MultiHeadAttention Layer (#1062)

cgarciae · web-flow · commit 3b0d97804c15 · 2020-03-10T18:21:07.000-04:00
* Add MultiHeadAttention Layer
diff --git a/.github/CODEOWNERS b/.github/CODEOWNERS
@@ -49,6 +49,7 @@
 
 /tensorflow_addons/layers/gelu*.py @aakashkumarnain
 /tensorflow_addons/layers/maxout*.py @failure-to-thrive
+/tensorflow_addons/layers/multihead_attention*.py @cgarciae
 /tensorflow_addons/layers/netvlad*.py @joel-shor
 /tensorflow_addons/layers/normalizations*.py @smokrow
 /tensorflow_addons/layers/optical_flow*.py @failure-to-thrive
diff --git a/tensorflow_addons/layers/BUILD b/tensorflow_addons/layers/BUILD
@@ -8,6 +8,7 @@ py_library(
         "__init__.py",
         "gelu.py",
         "maxout.py",
+        "multihead_attention.py",
         "netvlad.py",
         "normalizations.py",
         "optical_flow.py",
@@ -147,3 +148,15 @@ py_test(
         ":layers",
     ],
 )
+
+py_test(
+    name = "multihead_attention_test",
+    size = "small",
+    srcs = [
+        "multihead_attention_test.py",
+    ],
+    main = "multihead_attention_test.py",
+    deps = [
+        ":layers",
+    ],
+)
diff --git a/tensorflow_addons/layers/__init__.py b/tensorflow_addons/layers/__init__.py
@@ -16,6 +16,7 @@
 
 from tensorflow_addons.layers.gelu import GELU
 from tensorflow_addons.layers.maxout import Maxout
+from tensorflow_addons.layers.multihead_attention import MultiHeadAttention
 from tensorflow_addons.layers.normalizations import GroupNormalization
 from tensorflow_addons.layers.normalizations import InstanceNormalization
 from tensorflow_addons.layers.optical_flow import CorrelationCost
diff --git a/tensorflow_addons/layers/multihead_attention.py b/tensorflow_addons/layers/multihead_attention.py
@@ -0,0 +1,295 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# =============================================================================
+
+import typing
+
+import tensorflow as tf
+
+
+@tf.keras.utils.register_keras_serializable(package="Addons")
+class MultiHeadAttention(tf.keras.layers.Layer):
+    r"""
+    MultiHead Attention layer.
+
+    Defines the MultiHead Attention operation as defined in
+    [Attention Is All You Need](https://arxiv.org/abs/1706.03762) which takes
+    in a `query`, `key` and `value` tensors returns the dot-product attention
+    between them:
+
+        ```python
+        mha = MultiHeadAttention(head_size=128, num_heads=128)
+
+        query = tf.random.uniform((32, 20, 200)) # (batch_size, query_elements, query_depth)
+        key = tf.random.uniform((32, 15, 300)) # (batch_size, key_elements, key_depth)
+        value = tf.random.uniform((32, 15, 400)) # (batch_size, key_elements, value_depth)
+
+        attention = mha([query, key, value]) # (batch_size, query_elements, value_depth)
+        ```
+
+    If `value` is not given then internally `value = key` will be used:
+
+         ```python
+        mha = MultiHeadAttention(head_size=128, num_heads=128)
+
+        query = tf.random.uniform((32, 20, 200)) # (batch_size, query_elements, query_depth)
+        key = tf.random.uniform((32, 15, 300)) # (batch_size, key_elements, key_depth)
+
+        attention = mha([query, key]) # (batch_size, query_elements, key_depth)
+        ```
+
+    Arguments
+        head_size: int, dimensionality of the `query`, `key` and `value` tensors
+        after the linear transformation.
+        num_heads: int, number of attention heads.
+        output_size: int, dimensionality of the output space, if `None` then the
+        input dimension of
+        `value` or `key` will be used, default `None`.
+        dropout: float, `rate` parameter for the dropout layer that is
+        applied to attention after softmax,
+        default `0`.
+        use_projection_bias: bool, whether to use a bias term after the linear
+        output projection.
+        return_attn_coef: bool, if `True`, return the attention coefficients as
+        an additional output argument.
+        kernel_initializer: initializer, initializer for the kernel weights.
+        kernel_regularizer: regularizer, regularizer for the kernel weights.
+        kernel_constraint: constraint, constraint for the kernel weights.
+        bias_initializer: initializer, initializer for the bias weights.
+        bias_regularizer: regularizer, regularizer for the bias weights.
+        bias_constraint: constraint, constraint for the bias weights.
+
+    Call Arguments
+        inputs:  List of the following tensors:
+            * `query`: Tensor of shape `(..., query_elements, query_depth)`
+            * `key`: `Tensor of shape '(..., key_elements, key_depth)`
+            * `value`: Tensor of shape `(..., key_elements, value_depth)` (optional)
+        mask: a binary Tensor of shape `[batch_size?, num_heads?, query_elements, key_elements]`
+        which specifies which query elements can attendo to which key elements,
+        `1` indicates attention and `0` indicates no attention.
+
+    Output shape
+        - `(..., query_elements, output_size)` if `output_size` is given, else
+        - `(..., query_elements, value_depth)` if `value` is given, else
+        - `(..., query_elements, key_depth)`
+    """
+
+    def __init__(
+        self,
+        head_size: int,
+        num_heads: int,
+        output_size: int = None,
+        dropout: float = 0.0,
+        use_projection_bias: bool = True,
+        return_attn_coef: bool = False,
+        kernel_initializer: typing.Union[str, typing.Callable] = "glorot_uniform",
+        kernel_regularizer: typing.Union[str, typing.Callable] = None,
+        kernel_constraint: typing.Union[str, typing.Callable] = None,
+        bias_initializer: typing.Union[str, typing.Callable] = "zeros",
+        bias_regularizer: typing.Union[str, typing.Callable] = None,
+        bias_constraint: typing.Union[str, typing.Callable] = None,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        if output_size is not None and output_size < 1:
+            raise ValueError("output_size must be a positive number")
+
+        self.head_size = head_size
+        self.num_heads = num_heads
+        self.output_size = output_size
+        self.use_projection_bias = use_projection_bias
+        self.return_attn_coef = return_attn_coef
+
+        self.kernel_initializer = tf.keras.initializers.get(kernel_initializer)
+        self.kernel_regularizer = tf.keras.regularizers.get(kernel_regularizer)
+        self.kernel_constraint = tf.keras.constraints.get(kernel_constraint)
+        self.bias_initializer = tf.keras.initializers.get(bias_initializer)
+        self.bias_regularizer = tf.keras.regularizers.get(bias_regularizer)
+        self.bias_constraint = tf.keras.constraints.get(bias_constraint)
+
+        self.dropout = tf.keras.layers.Dropout(dropout)
+        self._droput_rate = dropout
+
+    def build(self, input_shape):
+
+        num_query_features = input_shape[0][-1]
+        num_key_features = input_shape[1][-1]
+        num_value_features = (
+            input_shape[2][-1] if len(input_shape) > 2 else num_key_features
+        )
+        output_size = (
+            self.output_size if self.output_size is not None else num_value_features
+        )
+
+        self.query_kernel = self.add_weight(
+            name="query_kernel",
+            shape=[self.num_heads, num_query_features, self.head_size],
+            initializer=self.kernel_initializer,
+            regularizer=self.kernel_regularizer,
+            constraint=self.kernel_constraint,
+        )
+        self.key_kernel = self.add_weight(
+            name="key_kernel",
+            shape=[self.num_heads, num_key_features, self.head_size],
+            initializer=self.kernel_initializer,
+            regularizer=self.kernel_regularizer,
+            constraint=self.kernel_constraint,
+        )
+        self.value_kernel = self.add_weight(
+            name="value_kernel",
+            shape=[self.num_heads, num_value_features, self.head_size],
+            initializer=self.kernel_initializer,
+            regularizer=self.kernel_regularizer,
+            constraint=self.kernel_constraint,
+        )
+        self.projection_kernel = self.add_weight(
+            name="projection_kernel",
+            shape=[self.num_heads, self.head_size, output_size],
+            initializer=self.kernel_initializer,
+            regularizer=self.kernel_regularizer,
+            constraint=self.kernel_constraint,
+        )
+
+        if self.use_projection_bias:
+            self.projection_bias = self.add_weight(
+                name="projection_bias",
+                shape=[output_size],
+                initializer=self.bias_initializer,
+                regularizer=self.bias_regularizer,
+                constraint=self.bias_constraint,
+            )
+        else:
+            self.projection_bias = None
+
