def create_model(self):
in_tensor = x
for i in range(assignments['num_conv']):
in_tensor = create_inception_block(in_tensor=in_tensor)
in_tensor = tf.layers.batch_normalization(in_tensor,
training=training_mode)
return create_conv(
in_tensor=in_tensor,
kernel=assignments.get('output_kernel', 5),
out_features=1,
activation=assignments.get('output_act', 'sigmoid'),
)
def create_model(self):
relu = activation_functions['relu']
sigmoid = activation_functions['sigmoid']
in_tensor = x
for i in range(assignments.get('num_conv', 1)):
block_features = assignments.get('block_features', 1)
shortcut = bn(conv2d(in_tensor, 1, block_features))
in_tensor = relu(bn(conv2d(
in_tensor=in_tensor,
kernel=assignments.get('conv1_kernel', 1),
output=assignments.get('conv1_features', 1),
)))
in_tensor = bn(conv2d(
in_tensor=in_tensor,
kernel=assignments.get('conv2_kernel', 1),
output=block_features,
))
in_tensor = in_tensor + shortcut
output_1 = relu(bn(conv2d(
in_tensor=in_tensor,
kernel=assignments.get('output1_kernel', 1),
output=assignments.get('output1_features', 1),
)))
activation = sigmoid(bn(conv2d(
in_tensor=output_1,
kernel=assignments.get('activation_kernel', 1),
output=1,
)))
return activation
def create_model(self):
conv1_kernel = assignments.get('conv1_kernel', 5)
conv1_output = assignments.get('conv1_output', 64)
w_c1, b_c1 = create_weights(conv1_kernel, 4, conv1_output)
conv1 = tf.nn.conv2d(x, w_c1, strides=[1, 1, 1, 1], padding='SAME')
conv1 = tf.add(conv1, b_c1)
conv1 = activation_functions[assignments.get('conv1_act',
'relu')](conv1)
conv2_kernel = assignments.get('conv2_kernel', 5)
w_c2, b_c2 = create_weights(conv2_kernel, conv1_output, 1)
conv2 = tf.nn.conv2d(conv1, w_c2, strides=[1, 1, 1, 1], padding='SAME')
conv2 = tf.add(conv2, b_c2)
conv2 = activation_functions[assignments.get('conv2_act',
'sigmoid')](conv2)
return conv2
def create_inception_block(in_tensor):
in_features = in_tensor.get_shape()[3].value
# 'conv_a' is a single 1x1 convolution
conv_a1 = create_conv(in_tensor, 1,
assignments.get('conv_a1_features', 16))
conv_a = conv_a1
# 'conv_b' is a 1x1 convolution followed by 3x3 convolution
conv_b1 = create_conv(in_tensor, 1,
assignments.get('conv_b1_features', 16))
conv_b2 = create_conv(conv_b1, 3, assignments.get('conv_b2_features', 16))
conv_b = conv_b2
# 'conv_c' is a 1x1 convolution followed by 5x5 convolution
conv_c1 = create_conv(in_tensor, 1,
assignments.get('conv_c1_features', 16))
conv_c2 = create_conv(conv_c1, 5, assignments.get('conv_c2_features', 16))
conv_c = conv_c2
# 'conv_d' is a 3x3 max pool followed by 1x1 convolution
conv_d1 = tf.nn.max_pool(
in_tensor,
ksize=[1, 3, 3, 1],
strides=[1, 1, 1, 1],
padding='SAME',
)
conv_d2 = create_conv(conv_c1, 1, assignments.get('conv_d2_features', 16))
conv_d = conv_d2
return activation_functions[assignments.get('block_act', 'relu')](
tf.concat([conv_a, conv_b, conv_c, conv_d], 3))
def create_model(self):
in_tensor = x
intermediate_layers = [x]
for i in range(assignments.get('num_conv', 2)):
in_tensor = create_inception_block(in_tensor=in_tensor)
in_tensor = tf.layers.batch_normalization(in_tensor,
training=training_mode)
intermediate_layers.append(in_tensor)
block_result = tf.concat(intermediate_layers, 3)
final_conv_1 = create_conv(block_result,
assignments.get('out_conv1_kernel', 11),
assignments.get('out_conv1_features', 100))
final_conv_2 = create_conv(final_conv_1,
assignments.get('out_conv2_kernel', 1),
assignments.get('out_conv2_features', 100))
out_final = create_conv(
in_tensor=final_conv_2,
kernel=assignments.get('activation_kernel', 1),
out_features=1,
activation='sigmoid',
)
return out_final