Source code for ASTROMER.core.attention

import tensorflow as tf


[docs]def scaled_dot_product_attention(q, k, v, mask): """ Calculate the attention weights. q, k, v must have matching leading dimensions. k, v must have matching penultimate dimension, i.e.: seq_len_k = seq_len_v. The mask has different shapes depending on its type(padding or look ahead) but it must be broadcastable for addition. :param q: query shape == (..., seq_len_q, depth) :param k: key shape == (..., seq_len_k, depth) :param v: value shape == (..., seq_len_v, depth_v) :param mask: Float tensor with shape broadcastable to (..., seq_len_q, seq_len_k). Defaults to None. :return: output, attention_weights """ matmul_qk = tf.matmul(q, k, transpose_b=True) # (..., seq_len_q, seq_len_k) # scale matmul_qk dk = tf.cast(tf.shape(k)[-1], tf.float32) scaled_attention_logits = matmul_qk / tf.math.sqrt(dk) steps = tf.shape(scaled_attention_logits)[2] mask_rshp = tf.tile(mask, [1,1,steps]) mask_rshp += tf.transpose(mask_rshp, [0,2,1]) mask_rshp = tf.minimum(1., mask_rshp) mask_rshp = tf.expand_dims(mask_rshp, 1) scaled_attention_logits += mask_rshp # softmax is normalized on the last axis (seq_len_k) so that the scores add up to 1. attention_weights = tf.nn.softmax(scaled_attention_logits, axis=-1, name='MaskedSoftMax') # (..., seq_len_q, seq_len_k) output = tf.matmul(attention_weights, v, name='Z') # (..., seq_len_q, depth_v) return output, attention_weights
[docs]class MultiHeadAttention(tf.keras.layers.Layer): """ """ def __init__(self, d_model, num_heads): super(MultiHeadAttention, self).__init__() self.num_heads = num_heads self.d_model = d_model assert d_model % self.num_heads == 0 self.depth = d_model // self.num_heads # final dimension self.wq = tf.keras.layers.Dense(d_model, name='WQ') self.wk = tf.keras.layers.Dense(d_model, name='WK') self.wv = tf.keras.layers.Dense(d_model, name='WV') self.dense = tf.keras.layers.Dense(d_model, name='MixerDense')
[docs] def split_heads(self, x, batch_size, name='qkv'): """Split the last dimension into (num_heads, depth). Transpose the result such that the shape is (batch_size, num_heads, seq_len, depth) """ x = tf.reshape(x, (batch_size, -1, self.num_heads, self.depth)) return tf.transpose(x, perm=[0, 2, 1, 3], name=name)
[docs] def call(self, x, mask): batch_size = tf.shape(x)[0] q = self.wq(x) # (batch_size, seq_len, d_model) k = self.wk(x) # (batch_size, seq_len, d_model) v = self.wv(x) # (batch_size, seq_len, d_model) q = self.split_heads(q, batch_size, name='Q') # (batch_size, num_heads, seq_len_q, depth) k = self.split_heads(k, batch_size, name='K') # (batch_size, num_heads, seq_len_k, depth) v = self.split_heads(v, batch_size, name='V') # (batch_size, num_heads, seq_len_v, depth) # scaled_attention.shape == (batch_size, num_heads, seq_len_q, depth) # attention_weights.shape == (batch_size, num_heads, seq_len_q, seq_len_k) scaled_attention, attention_weights = scaled_dot_product_attention(q, k, v, mask) scaled_attention = tf.transpose(scaled_attention, perm=[0, 2, 1, 3]) # (batch_size, seq_len_q, num_heads, depth) concat_attention = tf.reshape(scaled_attention, (batch_size, -1, self.d_model)) # (batch_size, seq_len_q, d_model) output = self.dense(concat_attention) # (batch_size, seq_len_q, d_model) return output, attention_weights