Implement Convolution Module in Conformer Using TensorFlow – TensorFlow Tutorial

By | March 29, 2022

In this tutorial, we will introduce how to use tensorflow to implement convolution module in conformer.

The structure of convolution module in conformer

The structure of this convolution module is:

Implement Convolution Module in Conformer Using TensorFlow - TensorFlow Tutorial

There are 9 steps in this module.

How to implement it using tensorflow?

Here is an example code for it using tensorflow.

class ConvolutionLayer():
    def __init__(self, expansion_factor = 2):
        self.expansion_factor = expansion_factor

    def __call__(self, *args, **kwargs):
        return self.forward(*args, **kwargs)
    def glu(self, inputs):
        mat1, mat2 = tf.split(inputs, 2, axis=-1)
        mat2 = tf.nn.sigmoid(mat2)

        return tf.math.multiply(mat1, mat2)
    def forward(self, inputs, filters = 512, trainable = True, dropout_p =  0.1):
        """
        Inputs: inputs
        inputs (batch, time, dim): Tensor contains input sequences
        Outputs: outputs
        outputs (batch, time, dim): Tensor produces by conformer convolution module.
        :param inputs:
        :return:
        """
        self.dropout_p = dropout_p
        # 1. LayerNormalization
        outputs = tf.contrib.layers.layer_norm(inputs=inputs, begin_norm_axis=-1, begin_params_axis=-1)
        # 2.pointwise conv
        outputs = tf.layers.conv1d(inputs, filters * self.expansion_factor, kernel_size = 1, strides=1)
        # 3.glu
        outputs = self.glu(outputs)
        # 4. 1D-Depthwise conv
        outputs = tf.layers.conv1d(inputs, filters, kernel_size=3, strides=1, padding='same')
        # 5. batch norm
        outputs = tf.layers.batch_normalization(outputs, axis=-1, training=trainable)
        # 6. swish
        outputs = tf.nn.swish(outputs)
        # 7. pointwise conv
        outputs = tf.layers.conv1d(inputs, filters, kernel_size=1, strides=1)
        # 8. dropout
        outputs = tf.layers.dropout(outputs ,rate = self.dropout_p)
        # 9. add
        return outputs + inputs

In this code, you can find each step.

Then, we can start to evalute it.

if __name__ == "__main__":
    import numpy as np

    batch_size, seq_len, dim = 3, 12, 80

    inputs = tf.random.uniform((batch_size, seq_len, dim), minval=-10, maxval=10)
    conv_layer = ConvolutionLayer()
    outputs = conv_layer(inputs, filters = 80)
    init = tf.global_variables_initializer()
    init_local = tf.local_variables_initializer()
    with tf.Session() as sess:
        sess.run([init, init_local])
        np.set_printoptions(precision=4, suppress=True)
        a = sess.run(outputs)
        print(a.shape)
        print(a)

In this code, the inputs is 3* 12 * 80, which means the batch_size = 3, time = 12 and dim = 80.

Run this code, we also will get a tensor with 3* 12* 80

(3, 12, 80)
[[[  3.8787  -6.4512  14.3416 ...  -7.5793  -5.3391   2.4872]
  [-13.6301  -5.2775   2.6721 ... -10.4354  -4.342   -3.813 ]
  [-19.9332   5.1199  22.6567 ...   4.5053  11.7114  -4.19  ]
  ...
  ...
  [ -4.1416  -8.6751  11.1841 ...  -4.1469  -6.7145  -5.4404]
  [ -7.6805   7.1907   7.1883 ...  -5.4993   8.4704  19.3019]
  [ -4.9997   3.0497 -12.2531 ...   2.2781  -5.6913  -6.9605]]]

Leave a Reply

Your email address will not be published. Required fields are marked *