def create(self):
# 1st Layer: Conv (w ReLu) -> Pool -> Lrn
conv1 = conv(self.tensor,
11,
11,
96,
4,
4,
padding='VALID',
name='conv1',
trainable=self.is_layer_trainable('conv1'))
norm1 = tf.nn.local_response_normalization(conv1,
depth_radius=2,
alpha=2e-05,
beta=0.75,
bias=1.0,
name='norm1')
pool1 = tf.nn.max_pool(norm1,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='VALID',
name='pool1')
# 2nd Layer: Conv (w ReLu) -> Pool -> Lrn with 2 groups
conv2 = conv(pool1,
5,
5,
256,
1,
1,
groups=2,
name='conv2',
trainable=self.is_layer_trainable('conv2'))
norm2 = tf.nn.local_response_normalization(conv2,
depth_radius=2,
alpha=2e-05,
beta=0.75,
bias=1.0,
name='norm2')
pool2 = tf.nn.max_pool(norm2,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='VALID',
name='pool2')
# 3rd Layer: Conv (w ReLu)
conv3 = conv(pool2,
3,
3,
384,
1,
1,
name='conv3',
trainable=self.is_layer_trainable('conv3'))
# 4th Layer: Conv (w ReLu) splitted into two groups
conv4 = conv(conv3,
3,
3,
384,
1,
1,
groups=2,
name='conv4',
trainable=self.is_layer_trainable('conv4'))
# 5th Layer: Conv (w ReLu) -> Pool splitted into two groups
conv5 = conv(conv4,
3,
3,
256,
1,
1,
groups=2,
name='conv5',
trainable=self.is_layer_trainable('conv5'))
pool5 = tf.nn.max_pool(conv5,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='VALID',
name='pool5')
# 6th Layer: Flatten -> FC (w ReLu) -> Dropout
pool5_out = int(np.prod(pool5.get_shape()[1:])) # 6 * 6 * 256 = 9216
pool5_flat = tf.reshape(
pool5,
[-1, pool5_out
]) # shape=(image count, 6, 6, 256) -> shape=(image count, 9216)
fc6 = fc(pool5_flat,
4096,
name='fc6',
trainable=self.is_layer_trainable('fc6'))
dropout6 = tf.nn.dropout(fc6, self.keep_prob)
# 7th Layer: FC (w ReLu) -> Dropout
fc7 = fc(dropout6,
4096,
name='fc7',
trainable=self.is_layer_trainable('fc7'))
dropout7 = tf.nn.dropout(fc7, self.keep_prob)
# 8th Layer: FC and return unscaled activations (for tf.nn.softmax_cross_entropy_with_logits)
fc8 = fc(dropout7,
self.num_classes,
relu=False,
name='fc8',
trainable=self.is_layer_trainable('fc8'))
# add layers to the endpoints dict
endpoints = OrderedDict()
endpoints['conv1'] = conv1
endpoints['pool1'] = pool1
endpoints['conv2'] = conv2
endpoints['pool2'] = pool2
endpoints['conv3'] = conv3
endpoints['conv4'] = conv4
endpoints['conv5'] = conv5
endpoints['pool5'] = pool5
endpoints['pool5/flat'] = pool5_flat # 9216
endpoints['fc6'] = fc6 # 4096
endpoints['fc7'] = fc7 # 4096
endpoints['fc8'] = fc8 # number of classes
return fc8, endpoints
def create(self):
# 1st Layer: Conv -> Conv -> Pool
# conv(tensor, filter_height, filter_width, num_filters, stride_y, stride_x, name, padding)
conv1_1 = conv(self.tensor,
3,
3,
64,
1,
1,
padding='SAME',
name='conv1_1',
trainable=self.is_layer_trainable('conv1_1'))
conv1_2 = conv(conv1_1,
3,
3,
64,
1,
1,
padding='SAME',
name='conv1_2',
trainable=self.is_layer_trainable('conv1_2'))
pool1 = tf.nn.max_pool(conv1_2,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool1')
# 2nd Layer: Conv -> Conv -> Pool
conv2_1 = conv(pool1,
3,
3,
128,
1,
1,
padding='SAME',
name='conv2_1',
trainable=self.is_layer_trainable('conv2_1'))
conv2_2 = conv(conv2_1,
3,
3,
128,
1,
1,
