# coding=utf-8
# Copyright 2020 The Trax Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION. 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.
"""PyTorch REFORMER model. """
import sys
from collections import namedtuple
from dataclasses import dataclass
from functools import reduce
from operator import mul
from typing import List, Optional, Tuple
import numpy as np
import torch
from torch import nn
from torch.autograd.function import Function
from torch.nn import CrossEntropyLoss, MSELoss
from ...activations import ACT2FN
from ...file_utils import (
DUMMY_INPUTS,
DUMMY_MASK,
ModelOutput,
add_code_sample_docstrings,
add_start_docstrings,
add_start_docstrings_to_model_forward,
)
from ...modeling_outputs import CausalLMOutput, MaskedLMOutput, QuestionAnsweringModelOutput, SequenceClassifierOutput
from ...modeling_utils import PreTrainedModel, apply_chunking_to_forward
from ...utils import logging
from .configuration_reformer import ReformerConfig
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "google/reformer-crime-and-punishment"
_CONFIG_FOR_DOC = "ReformerConfig"
_TOKENIZER_FOR_DOC = "ReformerTokenizer"
REFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = [
"google/reformer-crime-and-punishment",
"google/reformer-enwik8",
# See all Reformer models at https://huggingface.co/models?filter=reformer
]
# Define named tuples for nn.Modules here
LSHSelfAttentionOutput = namedtuple("LSHSelfAttentionOutput", ["hidden_states", "attention_probs", "buckets"])
LocalSelfAttentionOutput = namedtuple("LocalSelfAttentionOutput", ["hidden_states", "attention_probs"])
AttentionOutput = namedtuple("AttentionOutput", ["hidden_states", "attention_probs", "buckets"])
ReformerOutput = namedtuple("ReformerOutput", ["hidden_states", "attn_output", "attention_probs", "buckets"])
ReformerBackwardOutput = namedtuple(
"ReformerBackwardOutput", ["attn_output", "hidden_states", "grad_attn_output", "grad_hidden_states"]
)
ReformerEncoderOutput = namedtuple(
"ReformerEncoderOutput",
["hidden_states", "all_hidden_states", "all_attentions", "past_buckets_states"],
)
def _stable_argsort(vector, dim):
# this function scales the vector so that torch.argsort is stable.
# torch.argsort is not stable on its own
scale_offset = torch.arange(vector.shape[dim], device=vector.device).view(1, 1, -1)
scale_offset = scale_offset.expand(vector.shape)
scaled_vector = vector.shape[dim] * vector + (scale_offset % vector.shape[dim])
return torch.argsort(scaled_vector, dim=dim)
def _get_least_common_mult_chunk_len(config):
attn_types = config.attn_layers
attn_types_set = set(attn_types)
if len(attn_types_set) == 1 and attn_types[0] == "lsh":
return config.lsh_attn_chunk_length
elif len(attn_types_set) == 1 and attn_types[0] == "local":
return config.local_attn_chunk_length
elif len(attn_types_set) == 2 and attn_types_set == set(["lsh", "local"]):
return np.lcm(config.lsh_attn_chunk_length, config.local_attn_chunk_length)
else:
raise NotImplementedError(
"Only attn layer types 'lsh' and 'local' exist, but `config.attn_layers`: {}. Select attn layer types from ['lsh', 'local'] only.".format(
config.attn_layers
)
)
def _get_min_chunk_len(config):
attn_types = config.attn_layers
attn_types_set = set(attn_types)
if len(attn_types_set) == 1 and attn_types[0] == "lsh":
return config.lsh_attn_chunk_length
elif len(attn_types_set) == 1 and attn_types[0] == "local":
return config.local_attn_chunk_length
elif len(attn_types_set) == 2 and attn_types_set == set(["lsh", "local"]):
return min(config.lsh_attn_chunk_length, config.local_attn_chunk_length)
else:
raise NotImplementedError(
"Only attn layer types 'lsh' and 'local' exist, but `config.attn_layers`: {}. Select attn layer types from ['lsh', 'local'] only.".format(
config.attn_layers
)
)
class AxialPositionEmbeddings(nn.Module):
"""
Constructs axial position embeddings. Useful for very long input sequences to save memory and time.
"""
def __init__(self, config):
super().__init__()
self.axial_pos_shape = config.axial_pos_shape
self.axial_pos_embds_dim = config.axial_pos_embds_dim
self.dropout = config.hidden_dropout_prob
self.least_common_mult_chunk_length = _get_least_common_mult_chunk_len(config)
self.weights = nn.ParameterList()
assert (
sum(self.axial_pos_embds_dim) == config.hidden_size
), "Make sure that config.axial_pos_embds factors: {} sum to config.hidden_size: {}".format(
self.axial_pos_embds_dim, config.hidden_size
)
# create weights
for axis, axial_pos_embd_dim in enumerate(self.axial_pos_embds_dim):
# create expanded shapes
ax_shape = [1] * len(self.axial_pos_shape)
ax_shape[axis] = self.axial_pos_shape[axis]
ax_shape = tuple(ax_shape) + (axial_pos_embd_dim,)
# create tensor and init
self.weights.append(nn.Parameter(torch.ones(ax_shape, dtype=torch.float32)))
def forward(self, position_ids):
# broadcast weights to correct shape
batch_size = position_ids.shape[0]
sequence_length = position_ids.shape[1]
broadcasted_weights = [
weight.expand((batch_size,) + self.axial_pos_shape + weight.shape[-1:]) for weight in self.weights
]
if self.training is True:
assert (
reduce(mul, self.axial_pos_shape) == sequence_length
), "If training, make sure that config.axial_pos_shape factors: {} multiply to sequence length. Got prod({}) != sequence_length: {}. You might want to consider padding your sequence length to {} or changing config.axial_pos_shape.".format(
self.axial_pos_shape, self.axial_pos_shape, sequence_length, reduce(mul, self.axial_pos_shape)
)
if self.dropout > 0:
weights = torch.cat(broadcasted_weights, dim=-1)
# permute weights so that 2D correctly drops dims 1 and 2
transposed_weights = weights.transpose(2, 1)
# drop entire matrix of last two dims (prev dims 1 and 2)
dropped_transposed_weights = nn.functional.dropout2d(
transposed_weights, p=self.dropout, training=self.training
)
dropped_weights = dropped_transposed_weights.transpose(2, 1)
position_encodings = torch.reshape(dropped_weights, (batch_size, sequence_length, -1))
else:
position_encodings = torch.cat(
[torch.reshape(weight, (batch_size, sequence_length, -1)) for weight in broadcasted_weights],
dim=-1,
)
else:
assert (
reduce(mul, self.axial_pos_shape) >= sequence_length
), "Make sure that config.axial_pos_shape factors: {} multiply at least to max(sequence_length, least_common_mult_chunk_length): max({}, {})".format(
self.axial_pos_shape,
sequence_length,
self.least_common_mult_chunk_length,
)
# compute how many columns are needed
max_position_id = position_ids.max().item()
required_pos_encodings_columns = -(-(max_position_id + 1) // self.axial_pos_shape[1])
# cut to columns that are needed
position_encodings = torch.cat(
[weight[:, :required_pos_encodings_columns] for weight in broadcasted_weights], dim=-1
)
position_encodings = torch.reshape(position_encodings, (batch_size, -1, position_encodings.shape[-1]))
# select correct position encodings
position_encodings = torch.cat(
[
torch.index_select(position_encodings[i], 0, position_ids[i]).unsqueeze(0)
for i in range(batch_size)
],
dim=0,
)
return position_encodings
class PositionEmbeddings(nn.Module):
"""Constructs conventional position embeddings of shape `[max_pos_embeddings, hidden_size]`."""
def __init__(self, config):
super().__init__()
self.dropout = config.hidden_dropout_prob
self.embedding = nn.Embedding(config.max_position_embeddings, config.hidden_size)
def forward(self, position_ids):
position_embeddings = self.embedding(position_ids)
position_embeddings = nn.functional.dropout(position_embeddings, p=self.dropout, training=self.training)
return position_embeddings
class ReformerEmbeddings(nn.Module):
"""Construct the embeddings from word, position and token_type embeddings."""
def __init__(self, config):
super().__init__()
self.max_position_embeddings = config.max_position_embeddings
self.dropout = config.hidden_dropout_prob
self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
self.position_embeddings = (
AxialPositionEmbeddings(config) if config.axial_pos_embds else PositionEmbeddings(config)
)
def forward(self, input_ids=None, position_ids=None, inputs_embeds=None, start_idx_pos_encodings=0):
if input_ids is not None:
input_shape = input_ids.size()
device = input_ids.device
else:
input_shape = inputs_embeds.size()[:-1]
device = inputs_embeds.device
seq_length = input_shape[1]
if position_ids is None:
position_ids = torch.arange(
start_idx_pos_encodings, start_idx_pos_encodings + seq_length, dtype=torch.long, device=device
)
position_ids = position_ids.unsqueeze(0).expand(input_shape)
if inputs_embeds is None:
inputs_embeds = self.word_embeddings(input_ids)
assert (
position_ids.shape[-1] <= self.max_position_embeddings
), "Sequence Length: {} has to be larger equal than config.max_position_embeddings: {}".format(
position_ids.shape[-1], self.max_position_embeddings
)
# dropout
embeddings = nn.functional.dropout(inputs_embeds, p=self.dropout, training=self.training)
# add positional embeddings
position_embeddings = self.position_embeddings(position_ids)
embeddings = embeddings + position_embeddings
return embeddings
class EfficientAttentionMixin:
"""
A few utilities for nn.Modules in Reformer, to be used as a mixin.
"""
def _look_adjacent(self, vectors, num_chunks_before, num_chunks_after):
"""
Used to implement attention between consecutive chunks.
Args:
vectors: array of shape [batch_size, num_attention_heads, n_chunks, chunk_len, ...]
num_chunks_before: chunks before current chunk to include in attention
num_chunks_after: chunks after current chunk to include in attention
Returns:
tensor of shape [num_chunks, N * chunk_length, ...], where N = (1 + num_chunks_before + num_chunks_after).
"""
if num_chunks_before == 0 and num_chunks_after == 0:
return vectors
slices = []
for i in range(-num_chunks_before, num_chunks_after + 1):
if i == 0:
slices.append(vectors)
else:
slices.append(torch.cat([vectors[:, :, i:, ...], vectors[:, :, :i, ...]], dim=2))
return torch.cat(slices, dim=3)
def _split_hidden_size_dim(self, x, num_attn_heads, attn_head_size):
"""
splits hidden_size dim into attn_head_size and num_attn_heads
"""
new_x_shape = x.size()[:-1] + (num_attn_heads, attn_head_size)
x = x.view(*new_x_shape)
return x.transpose(2, 1)
def _merge_hidden_size_dims(self, x, num_attn_heads, attn_head_size):
"""
merges attn_head_size dim and num_attn_heads dim into hidden_size
"""
x = x.permute(0, 2, 1, 3)
return torch.reshape(x, (x.size()[0], -1, num_attn_heads * attn_head_size))
def _split_seq_length_dim_to(self, vectors, dim_factor_1, dim_factor_2, num_attn_heads, attn_head_size=None):
"""
splits sequence length dim of vectors into `dim_factor_1` and `dim_factor_2` dims
"""
batch_size = vectors.shape[0]
split_dim_shape = (batch_size, num_attn_heads, dim_factor_1, dim_factor_2)
if len(vectors.shape) == 4:
return torch.reshape(vectors, split_dim_shape + (attn_head_size,))
elif len(vectors.shape) == 3:
return torch.reshape(vectors, split_dim_shape)
else:
raise ValueError("Input vector rank should be one of [3, 4], but is: {}".format(len(vectors.shape)))
class LSHSelfAttention(nn.Module, EfficientAttentionMixin):
def __init__(self, config):
super().__init__()
self.config = config
self.chunk_length = config.lsh_attn_chunk_length
self.num_hashes = config.num_hashes
self.num_buckets = config.num_buckets
self.num_chunks_before = config.lsh_num_chunks_before
self.num_chunks_after = config.lsh_num_chunks_after
self.hash_seed = config.hash_seed
self.is_decoder = config.is_decoder
self.max_position_embeddings = config.max_position_embeddings
self.dropout = config.lsh_attention_probs_dropout_prob
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = config.attention_head_size
self.all_head_size = self.num_attention_heads * self.attention_head_size
self.hidden_size = config.hidden_size
# projection matrices
self.query_key = nn.Linear(self.hidden_size, self.all_head_size, bias=False)
self.value = nn.Linear(self.hidden_size, self.all_head_size, bias=False)
# save mask value here. Need fp32 and fp16 mask values
self.register_buffer("self_mask_value_float16", torch.tensor(-1e3))
self.register_buffer("self_mask_value_float32", torch.tensor(-1e5))
self.register_buffer("mask_value_float16", torch.tensor(-1e4))
self.register_buffer("mask_value_float32", torch.tensor(-1e9))
def forward(
self,
hidden_states,
attention_mask=None,
head_mask=None,
num_hashes=None,
buckets=None,
past_buckets_states=None,
use_cache=False,
output_attentions=False,
**kwargs
):
sequence_length = hidden_states.shape[1]
batch_size = hidden_states.shape[0]
# num hashes can optionally be overwritten by user
num_hashes = num_hashes if num_hashes is not None else self.num_hashes
do_cached_attention = use_cache and past_buckets_states[1] is not None
# check if cache shall be used and that hidden states are already cached
if do_cached_attention:
assert (
sequence_length == 1
), f"At the moment, auto-regressive language generation is only possible one word at a time. Make sure that input sequence length {sequence_length} equals 1, when `past_buckets_states` is passed."
past_buckets = past_buckets_states[0]
past_states = past_buckets_states[1]
# get query vector
query_vectors = self.query_key(hidden_states)
query_vectors = self._split_hidden_size_dim(
query_vectors, self.num_attention_heads, self.attention_head_size
)
if past_buckets is not None:
key_value_hidden_states, sorted_bucket_idx, buckets = self._get_relevant_hid_states_and_buckets(
query_vectors=query_vectors,
attention_mask=attention_mask,
num_hashes=num_hashes,
hidden_states=hidden_states,
past_states=past_states,
past_buckets=past_buckets,
)
query_key_vectors = self._query_per_attn_head(key_value_hidden_states)
value_vectors = self._value_per_attn_head(key_value_hidden_states)
# split key & value vectors by num hashes to apply
# self attention on each separately
query_key_vectors = self._split_seq_length_dim_to(
query_key_vectors,
num_hashes,
-1,
self.num_attention_heads,
self.attention_head_size,
)
value_vectors = self._split_seq_length_dim_to(
value_vectors,
num_hashes,
-1,
self.num_attention_heads,
self.attention_head_size,
)
# repeat query vectors across hash dimension
query_vectors = query_vectors.unsqueeze(2).repeat(1, 1, num_hashes, 1, 1)
else:
key_value_hidden_states = torch.cat([past_states, hidden_states], dim=1)
query_key_vectors = self.query_key(key_value_hidden_states)
value_vectors = self.value(key_value_hidden_states)
else:
# project hidden_states to query_key and value
query_vectors = None
query_key_vectors = self.query_key(hidden_states)
value_vectors = self.value(hidden_states)
# if query key is not already split
if not do_cached_attention or past_buckets is None:
query_key_vectors = self._split_hidden_size_dim(
query_key_vectors, self.num_attention_heads, self.attention_head_size
)
value_vectors = self._split_hidden_size_dim(
value_vectors, self.num_attention_heads, self.attention_head_size
)
# cache buckets for next incremental decoding
if do_cached_attention and past_buckets is None and key_value_hidden_states.shape[1] >= self.chunk_length:
buckets = self._hash_vectors(query_key_vectors, num_hashes, attention_mask)
# free memory
del hidden_states
assert (
query_key_vectors.shape[-1] == self.attention_head_size
), "last dim of query_key_vectors is {} but should be {}.".format(
query_key_vectors.shape[-1], self.attention_head_size
)
assert (
value_vectors.shape[-1] == self.attention_head_size
), "last dim of value_vectors is {} but should be {}.".format(
value_vectors.shape[-1], self.attention_head_size
)
do_standard_self_attention = (sequence_length <= self.chunk_length) or (
use_cache and past_buckets_states[1] is not None
)
# LSH attention only makes sense if chunked attention should be performed
if not do_standard_self_attention:
# set `num_buckets` on the fly, recommended way to do it
if self.num_buckets is None:
self._set_num_buckets(sequence_length)
# use cached buckets for backprop only
if buckets is None:
# hash query key vectors into buckets
buckets = self._hash_vectors(query_key_vectors, num_hashes, attention_mask)
else:
# make sure buckets has correct shape for LSH attention
buckets = buckets.view(batch_size, self.num_attention_heads, num_hashes * sequence_length)
assert (
int(buckets.shape[-1]) == num_hashes * sequence_length
), "last dim of buckets is {}, but should be {}".format(buckets.shape[-1], num_hashes * sequence_length)
sorted_bucket_idx, undo_sorted_bucket_idx = self._get_sorted_bucket_idx_and_undo_sorted_bucket_idx(
sequence_length, buckets, num_hashes
)
# make sure bucket idx is not longer then sequence length
sorted_bucket_idx_per_hash = sorted_bucket_idx % sequence_length
# cluster query key value vectors according to hashed buckets
query_key_vectors = self._gather_by_expansion(query_key_vectors, sorted_bucket_idx_per_hash, num_hashes)
value_vectors = self._gather_by_expansion(value_vectors, sorted_bucket_idx_per_hash, num_hashes)
query_key_vectors = self._split_seq_length_dim_to(
query_key_vectors,
-1,
self.chunk_length,
self.num_attention_heads,
self.attention_head_size,
)
value_vectors = self._split_seq_length_dim_to(
value_vectors,
-1,
self.chunk_length,
self.num_attention_heads,
self.attention_head_size,
)
if self.chunk_length is None:
assert (
self.num_chunks_before == 0 and self.num_chunks_after == 0
), "If `config.chunk_length` is `None`, make sure `config.num_chunks_after` and `config.num_chunks_before` are set to 0."
elif do_cached_attention and past_buckets is not None:
# use max sequence length
sorted_bucket_idx_per_hash = sorted_bucket_idx
else:
# get sequence length indices
sorted_bucket_idx_per_hash = torch.arange(sequence_length, device=query_key_vectors.device).repeat(
batch_size, self.num_attention_heads, 1
)
# scale key vectors
key_vectors = self._len_and_dim_norm(query_key_vectors)
# set query_vectors to query key vectors if LSH self attention
query_vectors = query_vectors if query_vectors is not None else query_key_vectors
# free memory
del query_key_vectors
# get attention probs
out_vectors, logits, attention_probs = self._attend(
query_vectors=query_vectors,
key_vectors=key_vectors,
value_vectors=value_vectors,
sorted_bucket_idx_per_hash=sorted_bucket_idx_per_hash,
attention_mask=attention_mask,
head_mask=head_mask,
do_standard_self_attention=do_standard_self_attention,
do_cached_attention=do_cached_attention,
)
# free memory
del key_vectors, value_vectors
# re-order out_vectors and logits
if not do_standard_self_attention:
# sort clusters back to correct ordering
out_vectors, logits = ReverseSort.apply(out_vectors, logits, sorted_bucket_idx, undo_sorted_bucket_idx)
if not do_standard_self_attention or (do_cached_attention and past_buckets is not None):
# sum up all hash rounds
if num_hashes > 1:
out_vectors = self._split_seq_length_dim_to(
out_vectors,
num_hashes,
sequence_length,
self.num_attention_heads,
self.attention_head_size,
)
logits = self._split_seq_length_dim_to(
logits,
num_hashes,
sequence_length,
self.num_attention_heads,
self.attention_head_size,
).unsqueeze(-1)
probs_vectors = torch.exp(logits - torch.logsumexp(logits, dim=2, keepdim=True))
out_vectors = torch.sum(out_vectors * probs_vectors, dim=2)
# free memory
del probs_vectors
# free memory
del logits
assert out_vectors.shape == (
batch_size,
self.num_attention_heads,
sequence_length,
self.attention_head_size,
), "out_vectors have be of shape `[batch_size, config.num_attention_heads, sequence_length, config.attention_head_size]`."
out_vectors = self._merge_hidden_size_dims(out_vectors, self.num_attention_heads, self.attention_head_size)
if output_attentions is False:
attention_probs = ()
if buckets is not None:
buckets = buckets.view(batch_size, self.num_attention_heads, num_hashes, -1)
return LSHSelfAttentionOutput(hidden_states=out_vectors, attention_probs=attention_probs, buckets=buckets)
def _query_per_attn_head(self, hidden_states):
per_head_query_key = self.query_key.weight.reshape(
self.num_attention_heads, self.attention_head_size, self.hidden_size
).transpose(-2, -1)
# only relevant for inference and no bias => we can use einsum here
query_key_vectors = torch.einsum("balh,ahr->balr", hidden_states, per_head_query_key)
return query_key_vectors
def _value_per_attn_head(self, hidden_states):
per_head_value = self.value.weight.reshape(
self.num_attention_heads, self.attention_head_size, self.hidden_size
).transpose(-2, -1)
# only relevant for inference and no bias => we can use einsum here
value_vectors = torch.einsum("balh,ahr->balr", hidden_states, per_head_value)
return value_vectors
def _hash_vectors(self, vectors, num_hashes, attention_mask, increase_num_buckets=False):
batch_size = vectors.shape[0]
# See https://arxiv.org/pdf/1509.02897.pdf
# We sample a different random rotation for each round of hashing to
# decrease the probability of hash misses.
if isinstance(self.num_buckets, int):
assert (
self.num_buckets % 2 == 0
), "There should be an even number of bucktes, but `self.num_bucktes`: {}".format(self.num_buckets)
rotation_size = self.num_buckets
num_buckets = self.num_buckets
else:
# Factorize the hash if self.num_buckets is a list or tuple
rotation_size, num_buckets = 0, 1
for bucket_factor in self.num_buckets:
assert bucket_factor % 2 == 0, "The number of buckets should be even, but `num_bucket`: {}".format(
bucket_factor
)
rotation_size = rotation_size + bucket_factor
num_buckets = num_buckets * bucket_factor
# remove gradient
vectors = vectors.detach()
if self.hash_seed is not None:
# for determinism
torch.manual_seed(self.hash_seed)
rotations_shape = (self.num_attention_heads, vectors.shape[-1], num_hashes, rotation_size // 2)
# create a random self.attention_head_size x num_hashes x num_buckets/2
random_rotations = torch.randn(rotations_shape, device=vectors.device, dtype=vectors.dtype)
# Output dim: Batch_Size x Num_Attn_Heads x Num_Hashes x Seq_Len x Num_Buckets/2
rotated_vectors = torch.einsum("bmtd,mdhr->bmhtr", vectors, random_rotations)
if isinstance(self.num_buckets, int) or len(self.num_buckets) == 1:
rotated_vectors = torch.cat([rotated_vectors, -rotated_vectors], dim=-1)
buckets = torch.argmax(rotated_vectors, dim=-1)
else:
# Get the buckets for them and combine.
buckets, cur_sum, cur_product = None, 0, 1
for bucket_factor in self.num_buckets:
rotated_vectors_factor = rotated_vectors[..., cur_sum : cur_sum + (bucket_factor // 2)]
cur_sum = cur_sum + bucket_factor // 2
rotated_vectors_factor = torch.cat([rotated_vectors_factor, -rotated_vectors_factor], dim=-1)
if buckets is None:
buckets = torch.argmax(rotated_vectors_factor, dim=-1)
else:
buckets = buckets + (cur_product * torch.argmax(rotated_vectors_factor, dim=-1))
cur_product = cur_product * bucket_factor
if attention_mask is not None and (attention_mask.sum().item() < batch_size * attention_mask.shape[-1]):
# add an extra bucket for padding tokens only
num_buckets = num_buckets + 1
# assign padding tokens extra bucket
buckets_mask = attention_mask.to(torch.uint8)[:, None, None, :].expand(buckets.shape)
buckets = torch.where(
buckets_mask, buckets, torch.tensor(num_buckets - 1, dtype=torch.long, device=buckets.device)
)
elif increase_num_buckets:
num_buckets = num_buckets + 1
# buckets is now (Batch_size x Num_Attn_Heads x Num_Hashes x Seq_Len).
# Next we add offsets so that bucket numbers from different hashing rounds don't overlap.
offsets = torch.arange(num_hashes, device=vectors.device)
offsets = (offsets * num_buckets).view((1, 1, -1, 1))
# expand to batch size and num attention heads
offsets = offsets.expand((batch_size, self.num_attention_heads) + offsets.shape[-2:])
offset_buckets = (buckets + offsets).flatten(start_dim=2, end_dim=3)
return offset_buckets
def _get_sorted_bucket_idx_and_undo_sorted_bucket_idx(self, sequence_length, buckets, num_hashes):
# no gradients are needed
with torch.no_grad():
# hash-based sort
sorted_bucket_idx = _stable_argsort(buckets, dim=-1)
# create simple indices to scatter to, to have undo sort
indices = (
torch.arange(sorted_bucket_idx.shape[-1], device=buckets.device)
.view(1, 1, -1)
.expand(sorted_bucket_idx.shape)
)
# get undo sort
undo_sorted_bucket_idx = sorted_bucket_idx.new(*sorted_bucket_idx.size())
undo_sorted_bucket_idx.scatter_(-1, sorted_bucket_idx, indices)
return sorted_bucket_idx, undo_sorted_bucket_idx
def _set_num_buckets(self, sequence_length):
# `num_buckets` should be set to 2 * sequence_length // chunk_length as recommended in paper
num_buckets_pow_2 = (2 * (sequence_length // self.chunk_length)).bit_length() - 1
# make sure buckets are power of 2
num_buckets = 2 ** num_buckets_pow_2
# factorize `num_buckets` if `num_buckets` becomes too large
num_buckets_limit = 2 * max(
int((self.max_position_embeddings // self.chunk_length) ** (0.5)),
self.chunk_length,
)
if num_buckets > num_buckets_limit:
num_buckets = [2 ** (num_buckets_pow_2 // 2), 2 ** (num_buckets_pow_2 - num_buckets_pow_2 // 2)]
logger.warning("config.num_buckets is not set. Setting config.num_buckets to {}...".format(num_buckets))
# set num buckets in config to be properly saved
self.config.num_buckets = num_buckets
self.num_buckets = num_buckets
def _attend(
self,
query_vectors,
key_vectors,
value_vectors,
sorted_bucket_idx_per_hash,
attention_mask,
head_mask,
do_standard_self_attention,
do_cached_attention,
):
# look at previous and following chunks if chunked attention
if not do_standard_self_attention:
key_vectors = self._look_adjacent(key_vectors, self.num_chunks_before, self.num_chunks_after)
value_vectors = self._look_adjacent(value_vectors, self.num_chunks_before, self.num_chunks_after)
# get logits and dots
# (BS, NumAttn, NumHash x NumChunk, Chunk_L x Hidden),(BS, NumAttn, NumHash x NumChunk, Chunk_L * (Num_bef + Num_aft + 1) x Hidden) -> (BS, NumAttn, NumHash x NumChunk, Chunk_L, Chunk_L * (1 + Num_bef + Num_aft))
query_key_dots = torch.matmul(query_vectors, key_vectors.transpose(-1, -2))
# free memory
del query_vectors, key_vectors
# if chunked attention split bucket idxs to query and key
if not do_standard_self_attention:
query_bucket_idx = self._split_seq_length_dim_to(
sorted_bucket_idx_per_hash, -1, self.chunk_length, self.num_attention_heads
)
key_value_bucket_idx = self._look_adjacent(query_bucket_idx, self.num_chunks_before, self.num_chunks_after)
elif do_cached_attention and query_key_dots.ndim > 4:
key_value_bucket_idx = sorted_bucket_idx_per_hash
query_bucket_idx = (
key_value_bucket_idx.new_ones(key_value_bucket_idx.shape[:-1] + (1,)) * key_value_bucket_idx.max()
)
elif do_cached_attention and query_key_dots.ndim <= 4:
query_bucket_idx = (query_key_dots.shape[-1] - 1) * torch.ones_like(query_key_dots)[:, :, :, -1]
key_value_bucket_idx = torch.arange(
query_key_dots.shape[-1], dtype=torch.long, device=query_key_dots.device
)[None, None, :].expand(query_bucket_idx.shape[:2] + (-1,))
else:
query_bucket_idx = key_value_bucket_idx = sorted_bucket_idx_per_hash
# get correct mask values depending on precision
if query_key_dots.dtype == torch.float16:
self_mask_value = self.self_mask_value_float16.half()
mask_value = self.mask_value_float16.half()
else:
self_mask_value = self.self_mask_value_float32
mask_value = self.mask_value_float32
if not do_cached_attention:
mask = self._compute_attn_mask(
query_bucket_idx,
key_value_bucket_idx,
attention_mask,
query_key_dots.shape,
do_standard_self_attention,
)
if mask is not None:
query_key_dots = torch.where(mask, query_key_dots, mask_value)
# free memory
del mask
# Self mask is ALWAYS applied.
# From the reformer paper (https://arxiv.org/pdf/2001.04451.pdf):
# " While attention to the future is not allowed, typical implementations of the
# Transformer do allow a position to attend to itself.
# Such behavior is undesirable in a shared-QK formulation because the dot-product
# of a query vector with itself will almost always be greater than the dot product of a
# query vector with a vector at another position. We therefore modify the masking
# to forbid a token from attending to itself, except in situations
# where a token has no other valid attention targets (e.g. the first token in a sequence) "
self_mask = torch.ne(query_bucket_idx.unsqueeze(-1), key_value_bucket_idx.unsqueeze(-2)).to(
query_bucket_idx.device
)
# apply self_mask
query_key_dots = torch.where(self_mask, query_key_dots, self_mask_value)
# free memory
del self_mask
logits = torch.logsumexp(query_key_dots, dim=-1, keepdim=True)
# dots shape is `[batch_size, num_attn_heads, num_hashes * seq_len // chunk_length, chunk_length, chunk_length * (1 + num_chunks_before + num_chunks_after)]`
attention_probs = torch.exp(query_key_dots - logits)
# free memory
del query_key_dots
# dropout
attention_probs = nn.functional.dropout(attention_probs, p=self.dropout, training=self.training)
# Mask heads if we want to
if head_mask is not None:
attention_probs = attention_probs * head_mask
# attend values
out_vectors = torch.matmul(attention_probs, value_vectors)
# free memory
del value_vectors
# merge chunk length
if out_vectors.ndim > 4:
logits = logits.flatten(start_dim=2, end_dim=3).squeeze(-1)
out_vectors = out_vectors.flatten(start_dim=2, end_dim=3)
return out_vectors, logits, attention_probs
def _compute_attn_mask(
self, query_indices, key_indices, attention_mask, query_key_dot_shape, do_standard_self_attention
):
# attention mask for LSH
if attention_mask is not None:
# if chunked attention, the attention mask has to correspond to LSH order
attention_mask = attention_mask.to(torch.uint8)[:, None, :]
if not do_standard_self_attention:
# expand attn_mask to fit with key_value_bucket_idx shape
attention_mask = attention_mask[:, None, :]
attention_mask = attention_mask.expand(query_indices.shape[:-1] + (-1,))
# extract attention mask from LSH sorted key_indices
attention_mask = torch.gather(attention_mask, -1, key_indices)
attention_mask = attention_mask.unsqueeze(-2).expand(query_key_dot_shape)
# Causal mask
if self.is_decoder is True:
causal_mask = torch.ge(query_indices.unsqueeze(-1), key_indices.unsqueeze(-2)).to(query_indices.device)
# add attention mask if not None
if attention_mask is not None:
attention_mask = causal_mask * attention_mask
else:
attention_mask = causal_mask
return attention_mask
def _get_relevant_hid_states_and_buckets(
self, query_vectors, attention_mask, num_hashes, hidden_states, past_states, past_buckets
):
# concat hidden states
hidden_states = torch.cat([past_states, hidden_states], dim=1)
# batch_size hidden
batch_size = hidden_states.shape[0]
sequence_length = hidden_states.shape[1]
# check if cached buckets include pad bucket
max_bucket = self.num_buckets if isinstance(self.num_buckets, int) else reduce(mul, self.num_buckets)
# if pad bucket was cached => need to increase num buckets for caching
increase_num_buckets = past_buckets.max() > num_hashes * max_bucket - 1
# retrieve query buckets
query_buckets = self._hash_vectors(
query_vectors, num_hashes, attention_mask, increase_num_buckets=increase_num_buckets
)
# concat buckets
concat_buckets = torch.cat([past_buckets, query_buckets.unsqueeze(-1)], dim=-1)
# hash-based sort
bucket_idx = _stable_argsort(concat_buckets, dim=-1)
# bucket_idx has shape: BatchSize x NumAttnHeads x NumHashes x SequenceLength
assert bucket_idx.shape == (
batch_size,
self.num_attention_heads,
num_hashes,
sequence_length,
), f"bucket_idx should have shape {(batch_size, self.num_attention_heads, num_hashes, sequence_length)}, but has shape {bucket_idx.shape}."
