def inception_mod(inputs, hyperparams: dict, name: str):
"""
Well-tuned inception module used in the subsequent GoogLeNet-like CNN
:param inputs: tensor
:param hyperparams: dictionnary of hyperparameters for the sizes of the kernels
:param name: for variable scope
"""
with tf.variable_scope(name):
# 1x1 pathway
x1 = Net.conv2d(layer_name='1x1_conv',
inputs=inputs,
kernel_shape=hyperparams["1x1_conv_kernel"],
strides=1,
activation_func=tf.nn.tanh,
padding='SAME')
# 1x1 to 3x3 pathway
x2 = Net.conv2d(layer_name='3x3_conv1',
inputs=inputs,
kernel_shape=hyperparams["3x3_conv_kernel2"],
strides=1,
activation_func=tf.nn.tanh,
padding='SAME',
kernel_shape_pre=hyperparams["3x3_conv_kernel1"])
# 1x1 to 5x5 pathway
x3 = Net.conv2d(layer_name='5x5_conv1',
inputs=inputs,
kernel_shape=hyperparams["5x5_conv_kernel2"],
strides=1,
activation_func=tf.nn.tanh,
padding='SAME',
kernel_shape_pre=hyperparams["5x5_conv_kernel1"])
# 3x3 to 1x1 pathway
x4 = tf.nn.max_pool(inputs,
ksize=[1, 3, 3, 1],
strides=[1, 1, 1, 1],
padding='SAME',
name="pooling1")
x4 = Net.conv2d(layer_name='pooling1_conv',
inputs=x4,
kernel_shape=hyperparams["pooling1_conv_kernel"],
strides=1,
activation_func=tf.nn.tanh,
padding='SAME')
x = tf.concat([x1, x2, x3, x4],
axis=3) # Concat in the 4th dim to stack
outputs = tf.tanh(x)
return outputs
def projection_shortcut(inputs):
return Net.conv2d(inputs=inputs,
kernel_shape=(1, 1, inputs.shape[-1],
filters_out),
strides=strides,
padding='SAME',
activation_func=tf.identity,
layer_name='projection')
def _building_block(inputs, filters, training, projection_shortcut,
strides, name: str):
"""
:param inputs: A tensor of size[batch, height_in, width_in, channels]
:param filters: The number of filters for the first convolution of the layer.
:param training: Either True or False, whether we are currently training the
model. Needed for batch norm.
:param projection_shortcut: The function to use for projection shortcuts
(typically a 1x1 convolution when downsampling the input).
strides: The block's stride. If greater than 1, this block will ultimately
downsample the input.
:param strides: The block's stride. If greater than 1, this block will ultimately
downsample the input.
:param name: A string name for the tensor output of the block layer.
:return: The output tensor of the block; shape should match inputs.
"""
with tf.variable_scope(name):
shortcut = inputs
inputs = Net.batch_normalization(inputs, training)
inputs = tf.nn.relu(inputs)
# The projection shortcut should come after the first batch norm and ReLU
# since it performs a 1x1 convolution.
if projection_shortcut is not None:
shortcut = projection_shortcut(inputs)
inputs = Net.conv2d(inputs=inputs,
kernel_shape=(3, 3, inputs.shape[-1], filters),
strides=strides,
padding='SAME',
activation_func=tf.identity,
layer_name='conv1_building_block')
inputs = Net.batch_normalization(inputs, training)
inputs = tf.nn.relu(inputs)
inputs = Net.conv2d(inputs=inputs,
kernel_shape=(3, 3, inputs.shape[-1], filters),
strides=1,
padding='SAME',
activation_func=tf.identity,
layer_name='conv2_building_block')
return inputs + shortcut