-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathvgg16.py
226 lines (190 loc) · 11.4 KB
/
vgg16.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
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
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
# VGG-16, 16-layer model from the paper:
# "Very Deep Convolutional Networks for Large-Scale Image Recognition"
# Original source: https://gist.github.com/ksimonyan/211839e770f7b538e2d8
# License: see http://www.robots.ox.ac.uk/~vgg/research/very_deep/
# Download pretrained weights from:
# https://s3.amazonaws.com/lasagne/recipes/pretrained/imagenet/vgg16.pkl
import numpy as np
import theano
import theano.tensor as T
import lasagne
from lasagne_models import LasagneModel, Lasagne_Conv_Deconv
import settings
import hyper_params
from utils import print_critical, print_error, print_warning, print_info, print_positive, log, logout
class VGG16_Model(LasagneModel):
def __init__(self, hyperparams = hyper_params.default_vgg16_hyper_params):
super(VGG16_Model, self).__init__(hyperparams = hyperparams)
self.input_prevgg = None
self.input_prevgg_out = None
self.target_prevgg = None
self.target_prevgg_out = None
self.input_vgg_model = None
self.input_vgg_model_out = None
self.target_vgg_model = None
self.target_vgg_model_out = None
def build(self):
pass
def build_network(self, input_var, target_var):
from lasagne.layers import InputLayer
from lasagne.layers import DenseLayer
from lasagne.layers import NonlinearityLayer
from lasagne.layers import DropoutLayer
from lasagne.layers import ReshapeLayer
from lasagne.layers import Pool2DLayer as PoolLayer
from lasagne.layers import TransposedConv2DLayer as Deconv2DLayer
from lasagne.nonlinearities import softmax, sigmoid, tanh
import cPickle as pickle
try:
from lasagne.layers.dnn import Conv2DDNNLayer as ConvLayer
except ImportError as e:
from lasagne.layers import Conv2DLayer as ConvLayer
print_warning("Cannot import 'lasagne.layers.dnn.Conv2DDNNLayer' as it requires GPU support and a functional cuDNN installation. Falling back on slower convolution function 'lasagne.layers.Conv2DLayer'.")
batch_size = settings.BATCH_SIZE
net = {}
net['input'] = InputLayer((batch_size, 3, 64, 64), input_var=input_var)
net['conv1'] = ConvLayer(net['input'], 64, 3, stride=1, pad='same') # 64x64
net['pool1'] = PoolLayer(net['conv1'], 2) # 32x32
net['conv2'] = ConvLayer(net['pool1'], 64, 3, stride=1, pad='same') # 32x32
net['dropout1'] = DropoutLayer(net['conv2'], p=0.5)
net['conv3'] = ConvLayer(net['dropout1'], 64, 3, stride=1, pad='same') # 32x32
net['dropout3'] = DropoutLayer(net['conv3'], p=0.5)
net['fc1'] = DenseLayer(net['dropout3'], 3*32*32)
net['output'] = ReshapeLayer(net['fc1'], (batch_size, 3, 32, 32))
# net['input'] = InputLayer((batch_size, 3, 64, 64), input_var=input_var)
# net['dropout1'] = DropoutLayer(net['input'], p=0.1)
# net['conv1'] = ConvLayer(net['dropout1'], 256, 5, stride=2, pad='same') # 32x32
# net['dropout2'] = DropoutLayer(net['conv1'], p=0.5)
# net['conv2'] = ConvLayer(net['dropout2'], 256, 7, stride=1, pad='same') # 32x32
# net['dropout3'] = DropoutLayer(net['conv2'], p=0.5)
# net['deconv1'] = Deconv2DLayer(net['dropout3'], 256, 7, stride=1, crop='same', output_size=32) # 32x32
# net['dropout4'] = DropoutLayer(net['deconv1'], p=0.5)
# net['deconv3'] = Deconv2DLayer(net['dropout4'], 256, 9, stride=1, crop='same', output_size=32) # 32x32
# net['dropout5'] = DropoutLayer(net['deconv3'], p=0.5)
# net['fc1'] = DenseLayer(net['dropout5'], 2048)
# net['dropout6'] = DropoutLayer(net['fc1'], p=0.5)
# net['fc2'] = DenseLayer(net['dropout6'], 2048)
# net['dropout7'] = DropoutLayer(net['fc2'], p=0.5)
# net['fc3'] = DenseLayer(net['dropout7'], 3*32*32)
# net['dropout8'] = DropoutLayer(net['fc3'], p=0.5)
# net['reshape'] = ReshapeLayer(net['dropout8'], ([0], 3, 32, 32))
# net['output'] = Deconv2DLayer(net['reshape'], 3, 9, stride=1, crop='same', output_size=32, nonlinearity=sigmoid)
self.network, self.network_out = net, net['output']
print ("Conv_Deconv network output shape: {}".format(self.network_out.output_shape))
# self.input_pad, self.input_pad_out = self.build_pad_model(self.network_out)
# self.target_pad, self.target_pad_out = self.build_pad_model(InputLayer((batch_size, 3, 32, 32), input_var=target_var))
self.input_scaled, self.input_scaled_out = self.build_scaled_model(self.network_out)
self.target_scaled, self.target_scaled_out = self.build_scaled_model(InputLayer((batch_size, 3, 32, 32), input_var=target_var))
print("(Input) scaled network output shape: {}".format(self.input_scaled_out.output_shape))
print("(Target) scaled network output shape: {}".format(self.target_scaled_out.output_shape))
self.vgg_scaled_var = T.tensor4('scaled_vars')
self.vgg_model, self.vgg_model_out = self.build_vgg_model(self.vgg_scaled_var)
print("VGG model conv1_1 output shape: {}".format(self.vgg_model['conv1_1'].output_shape))
print("VGG model conv2_1 output shape: {}".format(self.vgg_model['conv2_1'].output_shape))
print("VGG model conv3_1 output shape: {}".format(self.vgg_model['conv3_1'].output_shape))
def build_pad_model(self, previous_layer):
from lasagne.layers import PadLayer
padnet = {}
padnet['input'] = previous_layer
padnet['pad'] = PadLayer(padnet['input'], (224-32)/2)
return padnet, padnet['pad']
def build_scaled_model(self, previous_layer):
from lasagne.layers import TransformerLayer
b = np.zeros((2, 3), dtype='float32')
b[0, 0] = 7.0
b[1, 1] = 7.0
b = b.flatten() # identity transform
W = lasagne.init.Constant(0.0)
scalenet = {}
scalenet['input'] = previous_layer
scalenet['scale_init'] = lasagne.layers.DenseLayer(scalenet['input'], num_units=6, W=W, b=b, nonlinearity=None)
scalenet['scale'] = TransformerLayer(scalenet['input'], scalenet['scale_init'], downsample_factor=1.0/7.0) # Output should be 3x224x224
return scalenet, scalenet['scale']
def build_vgg_model(self, input_var):
from lasagne.layers import InputLayer
from lasagne.layers import DenseLayer
from lasagne.layers import NonlinearityLayer
from lasagne.layers import DropoutLayer
from lasagne.layers import Pool2DLayer as PoolLayer
from lasagne.layers import TransposedConv2DLayer as Deconv2DLayer
from lasagne.nonlinearities import softmax, sigmoid, tanh
import cPickle as pickle
try:
from lasagne.layers.dnn import Conv2DDNNLayer as ConvLayer
except ImportError as e:
from lasagne.layers import Conv2DLayer as ConvLayer
print_warning("Cannot import 'lasagne.layers.dnn.Conv2DDNNLayer' as it requires GPU support and a functional cuDNN installation. Falling back on slower convolution function 'lasagne.layers.Conv2DLayer'.")
print_info("Building VGG-16 model...")
