def __init__(self, x):
conv1 = layers.conv2d(x,
64, [7, 7],
stride=2,
scope='conv1',
activation_fn=tf.nn.relu)
pool1 = layers.max_pool2d(conv1, [3, 3], scope='pool1')
pool1_lrn = tf.nn.local_response_normalization(pool1,
name='pool1_lrn',
alpha=0.0001,
beta=0.75)
conv2_reduce = layers.conv2d(pool1_lrn,
64, [1, 1],
scope='conv2_reduce',
activation_fn=tf.nn.relu)
conv2 = layers.conv2d(conv2_reduce,
192, [3, 3],
scope='conv2',
activation_fn=tf.nn.relu)
conv2_lrn = tf.nn.local_response_normalization(conv2,
name='pool1_lrn',
alpha=0.0001,
beta=0.75)
pool2 = layers.max_pool2d(conv2_lrn, [3, 3], scope='pool2')
with tf.variable_scope('inception_3a'):
inception_3a = inception(pool2, 64, 96, 128, 16, 32, 32)
with tf.variable_scope('inception_3b'):
inception_3b = inception(inception_3a, 128, 128, 192, 32, 96, 64)
pool3 = layers.max_pool2d(inception_3b, [3, 3], scope='pool3')
'''
with tf.variable_scope('inception_4a'):
conv11 = layers.conv2d(pool3, 192, [1, 1], activation_fn=tf.nn.relu)
conv33_reduce = layers.conv2d(pool3, 96, [1, 1], activation_fn=tf.nn.relu)
conv33_reduce_param = layers.conv2d(conv33_reduce, 32, [1, 1], activation_fn=tf.nn.relu)
conv33_reduce_param2 = layers.conv2d(conv33_reduce_param, 32, [1, 1],
activation_fn=tf.nn.relu)
conv33_reduce_param3 = layers.conv2d(conv33_reduce_param2, 6, [1, 1])
conv33_spatial_transfo = conv_spatial_transfo(conv33_reduce, conv33_reduce_param3, 3)
conv33 = layers.conv2d(conv33_spatial_transfo, 208, [3, 3], stride=3)
conv55_reduce = layers.conv2d(pool3, 16, [1, 1], activation_fn=tf.nn.relu)
conv55_reduce_param = layers.conv2d(conv55_reduce, 32, [1, 1], activation_fn=tf.nn.relu)
conv55_reduce_param2 = layers.conv2d(conv55_reduce_param, 32, [1, 1],
activation_fn=tf.nn.relu)
conv55_reduce_param3 = layers.conv2d(conv55_reduce_param2, 6, [1, 1])
conv55_spatial_transfo = conv_spatial_transfo(conv55_reduce, conv55_reduce_param3, 5)
conv55 = layers.conv2d(conv55_spatial_transfo, 48, [5, 5], stride=5)
pool_proj = layers.max_pool2d(pool3, [3, 3], stride=1, padding='SAME')
pool11 = layers.conv2d(pool_proj, 64, [1, 1])
inception4a = tf.concat([conv11, conv33, conv55, pool11], 3)
'''
feature_unnorm = layers.conv2d(pool3,
128, [1, 1],
scope='feature_unnorm')
self.feature = tf.nn.l2_normalize(feature_unnorm, dim=3)
def create_model(opt):
"""Create neural network model, defining layer architecture."""
dim_y = int((opt['crop_max_y'] - opt['crop_min_y']) / opt['scale_factor'])
dim_x = int((opt['crop_max_x'] - opt['crop_min_x']) / opt['scale_factor'])
inputs = Input(shape=(dim_y, dim_x, 3))
conv_1_1 = inception(inputs, 64)
conv_1_2 = inception(conv_1_1, 64)
pool_1 = MaxPooling2D()(conv_1_2)
conv_2_1 = inception(pool_1, 128)
conv_2_2 = inception(conv_2_1, 128)
pool_2 = MaxPooling2D()(conv_2_2)
conv_3_1 = inception(pool_2, 128)
conv_3_2 = inception(conv_3_1, 128)
conv_3_3 = inception(conv_3_2, 128)
pool_3 = MaxPooling2D()(conv_3_3)
conv_4_1 = inception(pool_3, 256)
conv_4_2 = inception(conv_4_1, 256)
conv_4_3 = inception(conv_4_2, 256)
pool_4 = MaxPooling2D()(conv_4_3)
conv_5_1 = inception(pool_4, 256)
conv_5_2 = inception(conv_5_1, 256)
conv_5_3 = inception(conv_5_2, 256)
conv_5_4 = inception(conv_5_3, 256)
conv_5_5 = inception(conv_5_4, 256)
conv_5_6 = inception(conv_5_5, 256)
unpool_4 = UpSampling2D()(conv_5_6)
unpool_4 = Concatenate()([unpool_4, conv_4_3]) # skip layer
decode_4_1 = inception(unpool_4, 256)
decode_4_2 = inception(decode_4_1, 256)
decode_4_3 = inception(decode_4_2, 128)
unpool_3 = UpSampling2D()(decode_4_3)
unpool_3 = Concatenate()([unpool_3, conv_3_3]) # skip layer
decode_3_1 = inception(unpool_3, 128)
decode_3_2 = inception(decode_3_1, 128)
decode_3_3 = inception(decode_3_2, 128)
unpool_2 = UpSampling2D()(decode_3_3)
unpool_2 = Concatenate()([unpool_2, conv_2_2]) # skip layer
decode_2_1 = inception(unpool_2, 128)
decode_2_2 = inception(decode_2_1, 64)
unpool_1 = UpSampling2D()(decode_2_2)
unpool_1 = Concatenate()([unpool_1, conv_1_2]) # skip layer
decode_1_1 = inception(unpool_1, 64)
final_layer = Conv2D(1, 3, padding='same',
kernel_regularizer=l2(0.01))(decode_1_1)
final_layer = Activation('sigmoid')(final_layer)
final_layer = Reshape((dim_y, dim_x))(final_layer)
model = Model(inputs=inputs, outputs=final_layer, name="Road")
return model