|
1
|
|
|
import os |
|
2
|
|
|
import pytest |
|
3
|
|
|
|
|
4
|
|
|
from artificial_artwork.pretrained_model.model_loader import get_vgg_19_model_path, load_default_model_parameters |
|
5
|
|
|
|
|
6
|
|
|
my_dir = os.path.dirname(os.path.realpath(__file__)) |
|
7
|
|
|
|
|
8
|
|
|
# IMAGE_MODEL_FILE_NAME = 'imagenet-vgg-verydeep-19.mat' |
|
9
|
|
|
|
|
10
|
|
|
|
|
11
|
|
|
PRODUCTION_IMAGE_MODEL = os.environ.get('AA_VGG_19', 'PRETRAINED_MODEL_NOT_FOUND') |
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
@pytest.fixture |
|
15
|
|
|
def model_parameters(): |
|
16
|
|
|
from artificial_artwork.pretrained_model.model_loader import load_default_model_parameters |
|
17
|
|
|
return load_default_model_parameters() |
|
18
|
|
|
# return load_default_model_parameters() |
|
19
|
|
|
# from artificial_artwork.pretrained_model.model_loader import load_vgg_model_parameters |
|
20
|
|
|
# return load_vgg_model_parameters |
|
21
|
|
|
|
|
22
|
|
|
|
|
23
|
|
|
@pytest.fixture |
|
24
|
|
|
def vgg_layers(): |
|
25
|
|
|
"""Expected layers structure of the vgg image model.""" |
|
26
|
|
|
from artificial_artwork.pretrained_model.vgg_architecture import LAYERS |
|
27
|
|
|
return LAYERS |
|
28
|
|
|
|
|
29
|
|
|
|
|
30
|
|
|
@pytest.fixture |
|
31
|
|
|
def style_network_architecture(): |
|
32
|
|
|
from artificial_artwork.pretrained_model.image_model import LAYERS |
|
33
|
|
|
return LAYERS |
|
34
|
|
|
|
|
35
|
|
|
|
|
36
|
|
|
@pytest.fixture |
|
37
|
|
|
def graph_factory(): |
|
38
|
|
|
from artificial_artwork.pretrained_model import graph_factory |
|
39
|
|
|
return graph_factory |
|
40
|
|
|
|
|
41
|
|
|
|
|
42
|
|
|
@pytest.mark.xfail(not os.path.isfile(PRODUCTION_IMAGE_MODEL), |
|
43
|
|
|
reason="No file found to load the pretrained image (cv) model.") |
|
44
|
|
|
def test_pretrained_model(model_parameters, graph_factory, vgg_layers, style_network_architecture): |
|
45
|
|
|
layers = model_parameters['layers'] |
|
46
|
|
|
|
|
47
|
|
|
image_specs = type('ImageSpecs', (), { |
|
48
|
|
|
'width': 400, |
|
49
|
|
|
'height': 300, |
|
50
|
|
|
'color_channels': 3 |
|
51
|
|
|
})() |
|
52
|
|
|
|
|
53
|
|
|
# verify original/loaded neural network has 43 layers |
|
54
|
|
|
assert len(layers[0]) == 43 |
|
55
|
|
|
|
|
56
|
|
|
for i, name in enumerate(vgg_layers): |
|
57
|
|
|
assert layers[0][i][0][0][0][0] == name |
|
58
|
|
|
|
|
59
|
|
|
graph = graph_factory.create(image_specs, model_parameters=model_parameters) |
|
60
|
|
|
assert set(graph.keys()) == set(['input'] + list(style_network_architecture)) |
|
61
|
|
|
|
|
62
|
|
|
|
|
63
|
|
|
@pytest.fixture |
|
64
|
|
|
def graph_builder(): |
|
65
|
|
|
from artificial_artwork.pretrained_model.model_loader import GraphBuilder |
|
66
|
|
|
return GraphBuilder() |
|
67
|
|
|
|
|
68
|
|
|
|
|
69
|
|
|
def test_building_layers(graph_builder): |
|
70
|
|
|
import numpy as np |
|
71
|
|
|
height = 2 |
|
72
|
|
|
width = 6 |
|
73
|
|
|
channels = 2 |
|
74
|
|
|
expected_input_shape = (1, height, width, channels) |
|
75
|
|
|
|
|
76
|
|
|
graph_builder.input(width, height, nb_channels=channels) |
|
77
|
|
|
# assert previous layer is the 'input' layer we just added/created |
|
78
|
|
|
assert tuple(graph_builder._prev_layer.shape) == expected_input_shape |
|
79
|
|
|
for w in range(width): |
|
80
|
|
|
for h in range(height): |
|
81
|
|
|
for c in range(channels): |
|
82
|
|
|
assert graph_builder._prev_layer[0][h][w][c] == graph_builder.graph['input'][0][h][w][c] == 0 |
|
83
|
|
|
|
|
84
|
|
|
# create relu(convolution) layer |
|
85
|
|
|
W = np.array(np.random.rand(*expected_input_shape[1:], channels), dtype=np.float32) |
|
86
|
|
|
|
|
87
|
|
|
b_weight = 6.0 |
|
88
|
|
|
b = np.array([b_weight], dtype=np.float32) |
|
89
|
|
|
graph_builder.relu_conv_2d('convo1', (W, b)) |
|
90
|
|
|
|
|
91
|
|
|
# assert the previous layer is the relu(convolution) layer we just added |
|
92
|
|
|
assert tuple(graph_builder._prev_layer.shape) == expected_input_shape |
|
93
|
|
|
for w in range(width): |
|
94
|
|
|
for h in range(height): |
|
95
|
|
|
for c in range(channels): |
|
96
|
|
|
assert graph_builder._prev_layer[0][h][w][c] == graph_builder.graph['convo1'][0][h][w][c] |
|
97
|
|
|
assert graph_builder._prev_layer[0][h][w][c] == b_weight # W[h][w][c][c] + b[0] |
|
98
|
|
|
|
|
99
|
|
|
|
|
100
|
|
|
# create Average Pooling layer |
|
101
|
|
|
layer_id = 'avgpool1' |
|
102
|
|
|
graph_builder.avg_pool(layer_id) |
|
103
|
|
|
|
|
104
|
|
|
# assert previous layer is the layer we just added/created |
|
105
|
|
|
expected_avg_pool_shape = (1, 1, 2, channels) |
|
106
|
|
|
expected_avg_output = np.array( |
|
107
|
|
|
[[[[b_weight, b_weight, b_weight], |
|
108
|
|
|
[b_weight, b_weight, b_weight], |
|
109
|
|
|
[b_weight, b_weight, b_weight] |
|
110
|
|
|
]]] |
|
111
|
|
|
,dtype=np.float32) |
|
112
|
|
|
|
|
113
|
|
|
for i in range(expected_avg_pool_shape[2]): |
|
114
|
|
|
for c in range(channels): |
|
115
|
|
|
assert graph_builder._prev_layer[0][0][i][c] == graph_builder.graph[layer_id][0][0][i][c] |
|
116
|
|
|
assert graph_builder._prev_layer[0][0][i][c] == expected_avg_output[0][0][i][c] |
|
117
|
|
|
|
boromir674 /
neural-style-transfer