# find indices of new bucket indices
relevant_bucket_idx = (bucket_idx == (bucket_idx.shape[-1] - 1)).nonzero()
# expand relevant bucket indices to its chunks
relevant_bucket_idx_chunk = self._expand_to_indices_in_relevant_chunk(relevant_bucket_idx, sequence_length)
relevant_bucket_idx_chunk = bucket_idx[tuple(relevant_bucket_idx_chunk.transpose(0, 1))]
# adapt bucket_idx for batch and hidden states for index select
bucket_idx_batch_offset = sequence_length * (
batch_size
* torch.arange(relevant_bucket_idx_chunk.shape[-1], device=hidden_states.device, dtype=torch.long)
// relevant_bucket_idx_chunk.shape[-1]
)
# add batch offset
relevant_bucket_idx_chunk_all_batch = relevant_bucket_idx_chunk + bucket_idx_batch_offset
hidden_states = hidden_states.reshape((-1, self.hidden_size))
# select all relevant hidden states
relevant_hidden_states = hidden_states.index_select(0, relevant_bucket_idx_chunk_all_batch)
# reshape hidden states and bucket_idx to correct output
relevant_hidden_states = relevant_hidden_states.reshape(
batch_size, self.num_attention_heads, -1, self.hidden_size
)
relevant_bucket_idx_chunk = relevant_bucket_idx_chunk.reshape(
batch_size, self.num_attention_heads, num_hashes, -1
)
assert (
relevant_hidden_states.shape[2]
== (self.num_chunks_before + self.num_chunks_after + 1) * self.chunk_length * num_hashes
), f"There should be {(self.num_chunks_before + self.num_chunks_after + 1) * self.chunk_length * num_hashes} `hidden_states`, there are {relevant_hidden_states.shape[2]} `hidden_states`."
assert (
relevant_bucket_idx_chunk.shape[-1]
== (self.num_chunks_before + self.num_chunks_after + 1) * self.chunk_length
), f"There should be {(self.num_chunks_before + self.num_chunks_after + 1) * self.chunk_length} `hidden_states`, there are {relevant_bucket_idx_chunk.shape[-1]} `bucket_idx`."
return relevant_hidden_states, relevant_bucket_idx_chunk, query_buckets
def _expand_to_indices_in_relevant_chunk(self, indices, sequence_length):
# get relevant indices of where chunk starts and its size
start_indices_chunk = ((indices[:, -1] // self.chunk_length) - self.num_chunks_before) * self.chunk_length
total_chunk_size = self.chunk_length * (1 + self.num_chunks_before + self.num_chunks_after)
# expand start indices and add correct chunk offset via arange
expanded_start_indices = start_indices_chunk.unsqueeze(-1).expand(indices.shape[0], total_chunk_size)
chunk_sequence_indices = expanded_start_indices + torch.arange(
total_chunk_size, device=indices.device, dtype=torch.long
).unsqueeze(0).expand(indices.shape[0], total_chunk_size)
# make sure that circular logic holds via % seq len
chunk_sequence_indices = chunk_sequence_indices.flatten() % sequence_length
# expand indices and set indices correctly
indices = indices.unsqueeze(1).expand((indices.shape[0], total_chunk_size, -1)).flatten(0, 1).clone()
indices[:, -1] = chunk_sequence_indices
return indices
def _len_and_dim_norm(self, vectors):
"""
length and attention head size dim normalization
"""
vectors = self._len_norm(vectors)
vectors = vectors * torch.rsqrt(
torch.tensor(self.attention_head_size, device=vectors.device, dtype=vectors.dtype)
)
return vectors
def _len_norm(self, x, epsilon=1e-6):
"""
length normalization
"""
variance = torch.mean(x ** 2, -1, keepdim=True)
norm_x = x * torch.rsqrt(variance + epsilon)
return norm_x
def _gather_by_expansion(self, vectors, idxs, num_hashes):
"""
expand dims of idxs and vectors for all hashes and gather
"""
expanded_idxs = idxs.unsqueeze(-1).expand(-1, -1, -1, self.attention_head_size)
vectors = vectors.repeat(1, 1, num_hashes, 1)
return torch.gather(vectors, 2, expanded_idxs)
class ReverseSort(Function):
"""
After chunked attention is applied which sorted clusters, original ordering has to be restored. Since customized
backward function is used for Reformer, the gradients of the output vectors have to be explicitly sorted here.
"""
@staticmethod
def forward(ctx, out_vectors, logits, sorted_bucket_idx, undo_sorted_bucket_idx):
# save sorted_bucket_idx for backprop
with torch.no_grad():
ctx.sorted_bucket_idx = sorted_bucket_idx
# undo sort to have correct order for next layer
expanded_undo_sort_indices = undo_sorted_bucket_idx.unsqueeze(-1).expand(out_vectors.shape)
out_vectors = torch.gather(out_vectors, 2, expanded_undo_sort_indices)
logits = torch.gather(logits, 2, undo_sorted_bucket_idx)
return out_vectors, logits
@staticmethod
def backward(ctx, grad_out_vectors, grad_logits):
# get parameters saved in ctx
sorted_bucket_idx = ctx.sorted_bucket_idx
expanded_sort_indices = sorted_bucket_idx.unsqueeze(-1).expand(grad_out_vectors.shape)
# reverse sort of forward
grad_out_vectors = torch.gather(grad_out_vectors, 2, expanded_sort_indices)
grad_logits = torch.gather(grad_logits, 2, sorted_bucket_idx)
# return grad and `None` fillers for last 2 forward args
return grad_out_vectors, grad_logits, None, None
class LocalSelfAttention(nn.Module, EfficientAttentionMixin):
def __init__(self, config):
super().__init__()
self.num_attention_heads = config.num_attention_heads
self.chunk_length = config.local_attn_chunk_length
self.num_chunks_before = config.local_num_chunks_before
self.num_chunks_after = config.local_num_chunks_after
self.is_decoder = config.is_decoder
self.pad_token_id = config.pad_token_id
self.attention_head_size = config.attention_head_size
self.all_head_size = self.num_attention_heads * self.attention_head_size
self.hidden_size = config.hidden_size
# projection matrices
self.query = nn.Linear(self.hidden_size, self.all_head_size, bias=False)
self.key = nn.Linear(self.hidden_size, self.all_head_size, bias=False)
self.value = nn.Linear(self.hidden_size, self.all_head_size, bias=False)
self.dropout = config.local_attention_probs_dropout_prob
# save mask value here
self.register_buffer("mask_value_float16", torch.tensor(-1e4))
self.register_buffer("mask_value_float32", torch.tensor(-1e9))
def forward(
self,
hidden_states,
attention_mask=None,
head_mask=None,
past_buckets_states=None,
use_cache=False,
output_attentions=False,
**kwargs
):
sequence_length = hidden_states.shape[1]
batch_size = hidden_states.shape[0]
# check if cache shall be used and that hidden states are already cached
if use_cache and past_buckets_states[1] is not None:
assert (
past_buckets_states[0] is None
), "LocalSelfAttention should not make use of `buckets`. There seems to be an error when caching hidden_states_and_buckets."
key_value_hidden_states = self._retrieve_relevant_hidden_states(
past_buckets_states[1], self.chunk_length, self.num_chunks_before
)
key_value_hidden_states = torch.cat([key_value_hidden_states, hidden_states], dim=1)
# only query vector for last token
query_vectors = self.query(hidden_states)
# compute key and value for relevant chunk
key_vectors = self.key(key_value_hidden_states)
value_vectors = self.value(key_value_hidden_states)
# free memory
del key_value_hidden_states
else:
# project hidden_states to query, key and value
query_vectors = self.query(hidden_states)
key_vectors = self.key(hidden_states)
value_vectors = self.value(hidden_states)
# split last dim into `config.num_attention_heads` and `config.attention_head_size`
query_vectors = self._split_hidden_size_dim(query_vectors, self.num_attention_heads, self.attention_head_size)
key_vectors = self._split_hidden_size_dim(key_vectors, self.num_attention_heads, self.attention_head_size)
value_vectors = self._split_hidden_size_dim(value_vectors, self.num_attention_heads, self.attention_head_size)
assert (
query_vectors.shape[-1] == self.attention_head_size
), "last dim of query_key_vectors is {} but should be {}.".format(
query_vectors.shape[-1], self.attention_head_size
)
assert (
key_vectors.shape[-1] == self.attention_head_size
), "last dim of query_key_vectors is {} but should be {}.".format(
key_vectors.shape[-1], self.attention_head_size
)
assert (
value_vectors.shape[-1] == self.attention_head_size
), "last dim of query_key_vectors is {} but should be {}.".format(
value_vectors.shape[-1], self.attention_head_size
)
if self.chunk_length is None:
assert (
self.num_chunks_before == 0 and self.num_chunks_after == 0
), "If `config.chunk_length` is `None`, make sure `config.num_chunks_after` and `config.num_chunks_before` are set to 0."
