Last active
March 16, 2022 07:01
-
-
Save ariG23498/cba7faf111ed9bdcf971c160d313a5af to your computer and use it in GitHub Desktop.
Just a transformer block impelemented in TensorFlow Keras
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
class TransformerBlock(layers.Layer): | |
"""A generic Transformer block with MHSA and MLP layers. | |
Args: | |
config: The configuration of the architecture. | |
""" | |
def __init__(self, config, **kwargs): | |
super().__init__(**kwargs) | |
self.config = config | |
self.layer_norm1 = layers.LayerNormalization( | |
epsilon=self.config.eps | |
) | |
self.mhsa = layers.MultiHeadAttention( | |
num_heads=self.config.num_heads, | |
key_dim=self.config.projection_dim, | |
dropout=self.config.dropout_rate, | |
) | |
self.dropout = layers.Dropout(self.config.dropout_rate) | |
self.residual_connection = layers.Add() | |
self.layer_norm2 = layers.LayerNormalization( | |
epsilon=self.config.eps | |
) | |
self.mlp = keras.Sequential([ | |
layers.Dense( | |
units=4 * self.config.projection_dim, | |
activation=tf.nn.gelu, | |
), | |
layers.Dropout(self.config.dropout_rate), | |
layers.Dense( | |
units=self.config.projection_dim, | |
), | |
layers.Dropout(self.config.dropout_rate), | |
]) | |
def get_config(self): | |
config = super().get_config() | |
config.update(self.config) | |
return config | |
def call(self, inputs): | |
x1 = self.layer_norm1(inputs) | |
attention_outputs = self.mhsa( | |
query=x1, | |
key=x1, | |
value=x1, | |
) | |
attention_outputs = self.dropout(attention_outputs) | |
x2 = self.residual_connection([attention_outputs, inputs]) | |
x3 = self.layer_norm2(x2) | |
x4 = self.mlp(x3) | |
outputs = self.residual_connection([x4, x2]) | |
return outputs |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment