|
1
|
|
|
### Part of this code is due to the MatConvNet team and is used to load the parameters of the pretrained VGG19 model in the notebook ### |
|
2
|
|
|
|
|
3
|
|
|
import os |
|
4
|
|
|
import re |
|
5
|
|
|
from typing import Dict, Tuple, Any, Protocol |
|
6
|
|
|
|
|
7
|
|
|
import attr |
|
8
|
|
|
from numpy.typing import NDArray |
|
9
|
|
|
import numpy as np |
|
10
|
|
|
import scipy.io |
|
11
|
|
|
import tensorflow as tf |
|
12
|
|
|
|
|
13
|
|
|
from .layers_getter import VggLayersGetter |
|
14
|
|
|
from .image_model import LAYERS as NETWORK_DESIGN |
|
15
|
|
|
|
|
16
|
|
|
|
|
17
|
|
|
class ImageSpecs(Protocol): |
|
18
|
|
|
width: int |
|
19
|
|
|
height: int |
|
20
|
|
|
color_channels: int |
|
21
|
|
|
|
|
22
|
|
|
|
|
23
|
|
|
def load_vgg_model_parameters(path: str) -> Dict[str, NDArray]: |
|
24
|
|
|
return scipy.io.loadmat(path) |
|
25
|
|
|
|
|
26
|
|
|
|
|
27
|
|
|
class NoImageModelSpesifiedError(Exception): pass |
|
28
|
|
|
|
|
29
|
|
|
|
|
30
|
|
|
def get_vgg_19_model_path(): |
|
31
|
|
|
try: |
|
32
|
|
|
return os.environ['AA_VGG_19'] |
|
33
|
|
|
except KeyError as variable_not_found: |
|
34
|
|
|
raise NoImageModelSpesifiedError('No pretrained image model found. ' |
|
35
|
|
|
'Please download it and set the AA_VGG_19 environment variable with the' |
|
36
|
|
|
'path where ou stored the model (*.mat file), to indicate to wher to ' |
|
37
|
|
|
'locate and load it') from variable_not_found |
|
38
|
|
|
|
|
39
|
|
|
|
|
40
|
|
|
def load_default_model_parameters(): |
|
41
|
|
|
path = get_vgg_19_model_path() |
|
42
|
|
|
return load_vgg_model_parameters(path) |
|
43
|
|
|
|
|
44
|
|
|
|
|
45
|
|
|
def get_layers(model_parameters: Dict[str, NDArray]) -> NDArray: |
|
46
|
|
|
return model_parameters['layers'][0] |
|
47
|
|
|
|
|
48
|
|
|
|
|
49
|
|
|
class GraphBuilder: |
|
50
|
|
|
|
|
51
|
|
|
def __init__(self): |
|
52
|
|
|
self.graph = {} |
|
53
|
|
|
self._prev_layer = None |
|
54
|
|
|
|
|
55
|
|
|
def _build_layer(self, layer_id: str, layer): |
|
56
|
|
|
self.graph[layer_id] = layer |
|
57
|
|
|
self._prev_layer = layer |
|
58
|
|
|
return self |
|
59
|
|
|
|
|
60
|
|
|
def input(self, width: int, height: int, nb_channels=3, dtype='float32', layer_id='input'): |
|
61
|
|
|
self.graph = {} |
|
62
|
|
|
return self._build_layer(layer_id, tf.Variable(np.zeros((1, height, width, nb_channels)), dtype=dtype)) |
|
63
|
|
|
|
|
64
|
|
|
def avg_pool(self, layer_id: str): |
|
65
|
|
|
return self._build_layer(layer_id, tf.nn.avg_pool(self._prev_layer, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')) |
|
66
|
|
|
|
|
67
|
|
|
def relu_conv_2d(self, layer_id: str, layer_weights): |
|
68
|
|
|
"""A Relu wrapped around a convolutional layer. |
|
69
|
|
|
|
|
70
|
|
|
Will use the layer_id to find weight (for W and b matrices) values in |
|
71
|
|
|
the pretrained model (layer). |
|
72
|
|
|
|
|
73
|
|
|
Also uses the layer_id to as dict key to the output graph. |
|
74
|
|
|
""" |
|
75
|
|
|
W, b = layer_weights |
|
76
|
|
|
return self._build_layer(layer_id, tf.nn.relu(self._conv_2d(W, b))) |
|
77
|
|
|
|
|
78
|
|
|
def _conv_2d(self, W: NDArray, b: NDArray): |
|
79
|
|
|
W = tf.constant(W) |
|
80
|
|
|
b = tf.constant(np.reshape(b, (b.size))) |
|
81
|
|
|
return tf.compat.v1.nn.conv2d(self._prev_layer, filter=W, strides=[1, 1, 1, 1], padding='SAME') + b |
|
82
|
|
|
|
|
83
|
|
|
|
|
84
|
|
|
@attr.s |
|
85
|
|
|
class ModelParameters: |
|
86
|
|
|
params = attr.ib(default=attr.Factory(load_default_model_parameters)) |
|
87
|
|
|
|
|
88
|
|
|
|
|
89
|
|
|
class GraphFactory: |
|
90
|
|
|
builder = GraphBuilder() |
|
91
|
|
|
|
|
92
|
|
|
@classmethod |
|
93
|
|
|
def weights(cls, layer: NDArray) -> Tuple[NDArray, NDArray]: |
|
94
|
|
|
"""Get the weights and bias for a given layer of the VGG model.""" |
|
95
|
|
|
# wb = vgg_layers[0][layer][0][0][2] |
|
96
|
|
|
wb = layer[0][0][2] |
|
97
|
|
|
W = wb[0][0] |
|
98
|
|
|
b = wb[0][1] |
|
99
|
|
|
return W, b |
|
100
|
|
|
|
|
101
|
|
|
@classmethod |
|
102
|
|
|
def create(cls, config: ImageSpecs, model_parameters=None) -> Dict[str, Any]: |
|
103
|
|
|
"""Create a model for the purpose of 'painting'/generating a picture. |
|
104
|
|
|
|
|
105
|
|
|
Creates a Deep Learning Neural Network with most layers having weights |
|
106
|
|
|
(aka model parameters) with values extracted from a pre-trained model |
|
107
|
|
|
(ie another neural network trained on an image dataset suitably). |
|
108
|
|
|
|
|
109
|
|
|
Args: |
|
110
|
|
|
config ([type]): [description] |
|
111
|
|
|
model_parameters ([type], optional): [description]. Defaults to None. |
|
112
|
|
|
|
|
113
|
|
|
Returns: |
|
114
|
|
|
Dict[str, Any]: [description] |
|
115
|
|
|
""" |
|
116
|
|
|
|
|
117
|
|
|
vgg_model_parameters = ModelParameters(*list(filter(None, [model_parameters]))) |
|
118
|
|
|
|
|
119
|
|
|
vgg_layers = get_layers(vgg_model_parameters.params) |
|
120
|
|
|
|
|
121
|
|
|
layer_getter = VggLayersGetter(vgg_layers) |
|
122
|
|
|
|
|
123
|
|
|
def relu(layer_id: str): |
|
124
|
|
|
return cls.builder.relu_conv_2d(layer_id, cls.weights(layer_getter.id_2_layer[layer_id])) |
|
125
|
|
|
|
|
126
|
|
|
layer_callbacks = { |
|
127
|
|
|
'conv': relu, |
|
128
|
|
|
'avgpool': cls.builder.avg_pool |
|
129
|
|
|
} |
|
130
|
|
|
|
|
131
|
|
|
def layer(layer_id: str): |
|
132
|
|
|
matched_string = re.match(r'(\w+?)[\d_]*$', layer_id).group(1) |
|
133
|
|
|
return layer_callbacks[matched_string](layer_id) |
|
134
|
|
|
|
|
135
|
|
|
## Build Graph |
|
136
|
|
|
|
|
137
|
|
|
# each relu_conv_2d uses pretrained model's layer weights for W and b matrices |
|
138
|
|
|
# each average pooling layer uses custom weight values |
|
139
|
|
|
# all weights are guaranteed to remain constant (see GraphBuilder._conv_2d method) |
|
140
|
|
|
|
|
141
|
|
|
# cls.builder.input(config.image_width, config.image_height, nb_channels=config.color_channels) |
|
142
|
|
|
cls.builder.input(config.width, config.height, nb_channels=config.color_channels) |
|
143
|
|
|
for layer_id in NETWORK_DESIGN: |
|
144
|
|
|
layer(layer_id) |
|
145
|
|
|
|
|
146
|
|
|
return cls.builder.graph |
|
147
|
|
|
|
boromir674 /
neural-style-transfer
