
160+
W = K.reshape(W, (w_in*h_in*c_in, w_out*h_out*c_out))
161+
162+
n_in, n_out = W.shape
163+
y_flat = K.reshape(inputs[1], (-1, n_in, 1))
164+
result_flat = K.reshape(result_ref, (-1, n_out))
165+
result_toeplitz = K.sum(W[None]*y_flat, 1)
166+
output_test = K.sum((result_toeplitz-result_flat)**2)
167+
f_test = K.function(y, output_test)
168+
169+
output_test_ = f_test(inputs_[1])
170+
assert_almost_equal(
171+
output_test_,
172+
np.zeros_like(output_test_),
173+
decimal=decimal,
174+
err_msg="wrong toeplitz matrix",
175+
)
176+
177+
K.set_floatx("float{}".format(32))
178+
K.set_epsilon(eps)
179+
"""

0 commit comments

Comments
 (0)