# normalize key vectors
key_vectors = key_vectors / torch.sqrt(
torch.tensor(self.attention_head_size, device=key_vectors.device, dtype=key_vectors.dtype)
)
# get sequence length indices
indices = torch.arange(sequence_length, device=query_vectors.device).repeat(
batch_size, self.num_attention_heads, 1
)
# if one should do normal n^2 self-attention
do_standard_self_attention = sequence_length <= self.chunk_length
# if input should be chunked
if not do_standard_self_attention:
# chunk vectors
# B x Num_Attn_Head x Seq_Len // chunk_len x chunk_len x attn_head_size
query_vectors = self._split_seq_length_dim_to(
query_vectors,
-1,
self.chunk_length,
self.num_attention_heads,
self.attention_head_size,
)
key_vectors = self._split_seq_length_dim_to(
key_vectors,
-1,
self.chunk_length,
self.num_attention_heads,
self.attention_head_size,
)
value_vectors = self._split_seq_length_dim_to(
value_vectors,
-1,
self.chunk_length,
self.num_attention_heads,
self.attention_head_size,
)
# chunk indices
query_indices = self._split_seq_length_dim_to(indices, -1, self.chunk_length, self.num_attention_heads)
key_indices = self._split_seq_length_dim_to(indices, -1, self.chunk_length, self.num_attention_heads)
# append chunks before and after
key_vectors = self._look_adjacent(key_vectors, self.num_chunks_before, self.num_chunks_after)
value_vectors = self._look_adjacent(value_vectors, self.num_chunks_before, self.num_chunks_after)
key_indices = self._look_adjacent(key_indices, self.num_chunks_before, self.num_chunks_after)
else:
query_indices = key_indices = indices
# query-key matmul: QK^T
query_key_dots = torch.matmul(query_vectors, key_vectors.transpose(-1, -2))
# free memory
del query_vectors, key_vectors
mask = self._compute_attn_mask(
query_indices, key_indices, attention_mask, query_key_dots.shape, do_standard_self_attention
)
if mask is not None:
# get mask tensor depending on half precision or not
if query_key_dots.dtype == torch.float16:
mask_value = self.mask_value_float16.half()
else:
mask_value = self.mask_value_float32
query_key_dots = torch.where(mask, query_key_dots, mask_value)
# free memory
del mask
# softmax
logits = torch.logsumexp(query_key_dots, dim=-1, keepdim=True)
attention_probs = torch.exp(query_key_dots - logits)
# free memory
del logits
# dropout
attention_probs = nn.functional.dropout(attention_probs, p=self.dropout, training=self.training)
# Mask heads if we want to
if head_mask is not None:
attention_probs = attention_probs * head_mask
# attend values
out_vectors = torch.matmul(attention_probs, value_vectors)
# free memory
del value_vectors
# merge chunk length
if not do_standard_self_attention:
out_vectors = out_vectors.flatten(start_dim=2, end_dim=3)
assert out_vectors.shape == (
batch_size,
self.num_attention_heads,
sequence_length,
self.attention_head_size,
)
out_vectors = self._merge_hidden_size_dims(out_vectors, self.num_attention_heads, self.attention_head_size)
if output_attentions is False:
attention_probs = ()
return LocalSelfAttentionOutput(hidden_states=out_vectors, attention_probs=attention_probs)
def _compute_attn_mask(
self, query_indices, key_indices, attention_mask, query_key_dots_shape, do_standard_self_attention
):
# chunk attention mask and look before and after
if attention_mask is not None:
attention_mask = attention_mask.to(torch.uint8)[:, None, :]
if not do_standard_self_attention:
attention_mask = self._split_seq_length_dim_to(attention_mask, -1, self.chunk_length, 1)
attention_mask = self._look_adjacent(attention_mask, self.num_chunks_before, self.num_chunks_after)
# create attn_mask
attention_mask = attention_mask.unsqueeze(-2).expand(query_key_dots_shape)
# Causal mask
if self.is_decoder is True:
causal_mask = torch.ge(query_indices.unsqueeze(-1), key_indices.unsqueeze(-2)).to(query_indices.device)
# add attention mask if not None
if attention_mask is not None:
attention_mask = causal_mask * attention_mask
else:
attention_mask = causal_mask
return attention_mask
@staticmethod
def _retrieve_relevant_hidden_states(previous_hidden_states, chunk_length, num_chunks_before):
start_position = ((previous_hidden_states.shape[1] // chunk_length) - num_chunks_before) * chunk_length
return previous_hidden_states[:, start_position:]
class ReformerSelfOutput(nn.Module):
def __init__(self, config):
super().__init__()
all_head_size = config.num_attention_heads * config.attention_head_size
self.dropout = config.hidden_dropout_prob
self.dense = nn.Linear(all_head_size, config.hidden_size, bias=False)
def forward(self, hidden_states):
hidden_states = self.dense(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
return hidden_states
class ReformerAttention(nn.Module):
def __init__(self, config, layer_id=0):
super().__init__()
self.layer_id = layer_id
self.attn_layers = config.attn_layers
self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
if len(set(self.attn_layers)) == 1 and self.attn_layers[0] == "lsh":
self.self_attention = LSHSelfAttention(config)
elif len(set(self.attn_layers)) == 1 and self.attn_layers[0] == "local":
self.self_attention = LocalSelfAttention(config)
elif len(set(self.attn_layers)) == 2 and set(self.attn_layers) == set(["lsh", "local"]):
# get correct attn layers
if self.attn_layers[self.layer_id] == "lsh":
self.self_attention = LSHSelfAttention(config)
else:
self.self_attention = LocalSelfAttention(config)
else:
raise NotImplementedError(
"Only attn layer types 'lsh' and 'local' exist, but got `config.attn_layers`: {}. Select attn layer types from ['lsh', 'local'] only.".format(
self.attn_layers
)
)
self.output = ReformerSelfOutput(config)
def forward(
self,
hidden_states,
attention_mask=None,
head_mask=None,
num_hashes=None,
past_buckets_states=None,
use_cache=False,
orig_sequence_length=None,
output_attentions=False,
buckets=None,
):
hidden_states = self.layer_norm(hidden_states)
# make sure cached hidden states is set to None for backward pass
if past_buckets_states is not None:
past_buckets_states_layer = past_buckets_states[self.layer_id]
else:
past_buckets_states_layer = None
# use cached buckets for backprob if buckets not None for LSHSelfAttention
self_attention_outputs = self.self_attention(
hidden_states=hidden_states,
head_mask=head_mask,
attention_mask=attention_mask,
num_hashes=num_hashes,
past_buckets_states=past_buckets_states_layer,
use_cache=use_cache,
output_attentions=output_attentions,
buckets=buckets,
)
# add buckets if necessary
if hasattr(self_attention_outputs, "buckets"):
buckets = self_attention_outputs.buckets
else:
buckets = None
# cache hidden states for future use
if use_cache:
if past_buckets_states[self.layer_id][0] is None:
# padded input should not be cached
past_buckets = (
buckets[:, :, :, :orig_sequence_length]
if (buckets is not None and orig_sequence_length > 1)
else buckets
)
else:
past_buckets = torch.cat([past_buckets_states[self.layer_id][0], buckets], dim=-1)
if past_buckets_states[self.layer_id][1] is None:
# padded input should not be cached
past_states = hidden_states[:, :orig_sequence_length]
else:
past_states = torch.cat([past_buckets_states[self.layer_id][1], hidden_states], dim=1)
past_buckets_states[self.layer_id] = (past_buckets, past_states)
# compute attention feed forward output
attention_output = self.output(self_attention_outputs.hidden_states)
return AttentionOutput(
hidden_states=attention_output,
attention_probs=self_attention_outputs.attention_probs,
buckets=buckets,
)
class ReformerFeedForwardDense(nn.Module):
def __init__(self, config):
super().__init__()
self.dropout = config.hidden_dropout_prob
if isinstance(config.hidden_act, str):
self.act_fn = ACT2FN[config.hidden_act]
else:
self.act_fn = config.hidden_act
self.dense = nn.Linear(config.hidden_size, config.feed_forward_size)
def forward(self, hidden_states):
hidden_states = self.dense(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = self.act_fn(hidden_states)
return hidden_states
class ReformerFeedForwardOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dropout = config.hidden_dropout_prob
self.dense = nn.Linear(config.feed_forward_size, config.hidden_size)
def forward(self, hidden_states):
hidden_states = self.dense(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
return hidden_states
class ChunkReformerFeedForward(nn.Module):
def __init__(self, config):
super().__init__()
self.chunk_size_feed_forward = config.chunk_size_feed_forward
self.seq_len_dim = 1
self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
self.dense = ReformerFeedForwardDense(config)
self.output = ReformerFeedForwardOutput(config)
def forward(self, attention_output):
return apply_chunking_to_forward(
self.forward_chunk,
self.chunk_size_feed_forward,
self.seq_len_dim,
attention_output,
)
def forward_chunk(self, hidden_states):
hidden_states = self.layer_norm(hidden_states)
hidden_states = self.dense(hidden_states)
return self.output(hidden_states)
class ReformerLayer(nn.Module):
def __init__(self, config, layer_id=0):
super().__init__()
self.attention = ReformerAttention(config, layer_id)
# dropout requires to have the same
# seed for forward and backward pass
self.attention_seed = None
self.feed_forward_seed = None
self.feed_forward = ChunkReformerFeedForward(config)
def _init_attention_seed(self):
"""
This function sets a new seed for the attention layer to make dropout deterministic for both forward calls: 1
normal forward call and 1 forward call in backward to recalculate activations.
