Source code for mindspore.nn.transformer.transformer
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""
Note:
Transformer Networks. This is interface that is subject to change or deletion.
"""
import math
import numpy as np
from mindspore.common.tensor import Tensor
from mindspore.common.parameter import Parameter
from mindspore.common.initializer import initializer
from mindspore import nn
from mindspore import context
import mindspore.common.dtype as mstype
from mindspore.ops import operations as P
from mindspore.ops import functional as F
from mindspore.nn.cell import Cell
from mindspore._checkparam import Validator
from mindspore import log as logger
from mindspore.parallel._utils import _get_parallel_mode, _is_sharding_propagation
from mindspore.context import ParallelMode
from .layers import _LayerNorm, _Linear, _Dropout, _check_input_shape, \
_args_type_validator_check, _valid_type_checks, _valid_value_checks, \
_check_shape_equal, _check_past_none_input_none, _check_input_dtype, _check_input_shape_value
from .op_parallel_config import default_dpmp_config, _PipeLineConfig, OpParallelConfig, _Config, _check_config, \
MoEParallelConfig
from .moe import default_moe_config, MoE, _check_moe_config
__all__ = [
"AttentionMask",
"VocabEmbedding",
"MultiHeadAttention",
"FeedForward",
"TransformerEncoder",
"TransformerDecoder",
"TransformerEncoderLayer",
"TransformerDecoderLayer",
"Transformer",
"TransformerOpParallelConfig",
"EmbeddingOpParallelConfig",
"TransformerRecomputeConfig"]
[文档]class EmbeddingOpParallelConfig(_Config):
r"""
EmbeddingOpParallelConfig for the setting data parallel or row slice for the embedding table.
Args:
data_parallel(int): The data parallel way. The input data will be sliced into n parts for embedding layer
according to this value. Default 1.
model_parallel(int): The model parallel way. The embedding table parameters
will be sliced according to the model parallel way. Default: 1.
vocab_emb_dp(bool): Shard embedding in model parallel or data parallel. If true, the embedding lookup
will be a data parallel style training and model_parallel value will be ignored. If false, the
embedding table will be sharded into n parts at the 0-th dimension row slice of the embedding table,
where the n is the model parallel way determined by this parameter. Default: True
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> from mindspore.nn.transformer import EmbeddingOpParallelConfig
>>> config=EmbeddingOpParallelConfig(data_parallel=1, model_parallel=1, vocab_emb_dp=True)
"""
def __init__(self, data_parallel=1, model_parallel=1, vocab_emb_dp=True):
self._dp_mp_config = OpParallelConfig(data_parallel=data_parallel, model_parallel=model_parallel)
Validator.check_bool(vocab_emb_dp, "vocab_emb_dp")
self.vocab_emb_dp = vocab_emb_dp
@property
def data_parallel(self):
return self._dp_mp_config.data_parallel
@data_parallel.setter
def data_parallel(self, value):
self._dp_mp_config.data_parallel = value
@property
def model_parallel(self):
return self._dp_mp_config.model_parallel
@model_parallel.setter
def model_parallel(self, value):
self._dp_mp_config.model_parallel = value
@property
def vocab_emb_dp(self):
return self._vocab_emb_dp
@vocab_emb_dp.setter
def vocab_emb_dp(self, value):
Validator.check_bool(value, "vocab_emb_dp")
self._vocab_emb_dp = value
@property
def dp_mp_config(self):
return self._dp_mp_config
[文档]class TransformerRecomputeConfig(_Config):
r"""
TransformerRecomputeConfig for the setting recompute attributes for encoder/decoder layers.
Args:
recompute (bool): Enable recomputation of the transformer block or not. Default: False.
parallel_optimizer_comm_recompute (bool): Specifies whether the communication operator allgathers
introduced by optimizer shard are recomputed in auto parallel or semi auto parallel mode.
Default: False.
mp_comm_recompute (bool): Specifies whether the model parallel communication operators
in the cell are recomputed in auto parallel or semi auto parallel mode. Default: True.
recompute_slice_activation (bool): Slice the cell output which would remains in memory. Default: False.