net = {}
net['input'] = InputLayer(shape = (None, 3, 224, 224), input_var = input_var, name = 'vgg_input')
net['conv1_1'] = ConvLayer(
net['input'], 64, 3, pad=1, flip_filters=False)
net['conv1_2'] = ConvLayer(
net['conv1_1'], 64, 3, pad=1, flip_filters=False)
net['pool1'] = PoolLayer(net['conv1_2'], 2)
net['conv2_1'] = ConvLayer(
net['pool1'], 128, 3, pad=1, flip_filters=False)
net['conv2_2'] = ConvLayer(
net['conv2_1'], 128, 3, pad=1, flip_filters=False)
net['pool2'] = PoolLayer(net['conv2_2'], 2)
net['conv3_1'] = ConvLayer(
net['pool2'], 256, 3, pad=1, flip_filters=False)
net['conv3_2'] = ConvLayer(
net['conv3_1'], 256, 3, pad=1, flip_filters=False)
net['conv3_3'] = ConvLayer(
net['conv3_2'], 256, 3, pad=1, flip_filters=False)
net['pool3'] = PoolLayer(net['conv3_3'], 2)
net['conv4_1'] = ConvLayer(
net['pool3'], 512, 3, pad=1, flip_filters=False)
net['conv4_2'] = ConvLayer(
net['conv4_1'], 512, 3, pad=1, flip_filters=False)
net['conv4_3'] = ConvLayer(
net['conv4_2'], 512, 3, pad=1, flip_filters=False)
net['pool4'] = PoolLayer(net['conv4_3'], 2)
net['conv5_1'] = ConvLayer(
net['pool4'], 512, 3, pad=1, flip_filters=False)
net['conv5_2'] = ConvLayer(
net['conv5_1'], 512, 3, pad=1, flip_filters=False)
net['conv5_3'] = ConvLayer(
net['conv5_2'], 512, 3, pad=1, flip_filters=False)
net['pool5'] = PoolLayer(net['conv5_3'], 2)
net['fc6'] = DenseLayer(net['pool5'], num_units=4096)
net['fc6_dropout'] = DropoutLayer(net['fc6'], p=0.5)
net['fc7'] = DenseLayer(net['fc6_dropout'], num_units=4096)
net['fc7_dropout'] = DropoutLayer(net['fc7'], p=0.5)
net['fc8'] = DenseLayer(
net['fc7_dropout'], num_units=1000, nonlinearity=None)
net['prob'] = NonlinearityLayer(net['fc8'], softmax)
net_output = net['prob']
print_info("Loading VGG16 pre-trained weights from file 'vgg16.pkl'...")
with open('vgg16.pkl', 'rb') as f:
params = pickle.load(f)
#net_output.initialize_layers()
lasagne.layers.set_all_param_values(net['prob'], params['param values'])
print_info("Alright, pre-trained VGG16 model is ready!")
return net, net['prob']
def build_loss(self, input_var, target_var, deterministic=False):
from lasagne.layers import get_output
from lasagne.objectives import squared_error, aggregate
from lasagne.regularization import regularize_layer_params, l2, l1
# Compute good ol' L2-norm loss between prediction and target
network_output = get_output(self.network_out, deterministic=deterministic)
l2_loss = squared_error(network_output, target_var).mean() / 1000.0
# Penalties
#l2_penalty = regularize_layer_params(self.network, l2) / 1000.0
#l1_penalty = regularize_layer_params(self.network, l1) * 1e-8
# Compute loss from VGG's intermediate layers
x_scaled = get_output(self.input_scaled_out, deterministic=deterministic)
y_scaled = get_output(self.target_scaled_out, deterministic=deterministic)
#layers = [self.vgg_model['conv1_1'], self.vgg_model['conv2_1'], self.vgg_model['conv3_1'], self.vgg_model['conv4_2']]
#x_1, x_2, x_3, x_4 = get_output(layers, inputs=x_scaled, deterministic=deterministic)
#y_1, y_2, y_3, y_4 = get_output(layers, inputs=y_scaled, deterministic=deterministic)
#loss_conv1_1 = squared_error(x_1, y_1).mean() / 1000.0
#loss_conv2_1 = squared_error(x_2, y_3).mean() / 1000.0
#loss_conv3_1 = squared_error(x_3, y_3).mean() / 1000.0
#loss_conv4_2 = squared_error(x_4, y_4).mean() / 1000.0
x_4 = get_output(self.vgg_model['conv4_2'], x_scaled, deterministic=deterministic)
y_4 = get_output(self.vgg_model['conv4_2'], y_scaled, deterministic=deterministic)
loss_conv4_2 = squared_error(x_4, y_4).mean() / 1000.0
#return l2_loss + l2_penalty + l1_penalty + 0.001*loss_conv1_1 + 0.001*loss_conv2_1 + 0.005*loss_conv3_1 + 0.01*loss_conv4_2
return l2_loss + settings.RATIO_VGG_LOSS * loss_conv4_2