"""
# randomize seeds
# use cuda generator if available
if hasattr(torch.cuda, "default_generators") and len(torch.cuda.default_generators) > 0:
# GPU
device_idx = torch.cuda.current_device()
self.attention_seed = torch.cuda.default_generators[device_idx].seed()
else:
# CPU
self.attention_seed = int(torch.seed() % sys.maxsize)
torch.manual_seed(self.attention_seed)
def _init_feed_forward_seed(self):
"""
This function sets a new seed for the feed forward layer to make dropout deterministic for both forward calls:
1 normal forward call and 1 forward call in backward to recalculate activations.
"""
# randomize seeds
# use cuda generator if available
if hasattr(torch.cuda, "default_generators") and len(torch.cuda.default_generators) > 0:
# GPU
device_idx = torch.cuda.current_device()
self.feed_forward_seed = torch.cuda.default_generators[device_idx].seed()
else:
# CPU
self.feed_forward_seed = int(torch.seed() % sys.maxsize)
torch.manual_seed(self.feed_forward_seed)
def forward(
self,
prev_attn_output,
hidden_states,
attention_mask=None,
head_mask=None,
num_hashes=None,
past_buckets_states=None,
use_cache=False,
orig_sequence_length=None,
output_attentions=False,
):
with torch.no_grad():
# every forward pass we sample a different seed
# for dropout and save for forward fn in backward pass
# to have correct dropout
if self.training:
self._init_attention_seed()
attn_outputs = self.attention(
hidden_states=hidden_states,
head_mask=head_mask,
attention_mask=attention_mask,
num_hashes=num_hashes,
past_buckets_states=past_buckets_states,
use_cache=use_cache,
orig_sequence_length=orig_sequence_length,
output_attentions=output_attentions,
)
attn_output = attn_outputs.hidden_states
# Implementation of RevNet (see Fig. 6 in https://towardsdatascience.com/illustrating-the-reformer-393575ac6ba0)
# Y_1 = X_1 + f(X_2)
attn_output = prev_attn_output + attn_output
# free memory
del prev_attn_output
# every forward pass we sample a different seed
# for dropout and save seed for forward fn in backward
# to have correct dropout
if self.training:
self._init_feed_forward_seed()
# Y_2 = X_2 + g(Y_1)
hidden_states = hidden_states + self.feed_forward(attn_output)
return ReformerOutput(
attn_output=attn_output,
hidden_states=hidden_states,
attention_probs=attn_outputs.attention_probs,
buckets=attn_outputs.buckets,
)
def backward_pass(
self,
next_attn_output,
hidden_states,
grad_attn_output,
grad_hidden_states,
attention_mask=None,
head_mask=None,
buckets=None,
):
# Implements the backward pass for reversible ResNets.
# A good blog post on how this works can be found here:
# Implementation of RevNet (see Fig. 6 in https://towardsdatascience.com/illustrating-the-reformer-393575ac6ba0)
# This code is heavily inspired by https://github.com/lucidrains/reformer-pytorch/blob/master/reformer_pytorch/reversible.py
with torch.enable_grad():
next_attn_output.requires_grad = True
# set seed to have correct dropout
torch.manual_seed(self.feed_forward_seed)
# g(Y_1)
res_hidden_states = self.feed_forward(next_attn_output)
res_hidden_states.backward(grad_hidden_states, retain_graph=True)
with torch.no_grad():
# X_2 = Y_2 - g(Y_1)
hidden_states = hidden_states - res_hidden_states
del res_hidden_states
grad_attn_output = grad_attn_output + next_attn_output.grad
next_attn_output.grad = None
with torch.enable_grad():
hidden_states.requires_grad = True
# set seed to have correct dropout
torch.manual_seed(self.attention_seed)
# f(X_2)
# use cached buckets for backprob if buckets not None for LSHSelfAttention
output = self.attention(
hidden_states=hidden_states,
head_mask=head_mask,
attention_mask=attention_mask,
buckets=buckets,
).hidden_states
output.backward(grad_attn_output, retain_graph=True)
with torch.no_grad():
# X_1 = Y_1 - f(X_2)
attn_output = next_attn_output - output
del output, next_attn_output
grad_hidden_states = grad_hidden_states + hidden_states.grad
hidden_states.grad = None
hidden_states = hidden_states.detach()
return ReformerBackwardOutput(
attn_output=attn_output,
hidden_states=hidden_states,
grad_attn_output=grad_attn_output,
grad_hidden_states=grad_hidden_states,
)
class _ReversibleFunction(Function):
"""
To prevent PyTorch from performing the usual backpropagation, a customized backward function is implemented here.
This way it is made sure that no memory expensive activations are saved during the forward pass. This function is
heavily inspired by https://github.com/lucidrains/reformer-pytorch/blob/master/reformer_pytorch/reversible.py
"""
@staticmethod
def forward(
ctx,
hidden_states,
layers,
attention_mask,
head_mask,
num_hashes,
all_hidden_states,
all_attentions,
past_buckets_states,
use_cache,
orig_sequence_length,
output_hidden_states,
output_attentions,
):
all_buckets = ()
# split duplicated tensor
hidden_states, attn_output = torch.chunk(hidden_states, 2, dim=-1)
for layer_id, (layer, layer_head_mask) in enumerate(zip(layers, head_mask)):
if output_hidden_states is True:
all_hidden_states.append(hidden_states)
layer_outputs = layer(
prev_attn_output=attn_output,
hidden_states=hidden_states,
attention_mask=attention_mask,
head_mask=layer_head_mask,
num_hashes=num_hashes,
past_buckets_states=past_buckets_states,
use_cache=use_cache,
orig_sequence_length=orig_sequence_length,
output_attentions=output_attentions,
)
attn_output = layer_outputs.attn_output
hidden_states = layer_outputs.hidden_states
all_buckets = all_buckets + (layer_outputs.buckets,)
if output_attentions:
all_attentions.append(layer_outputs.attention_probs)
# Add last layer
if output_hidden_states is True:
all_hidden_states.append(hidden_states)
# attach params to ctx for backward
ctx.save_for_backward(attn_output.detach(), hidden_states.detach())
ctx.layers = layers
ctx.all_buckets = all_buckets
ctx.head_mask = head_mask
ctx.attention_mask = attention_mask
# Concatenate 2 RevNet outputs
return torch.cat([attn_output, hidden_states], dim=-1)
@staticmethod
def backward(ctx, grad_hidden_states):
grad_attn_output, grad_hidden_states = torch.chunk(grad_hidden_states, 2, dim=-1)
# retrieve params from ctx for backward
attn_output, hidden_states = ctx.saved_tensors
# create tuple
output = ReformerBackwardOutput(
attn_output=attn_output,
hidden_states=hidden_states,
grad_attn_output=grad_attn_output,
grad_hidden_states=grad_hidden_states,
)
# free memory
del grad_attn_output, grad_hidden_states, attn_output, hidden_states
layers = ctx.layers
all_buckets = ctx.all_buckets
head_mask = ctx.head_mask
attention_mask = ctx.attention_mask
for idx, layer in enumerate(layers[::-1]):
# pop last buckets from stack
buckets = all_buckets[-1]
all_buckets = all_buckets[:-1]
# backprop
output = layer.backward_pass(
next_attn_output=output.attn_output,
hidden_states=output.hidden_states,
grad_attn_output=output.grad_attn_output,
grad_hidden_states=output.grad_hidden_states,
head_mask=head_mask[len(layers) - idx - 1],
attention_mask=attention_mask,
buckets=buckets,
)
assert all_buckets == (), "buckets have to be empty after backpropagation"
grad_hidden_states = torch.cat([output.grad_attn_output, output.grad_hidden_states], dim=-1)
# num of return vars has to match num of forward() args
# return gradient for hidden_states arg and None for other args
return grad_hidden_states, None, None, None, None, None, None, None, None, None, None, None
class ReformerEncoder(nn.Module):
def __init__(self, config):
super().__init__()
self.dropout = config.hidden_dropout_prob
self.layers = nn.ModuleList([ReformerLayer(config, i) for i in range(config.num_hidden_layers)])
# Reformer is using Rev Nets, thus last layer outputs are concatenated and
# Layer Norm is done over 2 * hidden_size
self.layer_norm = nn.LayerNorm(2 * config.hidden_size, eps=config.layer_norm_eps)
def forward(
self,
hidden_states,
attention_mask=None,
head_mask=None,
num_hashes=None,
past_buckets_states=None,
use_cache=False,
orig_sequence_length=None,
output_hidden_states=False,
output_attentions=False,
):
# hidden_states and attention lists to be filled if wished
all_hidden_states = []
all_attentions = []
# init cached hidden states if necessary
if past_buckets_states is None:
past_buckets_states = [((None), (None)) for i in range(len(self.layers))]
# concat same tensor for reversible ResNet
hidden_states = torch.cat([hidden_states, hidden_states], dim=-1)
hidden_states = _ReversibleFunction.apply(
hidden_states,
self.layers,
attention_mask,
head_mask,
num_hashes,
all_hidden_states,
all_attentions,
past_buckets_states,