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> from mindspore.nn.transformer import TransformerRecomputeConfig
>>> config=TransformerRecomputeConfig(recompute=True, parallel_optimizer_comm_recompute=True, \
... mp_comm_recompute=True, recompute_slice_activation=True)
"""
def __init__(self, recompute=False, parallel_optimizer_comm_recompute=False,
mp_comm_recompute=True, recompute_slice_activation=False):
Validator.check_bool(recompute, "recompute")
Validator.check_bool(parallel_optimizer_comm_recompute, "parallel_optimizer_comm_recompute")
Validator.check_bool(mp_comm_recompute, "mp_comm_recompute")
Validator.check_bool(recompute_slice_activation, "recompute_slice_activation")
self._recompute = recompute
self._parallel_optimizer_comm_recompute = parallel_optimizer_comm_recompute
self._mp_comm_recompute = mp_comm_recompute
self._recompute_slice_activation = recompute_slice_activation
@property
def recompute(self):
return self._recompute
@recompute.setter
def recompute(self, value):
Validator.check_bool(value, "recompute")
self._recompute = value
@property
def parallel_optimizer_comm_recompute(self):
return self._parallel_optimizer_comm_recompute
@parallel_optimizer_comm_recompute.setter
def parallel_optimizer_comm_recompute(self, value):
Validator.check_bool(value, "parallel_optimizer_comm_recompute")
self._parallel_optimizer_comm_recompute = value
@property
def mp_comm_recompute(self):
return self._mp_comm_recompute
@mp_comm_recompute.setter
def mp_comm_recompute(self, value):
Validator.check_bool(value, "mp_comm_recompute")
self._mp_comm_recompute = value
@property
def recompute_slice_activation(self):
return self._recompute_slice_activation
@recompute_slice_activation.setter
def recompute_slice_activation(self, value):
Validator.check_bool(value, "recompute_slice_activation")
self._recompute_slice_activation = value
default_transformer_recompute_config = TransformerRecomputeConfig()
[文档]class TransformerOpParallelConfig(_Config):
r"""
TransformerOpParallelConfig for the setting global data parallel, model parallel and fusion group.
The parallel configure setting.
Note:
Except the recompute argument, other arguments will **not** be effective when the user doesn't set
auto_parallel_context to `SEMI_AUTO_PARALLEL` or `AUTO_PARALLEL`.
The micro_batch_num must be greater than or equal to pipeline_stage when training.
The data_parallel\*model_parallel \*pipeline_stage must be equal or less equal to the device. When setting
the pipeline stage and optimizer_shard, the config will overwrite the auto_parallel_context. When given the
8 devices and the data_parallel is 1 and model_parallel is 1, the calculation will be repeated on each
device.
Args:
data_parallel (int): The data parallel way. The input data will be sliced into n parts for each layer
according to the data parallel way. Default: 1.
model_parallel (int): The model parallel way. The parameters of dense layers in MultiheadAttention and
FeedForward layer will be sliced according to the model parallel way. Default: 1.
expert_parallel (int): The expert parallel way. This is effective only when MoE (Mixture of Experts)
is applied. This value specifies the number of partitions to split the experts into.
pipeline_stage (int): The number of the pipeline stage. Should be a positive value. Default: 1.
micro_batch_num (int): The micro size of the batches for the pipeline training. Default: 1.
optimizer_shard (bool): Whether to enable optimizer shard. Default False.
gradient_aggregation_group (int): The fusion group size of the optimizer state sharding. Default: 4.
recompute (Union[TransformerRecomputeConfig, bool]): The configuration of recomputation for
the transformer block. Default: The default configuration of TransformerRecomputeConfig.
vocab_emb_dp (bool): Shard embedding in model parallel or data parallel. Default: True.
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> from mindspore.nn.transformer import TransformerRecomputeConfig
>>> recompute_config=TransformerRecomputeConfig(recompute=True, parallel_optimizer_comm_recompute=True, \
... mp_comm_recompute=True, recompute_slice_activation=True)
>>> config=TransformerOpParallelConfig(data_parallel=1, model_parallel=1, recompute=recompute_config)
"""
def __init__(self, data_parallel=1, model_parallel=1, expert_parallel=1, pipeline_stage=1, micro_batch_num=1,
recompute=default_transformer_recompute_config,
optimizer_shard=False, gradient_aggregation_group=4, vocab_emb_dp=True):
self.recompute = recompute
self.optimizer_shard = optimizer_shard
self.gradient_aggregation_group = gradient_aggregation_group
self._embed_dp_mp_config = EmbeddingOpParallelConfig(data_parallel=data_parallel, model_parallel=model_parallel,
vocab_emb_dp=vocab_emb_dp)
self._pp_config = _PipeLineConfig(pipeline_stage=pipeline_stage, micro_batch_num=micro_batch_num)
self._moe_config = MoEParallelConfig(data_parallel=data_parallel, model_parallel=model_parallel,
expert_parallel=expert_parallel)
@property
def recompute(self):
return self._recompute
@recompute.setter
def recompute(self, value):
if not isinstance(value, TransformerRecomputeConfig) and not isinstance(value, bool):
raise TypeError(f"recompute should be a TransformerRecomputeConfig/bool, but got {type(value).__name__}.")
if isinstance(value, bool):
logger.warning(f"TransformerRecomputeConfig is recommended as the recompute configuration type.")
self._recompute = value
@property
def vocab_emb_dp(self):
return self._embed_dp_mp_config.vocab_emb_dp
@vocab_emb_dp.setter
def vocab_emb_dp(self, value):
self._embed_dp_mp_config.vocab_emb_dp = value
@property
def gradient_aggregation_group(self):
return self._gradient_aggregation_group
@gradient_aggregation_group.setter
def gradient_aggregation_group(self, value):
Validator.check_positive_int(value, "gradient_aggregation_group")
self._gradient_aggregation_group = value
@property
def micro_batch_num(self):
return self._pp_config.micro_batch_num
@micro_batch_num.setter
def micro_batch_num(self, value):
self._pp_config.micro_batch_num = value
@property
def model_parallel(self):
return self._embed_dp_mp_config.model_parallel
@model_parallel.setter
def model_parallel(self, value):
self._embed_dp_mp_config.model_parallel = value
self._moe_config.model_parallel = value
@property
def data_parallel(self):
return self._embed_dp_mp_config.data_parallel
@data_parallel.setter
def data_parallel(self, value):
self._embed_dp_mp_config.data_parallel = value
self._moe_config.data_parallel = value
@property
def expert_parallel(self):
return self._moe_config.expert_parallel
@expert_parallel.setter
def expert_parallel(self, value):
self._moe_config.expert_parallel = value
@property
def pipeline_stage(self):
return self._pp_config.pipeline_stage
@pipeline_stage.setter
def pipeline_stage(self, value):
self._pp_config.pipeline_stage = value
@property
def optimizer_shard(self):
return self._optimizer_shard
@optimizer_shard.setter
def optimizer_shard(self, value):
Validator.check_bool(value, "optimizer_shard")
self._optimizer_shard = value
context.set_auto_parallel_context(enable_parallel_optimizer=value)
@property
def embedding_dp_mp_config(self):
return self._embed_dp_mp_config
@property
def dp_mp_config(self):
return self._embed_dp_mp_config.dp_mp_config
@property
def moe_parallel_config(self):
return self._moe_config
default_transformer_config = TransformerOpParallelConfig()
default_embedding_parallel_config = EmbeddingOpParallelConfig()
[文档]class FeedForward(Cell):
r"""
The multilayer perceptron with two linear layers with dropout applied at final output. The first linear
will project the input dimension from hidden_size to ffn_hidden_size, the second linear will project the
dimension from ffn_hidden_size to hidden_size. The first linear is sharded on the relative dimension,
the second linear is sharded on the output dimension. The overview process can be
.. math::
Dropout((xW_1+b_1)W_2 + b_2))
where the :math:`W_1, W_2, b_1` and :math:`b_2` are trainable parameters.
Args:
hidden_size: (int): The dimension of the inputs.
ffn_hidden_size: (int): The intermediate hidden size.
dropout_rate: (float): The dropout rate for the second linear's output.
hidden_act: (str): The activation of the internal feedforward layer. Supports 'relu',
'relu6', 'tanh', 'gelu', 'fast_gelu', 'elu', 'sigmoid', 'prelu', 'leakyrelu', 'hswish',
'hsigmoid', 'logsigmoid' and so on. Default: gelu.
expert_num: (int): The number of experts used in Linear. For the case expert_num > 1, BatchMatMul is used
and the first dimension in BatchMatMul indicate expert_num. Default: 1.
param_init_type: (dtype.Number): The parameter initialization type. Should be dtype.float32 or
dtype.float16. Default: dtype.float32.
parallel_config (OpParallelConfig, MoEParallelConfig): The config of parallel setting, see
`OpParallelConfig` or `MoEParallelConfig`. When MoE is applied, MoEParallelConfig is effective,
otherwise OpParallelConfig is effective. Default `default_dpmp_config`,
an instance of `OpParallelConfig` with default args.
Inputs:
- **x** (Tensor) - should be `[batch, seq_length, hidden_size] or [batch * seq_length, hidden_size]`.
Float tensor.
Outputs:
Tensor, the output of this layer after mapping. The shape is `[batch, seq_length, hidden_size] or
[batch * seq_length, hidden_size]`.
Raises:
ValueError: `hidden_act` is not a string.
TypeError: `parallel_config` is not a subclass of OpParallelConfig.
ValueError: `ffn_hidden_size` is not a multiple of the model parallel way.
ValueError: `hidden_size` is not a multiple of the model parallel way.
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> import numpy as np
>>> from mindspore.nn.transformer import FeedForward
>>> from mindspore import dtype as mstype
>>> from mindspore import Tensor
>>> model = FeedForward(hidden_size=15, ffn_hidden_size=30, dropout_rate=0.1)
>>> tensor = Tensor(np.ones((2, 20, 15)), mstype.float32)
>>> output = model(tensor)
>>> print(output.shape)
(2, 20, 15)
"""
@_args_type_validator_check(hidden_size=Validator.check_positive_int,
ffn_hidden_size=Validator.check_positive_int,
dropout_rate=Validator.check_non_negative_float,
hidden_act=_valid_type_checks([str], "FeedForward"),
param_init_type=_valid_value_checks([mstype.float32, mstype.float16],
"FeedForward"),
parallel_config=_valid_type_checks([OpParallelConfig, MoEParallelConfig],
"FeedForward"))
def __init__(self, hidden_size,
ffn_hidden_size,
dropout_rate,
hidden_act='gelu',
expert_num=1,
param_init_type=mstype.float32,
parallel_config=default_dpmp_config):
super(FeedForward, self).__init__()
if _get_parallel_mode() in (ParallelMode.AUTO_PARALLEL,) and _is_sharding_propagation():
_check_config(parallel_config)
mp = parallel_config.model_parallel
if expert_num > 1:
ep = parallel_config.expert_parallel
else:
ep = 1
# ffn use less dp than other ops when use_moe, due to there are ops use dp and ep.
dp = int(parallel_config.data_parallel / ep)
if ffn_hidden_size % mp != 0:
raise ValueError("For 'FeedForward', the class variable 'ffn_hidden_size' must be a multiple of the"
"num of model parallel, but got the ffn_hidden_size is {} and the num of model "
"parallel is {}.".format(ffn_hidden_size, mp))
if hidden_size % mp != 0:
raise ValueError("For 'FeedForward', the class variable 'hidden_size' must be a multiple of the num of "
"model parallel, but got the hidden_size is {} and the num of model parallel is {}."
.format(hidden_size, mp))
if dropout_rate < 0 or dropout_rate >= 1:
raise ValueError("For 'FeedForward', the class variable 'dropout_rate' must be in the range [0, 1.0), "
"but got the value : {}.".format(dropout_rate))
input_size = hidden_size
output_size = ffn_hidden_size
# Project to ffn_hidden_size
self.mapping = _Linear(in_channels=input_size,
out_channels=output_size,
activation=hidden_act,
transpose_b=False,
expert_num=expert_num,
outer_batch=dp,
param_init_type=param_init_type)
# Project back to hidden_size
self.projection = _Linear(in_channels=output_size,
out_channels=input_size,
transpose_b=False,
expert_num=expert_num,
outer_batch=dp,
param_init_type=param_init_type)
if expert_num > 1:
self.projection.shard(strategy_matmul=((dp, ep, 1, mp), (ep, mp, 1)))
else:
self.projection.shard(strategy_matmul=((dp, mp), (mp, 1)))
self.projection.bias.parallel_optimizer = False
self.dropout = _Dropout(1 - dropout_rate)
self.dropout_3d = _Dropout(1 - dropout_rate)
self.dropout_4d = _Dropout(1 - dropout_rate)
self.cast = P.Cast()
else:
_check_config(parallel_config)
mp = parallel_config.model_parallel
if expert_num > 1:
ep = parallel_config.expert_parallel
else:
ep = 1
# ffn use less dp than other ops when use_moe, due to there are ops use dp and ep.
dp = int(parallel_config.data_parallel / ep)
if ffn_hidden_size % mp != 0:
raise ValueError("For 'FeedForward', the class variable 'ffn_hidden_size' must be a multiple of the"
"num of model parallel, but got the ffn_hidden_size is {} and the num of model "
"parallel is {}.".format(ffn_hidden_size, mp))
if hidden_size % mp != 0:
raise ValueError("For 'FeedForward', the class variable 'hidden_size' must be a multiple of the num of "
"model parallel, but got the hidden_size is {} and the num of model parallel is {}."
