Function block_reduction_b

tests/utils/custom_layers/inception_blocks.py:169–209

Builds Reduction-B block for Inception v4 network.

(inputs, scope=None, is_train=False)

def block_reduction_b(inputs, scope=None, is_train=False):
    """Builds Reduction-B block for Inception v4 network."""
    # By default use stride=1 and SAME padding

    with tf.variable_scope(scope, 'BlockReductionB', [inputs]):
        with tf.variable_scope('Branch_0'):
            branch_0, _ = conv_module(
                inputs, n_out_channel=192, filter_size=(1, 1), strides=(1, 1), padding='SAME', batch_norm_init=None,
                is_train=is_train, use_batchnorm=True, activation_fn='ReLU', name='Conv2d_0a_1x1'
            )

            branch_0, _ = conv_module(
                branch_0, n_out_channel=192, filter_size=(3, 3), strides=(2, 2), padding='VALID', batch_norm_init=None,
                is_train=is_train, use_batchnorm=True, activation_fn='ReLU', name='Conv2d_1a_3x3'
            )

        with tf.variable_scope('Branch_1'):
            branch_1, _ = conv_module(
                inputs, n_out_channel=256, filter_size=(1, 1), strides=(1, 1), padding='SAME', batch_norm_init=None,
                is_train=is_train, use_batchnorm=True, activation_fn='ReLU', name='Conv2d_0a_1x1'
            )

            branch_1, _ = conv_module(
                branch_1, n_out_channel=256, filter_size=(1, 7), strides=(1, 1), padding='SAME', batch_norm_init=None,
                is_train=is_train, use_batchnorm=True, activation_fn='ReLU', name='Conv2d_0b_1x7'
            )

            branch_1, _ = conv_module(
                branch_1, n_out_channel=320, filter_size=(7, 1), strides=(1, 1), padding='SAME', batch_norm_init=None,
                is_train=is_train, use_batchnorm=True, activation_fn='ReLU', name='Conv2d_0c_7x1'
            )

            branch_1, _ = conv_module(
                branch_1, n_out_channel=320, filter_size=(3, 3), strides=(2, 2), padding='VALID', batch_norm_init=None,
                is_train=is_train, use_batchnorm=True, activation_fn='ReLU', name='Conv2d_1a_3x3'
            )

        with tf.variable_scope('Branch_2'):
            branch_2 = tl.layers.MaxPool2d(inputs, (3, 3), strides=(2, 2), padding='VALID', name='MaxPool_1a_3x3')

        return tl.layers.ConcatLayer([branch_0, branch_1, branch_2], concat_dim=3, name='concat_layer')


def block_inception_c(inputs, scope=None, is_train=False):