padding='SAME',
name='conv2_2',
trainable=self.is_layer_trainable('conv2_2'))
pool2 = tf.nn.max_pool(conv2_2,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool2')
# 3rd Layer: Conv -> Conv -> Conv -> Pool
conv3_1 = conv(pool2,
3,
3,
256,
1,
1,
padding='SAME',
name='conv3_1',
trainable=self.is_layer_trainable('conv3_1'))
conv3_2 = conv(conv3_1,
3,
3,
256,
1,
1,
padding='SAME',
name='conv3_2',
trainable=self.is_layer_trainable('conv3_2'))
conv3_3 = conv(conv3_2,
3,
3,
256,
1,
1,
padding='SAME',
name='conv3_3',
trainable=self.is_layer_trainable('conv3_3'))
pool3 = tf.nn.max_pool(conv3_3,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool3')
# 4th Layer: Conv -> Conv -> Conv -> Pool
conv4_1 = conv(pool3,
3,
3,
512,
1,
1,
padding='SAME',
name='conv4_1',
trainable=self.is_layer_trainable('conv4_1'))
conv4_2 = conv(conv4_1,
3,
3,
512,
1,
1,
padding='SAME',
name='conv4_2',
trainable=self.is_layer_trainable('conv4_2'))
conv4_3 = conv(conv4_2,
3,
3,
512,
1,
1,
padding='SAME',
name='conv4_3',
trainable=self.is_layer_trainable('conv4_3'))
pool4 = tf.nn.max_pool(conv4_3,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool4')
# 5th Layer: Conv -> Conv -> Conv -> Pool
conv5_1 = conv(pool4,
3,
3,
512,
1,
1,
padding='SAME',
name='conv5_1',
trainable=self.is_layer_trainable('conv5_1'))
conv5_2 = conv(conv5_1,
3,
3,
512,
1,
1,
padding='SAME',
name='conv5_2',
trainable=self.is_layer_trainable('conv5_2'))
conv5_3 = conv(conv5_2,
3,
3,
512,
1,
1,
padding='SAME',
name='conv5_3',
trainable=self.is_layer_trainable('conv5_3'))
pool5 = tf.nn.max_pool(conv5_3,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool5')
# 6th Layer: FC -> DropOut
# [1:] cuts away the first element
pool5_out = int(np.prod(pool5.get_shape()[1:])) # 7 * 7 * 512 = 25088
pool5_flat = tf.reshape(
pool5,
[-1, pool5_out
]) # shape=(image count, 7, 7, 512) -> shape=(image count, 25088)
fc6 = fc(pool5_flat,
num_out=4096,
name='fc6',
relu=True,
trainable=self.is_layer_trainable('fc6'))
dropout1 = tf.nn.dropout(fc6, self.keep_prob)
# 7th Layer: FC
fc7 = fc(dropout1,
num_out=4096,
name='fc7',
relu=True,
trainable=self.is_layer_trainable('fc7'))
dropout2 = tf.nn.dropout(fc7, self.keep_prob)
# 8th Layer: FC
fc8 = fc(dropout2,
num_out=self.num_classes,
name='fc8',
relu=False,
trainable=self.is_layer_trainable('fc8'))
# add layers to the endpoints dict
endpoints = OrderedDict()
endpoints['conv1/conv1_1'] = conv1_1
endpoints['conv1/conv1_2'] = conv1_2
endpoints['pool1'] = pool1
endpoints['conv2/conv2_1'] = conv2_1
endpoints['conv2/conv2_2'] = conv2_2
endpoints['pool2'] = pool2
endpoints['conv3/conv3_1'] = conv3_1
endpoints['conv3/conv3_2'] = conv3_2
endpoints['conv3/conv3_3'] = conv3_3
endpoints['pool3'] = pool3
endpoints['conv4/conv4_1'] = conv4_1
endpoints['conv4/conv4_2'] = conv4_2
endpoints['conv4/conv4_3'] = conv4_3
endpoints['pool4'] = pool4
endpoints['conv5/conv5_1'] = conv5_1
endpoints['conv5/conv5_2'] = conv5_2
endpoints['conv5/conv5_3'] = conv5_3
endpoints['pool5'] = pool5
endpoints['pool5/flat'] = pool5_flat # 25088
endpoints['fc6'] = fc6 # 4096
endpoints['fc7'] = fc7 # 4096
endpoints['fc8'] = fc8 # number of output classes
return fc8, endpoints