.format(hidden_size, mp))
if dropout_rate < 0 or dropout_rate >= 1:
raise ValueError("For 'FeedForward', the class variable 'dropout_rate' must be in the range [0, 1.0), "
"but got the value : {}.".format(dropout_rate))
input_size = hidden_size
output_size = ffn_hidden_size
# Project to ffn_hidden_size
self.mapping = _Linear(in_channels=input_size,
out_channels=output_size,
activation=hidden_act,
transpose_b=False,
expert_num=expert_num,
outer_batch=dp,
param_init_type=param_init_type)
if expert_num > 1:
self.mapping.shard(strategy_matmul=((dp, ep, 1, 1), (ep, 1, mp)),
strategy_bias=((dp, ep, 1, mp), (dp, ep, 1, mp)),
strategy_activation=((dp, ep, 1, mp),))
else:
self.mapping.shard(strategy_matmul=((dp, 1), (1, mp)),
strategy_bias=((dp, mp), (mp,)),
strategy_activation=((dp, mp),))
# Project back to hidden_size
self.projection = _Linear(in_channels=output_size,
out_channels=input_size,
transpose_b=False,
expert_num=expert_num,
outer_batch=dp,
param_init_type=param_init_type)
if expert_num > 1:
self.projection.shard(strategy_matmul=((dp, ep, 1, mp), (ep, mp, 1)),
strategy_bias=((dp, ep, 1, 1), (dp, ep, 1, 1)))
else:
self.projection.shard(strategy_matmul=((dp, mp), (mp, 1)),
strategy_bias=((dp, 1), (1,)))
self.projection.bias.parallel_optimizer = False
self.dropout = _Dropout(1 - dropout_rate)
self.dropout.shard(((dp, 1),))
self.dropout_3d = _Dropout(1 - dropout_rate)
self.dropout_3d.shard(((dp, 1, 1),))
self.dropout_4d = _Dropout(1 - dropout_rate)
self.dropout_4d.shard(((dp, ep, 1, 1),))
self.cast = P.Cast()
def construct(self, x):
_check_input_shape(F.shape(x), "x", self.cls_name, [2, 3])
_check_input_dtype(F.dtype(x), "x", [mstype.float32, mstype.float16], self.cls_name)
x = self.cast(x, mstype.float16)
# returned shape is [bs, seq_length, ffn_hidden_size] or [bs * seq_length, ffn_hidden_size]
hidden = self.mapping(x)
output = self.projection(hidden)
# returned shape is [bs, seq_length, ffn_hidden_size] or [bs * seq_length, ffn_hidden_size]
if len(F.shape(output)) == 3:
output = self.dropout_3d(output)
elif len(F.shape(output)) == 2:
output = self.dropout(output)
else:
output = self.dropout_4d(output)
return output
[文档]class AttentionMask(Cell):
r"""
Get the Lower triangular matrix from the input mask. The input mask is a 2D tensor (batch_size, seq_length)
with 1 and 0, where 1 indicates the current position is a valid token, otherwise not.
Args:
seq_length(int): The sequence length of the input tensor.
parallel_config(OpParallelConfig): The parallel configure. Default `default_dpmp_config`,
an instance of `OpParallelConfig` with default args.
Inputs:
- **input_mask** (Tensor) - The mask indicating whether each position is a valid input with
(batch_size, seq_length).
Outputs:
Tensor. The attention mask matrix with shape (batch_size, seq_length, seq_length).
Raises:
TypeError: `seq_length` is not an integer.
ValueError: `seq_length` is not a positive value.
TypeError: `parallel_config` is not a subclass of OpParallelConfig.