use_cache,
orig_sequence_length,
output_hidden_states,
output_attentions,
)
# Apply layer norm to concatenated hidden states
hidden_states = self.layer_norm(hidden_states)
# Apply dropout
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
return ReformerEncoderOutput(
hidden_states=hidden_states,
all_hidden_states=all_hidden_states,
all_attentions=all_attentions,
past_buckets_states=past_buckets_states,
)
class ReformerOnlyLMHead(nn.Module):
def __init__(self, config):
super().__init__()
# Reformer is using Rev Nets, thus last layer outputs are concatenated and
# Layer Norm is done over 2 * hidden_size
self.seq_len_dim = 1
self.chunk_size_lm_head = config.chunk_size_lm_head
self.decoder = nn.Linear(2 * config.hidden_size, config.vocab_size, bias=False)
self.bias = nn.Parameter(torch.zeros(config.vocab_size))
# Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
self.decoder.bias = self.bias
def forward(self, hidden_states):
return apply_chunking_to_forward(self.forward_chunk, self.chunk_size_lm_head, self.seq_len_dim, hidden_states)
def forward_chunk(self, hidden_states):
hidden_states = self.decoder(hidden_states)
return hidden_states
class ReformerPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = ReformerConfig
base_model_prefix = "reformer"
@property
def dummy_inputs(self):
input_ids = torch.tensor(DUMMY_INPUTS)
input_mask = torch.tensor(DUMMY_MASK)
dummy_inputs = {
"input_ids": input_ids,
"attention_mask": input_mask,
}
return dummy_inputs
def _init_weights(self, module):
""" Initialize the weights """
if isinstance(module, AxialPositionEmbeddings):
for weight in module.weights:
torch.nn.init.normal_(weight, std=self.config.axial_norm_std)
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
elif isinstance(module, nn.Linear):
# Slightly different from the TF version which uses truncated_normal for initialization
# cf https://github.com/pytorch/pytorch/pull/5617
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
@dataclass
class ReformerModelOutput(ModelOutput):
"""
Output type of :class:`~transformers.ReformerModel`.
Args:
last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_predict, hidden_size)`):
Sequence of hidden-states at the last layer of the model.
``num_predict`` corresponds to ``target_mapping.shape[1]``. If ``target_mapping`` is ``None``, then
``num_predict`` corresponds to ``sequence_length``.
past_buckets_states (:obj:`List[Tuple(torch.LongTensor, torch.FloatTensor)]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
List of :obj:`Tuple(torch.LongTensor, torch.FloatTensor` of length :obj:`config.n_layers`, with the first
element being the previous `buckets` of shape :obj:`(batch_size, num_heads, num_hashes, sequence_length)`)
and the second being the previous `hidden_states` of shape :obj:`(batch_size, sequence_length,
hidden_size)`).
Contains precomputed buckets and hidden-states that can be used (see ``past_buckets_states`` input) to
speed up sequential decoding.
hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings and one for the output of each
layer) of shape :obj:`(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
sequence_length, sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
last_hidden_state: torch.FloatTensor
past_buckets_states: Optional[List[Tuple[torch.LongTensor, torch.FloatTensor]]] = None
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
attentions: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class ReformerModelWithLMHeadOutput(ModelOutput):
"""
Output type of :class:`~transformers.ReformerModelWithLMHead`.
Args:
loss (:obj:`torch.FloatTensor` of shape `(1,)`, `optional`, returned when ``labels`` is provided)
Language modeling loss (for next-token prediction).
logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_predict, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
``num_predict`` corresponds to ``target_mapping.shape[1]``. If ``target_mapping`` is ``None``, then
``num_predict`` corresponds to ``sequence_length``.
past_buckets_states (:obj:`List[Tuple(torch.LongTensor, torch.FloatTensor)]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
List of :obj:`Tuple(torch.LongTensor, torch.FloatTensor` of length :obj:`config.n_layers`, with the first
element being the previous `buckets` of shape :obj:`(batch_size, num_heads, num_hashes, sequence_length)`)
and the second being the previous `hidden_states` of shape :obj:`(batch_size, sequence_length,
hidden_size)`).
Contains precomputed buckets and hidden-states that can be used (see ``past_buckets_states`` input) to
speed up sequential decoding.
hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
TTuple of :obj:`torch.FloatTensor` (one for the output of the embeddings and one for the output of each
layer) of shape :obj:`(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
sequence_length, sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
loss: Optional[torch.FloatTensor] = None
logits: torch.FloatTensor = None
past_buckets_states: Optional[List[Tuple[torch.LongTensor, torch.FloatTensor]]] = None
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
attentions: Optional[Tuple[torch.FloatTensor]] = None
REFORMER_START_DOCSTRING = r"""
Reformer was proposed in `Reformer: The Efficient Transformer <https://arxiv.org/abs/2001.04451>`__ by Nikita
Kitaev, Łukasz Kaiser, Anselm Levskaya.
This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
pruning heads etc.)
This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__
subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
general usage and behavior.
Parameters:
config (:class:`~transformers.ReformerConfig`): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
weights.
"""
REFORMER_INPUTS_DOCSTRING = r"""
Args:
input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. During training the input_ids sequence_length has to be
a multiple of the relevant model's chunk lengths (lsh's, local's or both). During evaluation, the indices
are automatically padded to be a multiple of the chunk length.
Indices can be obtained using :class:`~transformers.ReformerTokenizer`. See
:meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
details.
`What are input IDs? <../glossary.html#input-ids>`__
attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
`What are attention masks? <../glossary.html#attention-mask>`__
position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
config.max_position_embeddings - 1]``.
`What are position IDs? <../glossary.html#position-ids>`__
head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
vectors than the model's internal embedding lookup matrix.
num_hashes (:obj:`int`, `optional`):
The number of hashing rounds that should be performed during bucketing. Setting this argument overwrites
the default defined in :obj:`config.num_hashes`.
For more information, see :obj:`num_hashes` in :class:`~transformers.ReformerConfig`.
past_buckets_states (:obj:`List[Tuple(torch.LongTensor, torch.FloatTensor)]`, `optional`):
List of :obj:`Tuple(torch.LongTensor, torch.FloatTensor` of length :obj:`config.n_layers`, with the first
element being the previous `buckets` of shape :obj:`(batch_size, num_heads, num_hashes, sequence_length)`)
and the second being the previous `hidden_states` of shape :obj:`(batch_size, sequence_length,
hidden_size)`).
Contains precomputed hidden-states and buckets (only relevant for LSH Self-Attention). Can be used to speed
up sequential decoding.
use_cache (:obj:`bool`, `optional`):
If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
decoding (see :obj:`past_key_values`).
output_attentions (:obj:`bool`, `optional`):
Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
tensors for more detail.
output_hidden_states (:obj:`bool`, `optional`):
Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
more detail.
return_dict (:obj:`bool`, `optional`):
Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
"""
class ReformerClassificationHead(nn.Module):
"""Head for sentence-level classification tasks."""
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(2 * config.hidden_size, config.hidden_size)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
def forward(self, hidden_states, **kwargs):
hidden_states = hidden_states[:, 0, :] # take <s> token (equiv. to [CLS])
hidden_states = self.dropout(hidden_states)
hidden_states = self.dense(hidden_states)
hidden_states = torch.tanh(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.out_proj(hidden_states)
return hidden_states