ray/rllib/models/tf/layers/relative_multi_head_attention.py

139 lines
5.5 KiB
Python

from typing import Optional
from ray.rllib.utils.framework import try_import_tf
from ray.rllib.utils.typing import TensorType
tf1, tf, tfv = try_import_tf()
class RelativeMultiHeadAttention(tf.keras.layers.Layer if tf else object):
"""A RelativeMultiHeadAttention layer as described in [3].
Uses segment level recurrence with state reuse.
"""
def __init__(self,
out_dim: int,
num_heads: int,
head_dim: int,
input_layernorm: bool = False,
output_activation: Optional["tf.nn.activation"] = None,
**kwargs):
"""Initializes a RelativeMultiHeadAttention keras Layer object.
Args:
out_dim (int): The output dimensions of the multi-head attention
unit.
num_heads (int): The number of attention heads to use.
Denoted `H` in [2].
head_dim (int): The dimension of a single(!) attention head within
a multi-head attention unit. Denoted as `d` in [3].
input_layernorm (bool): Whether to prepend a LayerNorm before
everything else. Should be True for building a GTrXL.
output_activation (Optional[tf.nn.activation]): Optional tf.nn
activation function. Should be relu for GTrXL.
**kwargs:
"""
super().__init__(**kwargs)
# No bias or non-linearity.
self._num_heads = num_heads
self._head_dim = head_dim
# 3=Query, key, and value inputs.
self._qkv_layer = tf.keras.layers.Dense(
3 * num_heads * head_dim, use_bias=False)
self._linear_layer = tf.keras.layers.TimeDistributed(
tf.keras.layers.Dense(
out_dim, use_bias=False, activation=output_activation))
self._uvar = self.add_weight(shape=(num_heads, head_dim))
self._vvar = self.add_weight(shape=(num_heads, head_dim))
# Constant (non-trainable) sinusoid rel pos encoding matrix, which
# depends on this incoming time dimension.
# For inference, we prepend the memory to the current timestep's
# input: Tau + 1. For training, we prepend the memory to the input
# sequence: Tau + T.
self._pos_embedding = PositionalEmbedding(out_dim)
self._pos_proj = tf.keras.layers.Dense(
num_heads * head_dim, use_bias=False)
self._input_layernorm = None
if input_layernorm:
self._input_layernorm = tf.keras.layers.LayerNormalization(axis=-1)
def call(self, inputs: TensorType,
memory: Optional[TensorType] = None) -> TensorType:
T = tf.shape(inputs)[1] # length of segment (time)
H = self._num_heads # number of attention heads
d = self._head_dim # attention head dimension
# Add previous memory chunk (as const, w/o gradient) to input.
# Tau (number of (prev) time slices in each memory chunk).
Tau = tf.shape(memory)[1]
inputs = tf.concat([tf.stop_gradient(memory), inputs], axis=1)
# Apply the Layer-Norm.
if self._input_layernorm is not None:
inputs = self._input_layernorm(inputs)
qkv = self._qkv_layer(inputs)
queries, keys, values = tf.split(qkv, 3, -1)
# Cut out memory timesteps from query.
queries = queries[:, -T:]
# Splitting up queries into per-head dims (d).
queries = tf.reshape(queries, [-1, T, H, d])
keys = tf.reshape(keys, [-1, Tau + T, H, d])
values = tf.reshape(values, [-1, Tau + T, H, d])
R = self._pos_embedding(Tau + T)
R = self._pos_proj(R)
R = tf.reshape(R, [Tau + T, H, d])
# b=batch
# i and j=time indices (i=max-timesteps (inputs); j=Tau memory space)
# h=head
# d=head-dim (over which we will reduce-sum)
score = tf.einsum("bihd,bjhd->bijh", queries + self._uvar, keys)
pos_score = tf.einsum("bihd,jhd->bijh", queries + self._vvar, R)
score = score + self.rel_shift(pos_score)
score = score / d**0.5
# Causal mask of the same length as the sequence.
mask = tf.sequence_mask(
tf.range(Tau + 1, Tau + T + 1), dtype=score.dtype)
mask = mask[None, :, :, None]
masked_score = score * mask + 1e30 * (mask - 1.)
wmat = tf.nn.softmax(masked_score, axis=2)
out = tf.einsum("bijh,bjhd->bihd", wmat, values)
out = tf.reshape(out, tf.concat((tf.shape(out)[:2], [H * d]), axis=0))
return self._linear_layer(out)
@staticmethod
def rel_shift(x: TensorType) -> TensorType:
# Transposed version of the shift approach described in [3].
# https://github.com/kimiyoung/transformer-xl/blob/
# 44781ed21dbaec88b280f74d9ae2877f52b492a5/tf/model.py#L31
x_size = tf.shape(x)
x = tf.pad(x, [[0, 0], [0, 0], [1, 0], [0, 0]])
x = tf.reshape(x, [x_size[0], x_size[2] + 1, x_size[1], x_size[3]])
x = x[:, 1:, :, :]
x = tf.reshape(x, x_size)
return x
class PositionalEmbedding(tf.keras.layers.Layer if tf else object):
def __init__(self, out_dim, **kwargs):
super().__init__(**kwargs)
self.inverse_freq = 1 / (10000**(tf.range(0, out_dim, 2.0) / out_dim))
def call(self, seq_length):
pos_offsets = tf.cast(tf.range(seq_length - 1, -1, -1), tf.float32)
inputs = pos_offsets[:, None] * self.inverse_freq[None, :]
return tf.concat((tf.sin(inputs), tf.cos(inputs)), axis=-1)