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> import numpy as np
>>> from mindspore.nn.transformer import AttentionMask
>>> from mindspore import Tensor
>>> mask = AttentionMask(seq_length=4)
>>> mask_array = np.array([[1, 1, 1, 0]], np.float32)
>>> inputs = Tensor(mask_array)
>>> res = mask(inputs)
>>> print(res)
[[[1. 0. 0. 0]
[1. 1. 0. 0]
[1. 1. 1. 0]
[0. 0. 0. 0]]]
"""
@_args_type_validator_check(seq_length=Validator.check_positive_int,
parallel_config=_valid_type_checks([OpParallelConfig], "AttentionMask"))
def __init__(self, seq_length, parallel_config=default_dpmp_config):
super(AttentionMask, self).__init__()
self.seq_length = seq_length
self.not_equal = P.NotEqual().shard(((parallel_config.data_parallel, 1), ()))
self.reshape = P.Reshape()
self.mul = P.BatchMatMul().shard(
((parallel_config.data_parallel, 1, 1), (parallel_config.data_parallel, 1, 1)))
self.expand_dim = P.ExpandDims().shard(((1, 1),))
ones = np.ones(shape=(seq_length, seq_length))
# Default lower triangle mask matrix
self.lower_triangle_mask = Tensor(np.tril(ones), mstype.float32)
self.multiply = P.Mul().shard(((parallel_config.data_parallel, 1, 1), (1, 1, 1)))
def construct(self, input_mask):
_check_input_shape(F.shape(input_mask), "input_mask", self.cls_name, 2)
_check_input_dtype(F.dtype(input_mask), "input_mask", [mstype.float32, mstype.float16], self.cls_name)
_check_input_shape_value(F.shape(input_mask), 1, "input_mask", self.cls_name, self.seq_length)
input_mask = P.Cast()(self.not_equal(input_mask, 0), mstype.float16)
input_shape = P.Shape()(input_mask)
shape_right = (input_shape[0], 1, input_shape[1])
shape_left = input_shape + (1,)
# Mask the padded inputs
mask_left = self.reshape(input_mask, shape_left)
mask_right = self.reshape(input_mask, shape_right)
attention_mask = self.mul(mask_left, mask_right)
lower_traiangle = self.expand_dim(self.lower_triangle_mask, 0)
# the returned shape is [bs, seq_length, seq_length]
attention_mask = self.multiply(
attention_mask, lower_traiangle)
return attention_mask
[文档]class VocabEmbedding(Cell):
"""
The embedding lookup table from the 0-th dim of the parameter table. When the parallel_config.vocab_emb_dp is
True and in the `AUTO_PARALLEL` mode, the embedding lookup will be trained by the data parallel way, as the
parameters will be repeated on each device. If false, the embedding table will be sharded into n parts at
the 0-th dimension of the embedding table, where the n is the model parallel way determined by the
parallel_config(EmbeddingOpParallelConfig).
Note:
When `AUTO_PARALLEL` or `SEMI_AUTO_PARALLEL` mode is enabled, this layer support only 2-d dimension inputs,
as the shard is designed for 2d inputs.
Args:
vocab_size: (int): Size of the dictionary of embeddings.
embedding_size: (int): The size of each embedding vector.
parallel_config(EmbeddingOpParallelConfig): The parallel config of network. Default
`default_embedding_parallel_config`, an instance of `EmbeddingOpParallelConfig` with default args.
param_init: (Union[Tensor, str, Initializer, numbers.Number]): Initializer for the embedding_table.
Refer to class `initializer` for the values of string when a string
is specified. Default: 'normal'.
Inputs:
- **input_ids** (Tensor) - The tokenized inputs with datatype int32 with shape (batch_size, seq_length)
Outputs:
Tuple, a tuple contains (`output`, `embedding_table`)
- **output** (Tensor) - The embedding vector for the input with shape (batch_size,
seq_length, embedding_size).
- **embedding_table** (Tensor) - The embedding table with shape (vocab_size, embedding_size).
Raises:
ValueError: If the parallel_config.vocab_emb_dp is True, the vocab size is not a multiple of
parallel_config.model_parallel
ValueError: `vocab_size` is not a positive value.
ValueError: `embedding_size` is not a positive value.
TypeError: `parallel_config` is not a subclass of OpParallelConfig.
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> import numpy as np
>>> from mindspore.nn.transformer import VocabEmbedding
>>> from mindspore import Tensor
>>> from mindspore import dtype as mstype
>>> model = VocabEmbedding(vocab_size=30, embedding_size=30)
>>> tensor = Tensor(np.ones((20, 15)), mstype.int32)
>>> output, table = model(tensor)
>>> print(output.shape)
(20, 15, 30)
>>> print(table.shape)
(30, 30)
"""
@_args_type_validator_check(vocab_size=Validator.check_positive_int,
embedding_size=Validator.check_positive_int,
parallel_config=_valid_type_checks([EmbeddingOpParallelConfig], "VocabEmbedding"))
def __init__(self, vocab_size, embedding_size, parallel_config=default_embedding_parallel_config,
param_init='normal'):
super(VocabEmbedding, self).__init__()
_check_config(parallel_config)
self.vocab_size = vocab_size
self.embedding_size = embedding_size
self.embedding_table = Parameter(initializer(param_init, [self.vocab_size, self.embedding_size]),
name='embedding_table', parallel_optimizer=False)
if parallel_config.vocab_emb_dp:
self.gather = P.GatherV2().shard(((1, 1), (parallel_config.data_parallel, 1)))
logger.info(f"Using {parallel_config.data_parallel} data parallel for the embedding lookup.")