+        super().build(input_shape)
+
+    def call(self, inputs, training=None, mask=None):
+
+        # einsum nomenclature
+        # ------------------------
+        # N = query elements
+        # M = key/value elements
+        # H = heads
+        # I = input features
+        # O = output features
+
+        query = inputs[0]
+        key = inputs[1]
+        value = inputs[2] if len(inputs) > 2 else key
+
+        # verify shapes
+        if mask is not None:
+            if len(mask.shape) < 2:
+                raise ValueError("'mask' must have atleast 2 dimensions")
+            if query.shape[-2] != mask.shape[-2]:
+                raise ValueError(
+                    "mask's second to last dimension must be equal to the number of elements in 'query'"
+                )
+            if key.shape[-2] != mask.shape[-1]:
+                raise ValueError(
+                    "mask's last dimension must be equal to the number of elements in 'key'"
+                )
+            if key.shape[-2] != value.shape[-2]:
+                raise ValueError(
+                    "the number of elements in 'key' must be equal to the same as the number of elements in 'value'"
+                )
+
+        # Linear transformations
+        query = tf.einsum("...NI , HIO -> ...NHO", query, self.query_kernel)
+        key = tf.einsum("...MI , HIO -> ...MHO", key, self.key_kernel)
+        value = tf.einsum("...MI , HIO -> ...MHO", value, self.value_kernel)
+
+        # Scale dot-product, doing the division to either query or key
+        # instead of their product saves some computation
+        depth = tf.constant(self.head_size, dtype=tf.float32)
+        query /= tf.sqrt(depth)
+
+        # Calculate dot product attention
+        logits = tf.einsum("...NHO,...MHO->...HNM", query, key)
+
+        # apply mask
+        if mask is not None:
+            mask = tf.cast(mask, tf.float32)
+
+            # possibly expand on the head dimension so broadcasting works
+            if len(mask.shape) != len(logits.shape):
+                mask = tf.expand_dims(mask, -3)
+
+            logits += -10e9 * (1.0 - mask)
+
+        attn_coef = tf.nn.softmax(logits)
+
+        # attention dropout
+        attn_coef_dropout = self.dropout(attn_coef, training=training)
+
+        # attention * value
+        multihead_output = tf.einsum("...HNM,...MHI->...NHI", attn_coef_dropout, value)
+
+        # Run the outputs through another linear projection layer. Recombining heads
+        # is automatically done.
+        output = tf.einsum(
+            "...NHI,HIO->...NO", multihead_output, self.projection_kernel
+        )
+
+        if self.projection_bias is not None:
+            output += self.projection_bias
+
+        if self.return_attn_coef:
+            return output, attn_coef
+        else:
+            return output
+
+    def compute_output_shape(self, input_shape):
+        num_value_features = (
+            input_shape[2][-1] if len(input_shape) > 2 else input_shape[1][-1]
+        )
+        output_size = (
+            self.output_size if self.output_size is not None else num_value_features
+        )
+
+        output_shape = input_shape[0][:-1] + (output_size,)
+
+        if self.return_attn_coef:
+            num_query_elements = input_shape[0][-2]
+            num_key_elements = input_shape[1][-2]
+            attn_coef_shape = input_shape[0][:-2] + (
+                self.num_heads,
+                num_query_elements,
+                num_key_elements,
+            )
+
+            return output_shape, attn_coef_shape
+        else:
+            return output_shape
+
+    def get_config(self):
+        config = super().get_config()
+
+        config.update(
+            head_size=self.head_size,
+            num_heads=self.num_heads,
+            output_size=self.output_size,
+            dropout=self._droput_rate,
+            use_projection_bias=self.use_projection_bias,
+            return_attn_coef=self.return_attn_coef,
+            kernel_initializer=tf.keras.initializers.serialize(self.kernel_initializer),
+            kernel_regularizer=tf.keras.regularizers.serialize(self.kernel_regularizer),
+            kernel_constraint=tf.keras.constraints.serialize(self.kernel_constraint),
+            bias_initializer=tf.keras.initializers.serialize(self.bias_initializer),
+            bias_regularizer=tf.keras.regularizers.serialize(self.bias_regularizer),
+            bias_constraint=tf.keras.constraints.serialize(self.bias_constraint),
+        )
+
+        return config
diff --git a/tensorflow_addons/layers/multihead_attention_test.py b/tensorflow_addons/layers/multihead_attention_test.py