else:
if self.vocab_size % parallel_config.model_parallel != 0:
raise ValueError(f"The vocab size of the embedding {self.vocab_size} must be a "
f"multiple of parallel_config.model_parallel {parallel_config.model_parallel}.")
self.gather = P.GatherV2().shard(((parallel_config.model_parallel, 1), (1, 1)))
logger.info(f"Using {parallel_config.data_parallel} model parallel for the embedding lookup.")
def construct(self, input_ids):
_check_input_shape(F.shape(input_ids), "input_ids", self.cls_name, 2)
_check_input_dtype(F.dtype(input_ids), "input_ids", [mstype.int32], self.cls_name)
output = self.gather(self.embedding_table, input_ids, 0)
return output, self.embedding_table
[文档]class MultiHeadAttention(Cell):
r"""
This is an implementation of multihead attention in the paper `Attention is all you need
<https://arxiv.org/pdf/1706.03762v5.pdf>`_. Given the query vector with source length, and the
key and value vector with target length, the attention will be performed as the following
.. math::
MultiHeadAttention(query, key, vector) = Concat(head_1, \dots, head_h)W^O
where :math:`head_i = Attention(QW_i^Q, KW_i^K, VW_i^V)`. The default is with a bias.
if query, key and value tensor is same, then it will be self attention.
Args:
batch_size(int): The batch size of the input tensor.
src_seq_length(int): The sequence length of the query vector.
tgt_seq_length(int): The sequence length of the key and value vector.
hidden_size(int): The hidden size of the input.
num_heads(int): The number of the heads.
hidden_dropout_rate(float): The dropout rate of the final output of the layer. Default:0.1.
attention_dropout_rate(float): The dropout rate of the attention scores. Default:0.1.
compute_dtype(dtype.Number): The computation type of dense. Default dtype.float16.
Should be dtype.float32 or dtype.float16.
softmax_compute_type(dtype.Number): The type of softmax computation module. Default dtype.float32.
Should be dtype.float32 or dtype.float16.
param_init_type(dtype.Number): The parameter initialization type of the module. Default dtype.float32.
Should be dtype.float32 or dtype.float16.
use_past(bool): Use the past state to compute, used for incremental prediction. For example, if we have two
words and want to generate the ten more words. We just need to compute the two words' state only once,
and generate the next word one by one. When use_past is True, there are two steps to run the prediction.
In the first step, set the is_first_iteration to be True by
`model.add_flags_recursive(is_first_iteration=True)`, and pass the full inputs. Then, set the
is_first_iteration to be False by `model.add_flags_recursive(is_first_iteration=False)`. At this moment,
pass the single step's input tensor, and loop it. Default False.
parallel_config(OpParallelConfig): The parallel configure. Default `default_dpmp_config`,
an instance of `OpParallelConfig` with default args.
Inputs:
- **query_tensor** (Tensor) - The query vector with shape (batch_size, src_seq_length, hidden_size) or
(batch_size * src_seq_length, hidden_size), if the use_past is False or is_first_iteration=True.
Otherwise, must be (batch_size, 1, hidden_size)
- **key_tensor** (Tensor) - The key vector with shape (batch_size, tgt_seq_length, hidden_size) or
(batch_size * tgt_seq_length, hidden_size), if the use_past is False or is_first_iteration=True.
Otherwise, must be (batch_size, 1, hidden_size)
- **value_tensor** (Tensor) - The value vector with shape (batch_size, tgt_seq_length, hidden_size) or
(batch_size * tgt_seq_length, hidden_size), if the use_past is False or is_first_iteration=True.
Otherwise, must be (batch_size, 1, hidden_size)
- **attention_mask** (Tensor) - The attention mask matrix with shape (batch_size, src_seq_length,
tgt_seq_length), if the use_past is False or is_first_iteration=True. Otherwise,
must be (batch_size, 1, tgt_seq_length)
- **key_past** (Tensor) - Float16 tensor with shape (batch_size, num_heads, size_per_head, tgt_seq_length).
The past calculated key vector. Used for incremental prediction when the use_past is True.
Default None.
- **value_past** (Tensor) - Float16 tensor with shape
(batch_size, num_heads, tgt_seq_length, size_per_head).
The past calculated value vector. Used for incremental prediction when the use_past is True.
Default None.
- **batch_valid_length** (Tensor) - Int32 tensor with shape (batch_size,) the past calculated the index.
Used for incremental prediction when the use_past is True. Default None.
Outputs:
Tuple, a tuple contains(`output`, `layer_present`)
- **output** (Tensor) - Tensor, the float tensor of the output of the layer with
shape (batch_size, src_seq_length, hidden_size) or (batch_size * src_seq_length, hidden_size),
if the use_past is False or is_first_iteration=True. Otherwise, it will be (batch_size, 1, hidden_size).
- **layer_present** (Tuple) - A tuple of the Tensor of the projected key and value vector with
((batch_size, num_heads, size_per_head, tgt_seq_length),
(batch_size, num_heads, tgt_seq_length, size_per_head)).
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> import numpy as np
>>> from mindspore.nn.transformer import MultiHeadAttention
>>> from mindspore import dtype as mstype
>>> from mindspore import Tensor
>>> model = MultiHeadAttention(batch_size=2, hidden_size=15, src_seq_length=20, tgt_seq_length=20,
... num_heads=3)
>>> from_tensor = Tensor(np.ones((2, 20, 15)), mstype.float32)
>>> to_tensor = Tensor(np.ones((2, 20, 15)), mstype.float16)
>>> attention_mask = Tensor(np.ones((2, 20, 20)), mstype.float16)
>>> attn_out, past = model(from_tensor, to_tensor, to_tensor, attention_mask)
>>> print(attn_out.shape)
(2, 20, 15)
>>> print(past[0].shape)
(2, 3, 5, 20)
>>> print(past[1].shape)
(2, 3, 20, 5)
>>> # When use use_past=True, it includes two steps to implement the incremental prediction.
>>> # Step 1: set is_first_iteration=True, and input the full sequence length's state.
>>> # We need to prepare the memory parameters for saving key and value states firstly.
>>> model = MultiHeadAttention(batch_size=2, hidden_size=15, src_seq_length=20, tgt_seq_length=20,
... num_heads=3, use_past=True)
>>> key_past = Tensor(np.zeros(shape=(2, 3, 5, 20)), mstype.float16)
>>> value_past = Tensor(np.zeros(shape=(2, 3, 20, 5)), mstype.float16)
>>> batch_valid_length = Tensor(np.ones((2,)), mstype.int32)
>>> # Set is_first_iteration=True to generate the full memory states
>>> model.add_flags_recursive(is_first_iteration=True)
>>> attn_out, past = model(from_tensor, to_tensor, to_tensor, attention_mask, key_past, value_past,
... batch_valid_length)
>>> print(attn_out.shape)
(2, 20, 15)
>>> print(past[0].shape)
(2, 3, 5, 20)
>>> print(past[1].shape)
(2, 3, 20, 5)
>>> from_tensor = Tensor(np.ones((2, 1, 15)), mstype.float32)
>>> to_tensor = Tensor(np.ones((2, 1, 15)), mstype.float16)
>>> attention_mask = Tensor(np.ones((2, 1, 20)), mstype.float16)
>>> # Step 2: set is_first_iteration=False, and pass the single word to run the prediction rather than the
>>> # full sequence.
>>> model.add_flags_recursive(is_first_iteration=False)
>>> attn_out, past = model(from_tensor, to_tensor, to_tensor, attention_mask, key_past, value_past,
... batch_valid_length)
>>> print(attn_out.shape)
(2, 1, 15)
>>> print(past[0].shape)
(2, 3, 5, 20)
>>> print(past[1].shape)
(2, 3, 20, 5)
"""
@_args_type_validator_check(batch_size=Validator.check_positive_int,
hidden_size=Validator.check_positive_int,
num_heads=Validator.check_positive_int,
src_seq_length=Validator.check_positive_int,
tgt_seq_length=Validator.check_positive_int,
attention_dropout_rate=Validator.check_non_negative_float,
hidden_dropout_rate=Validator.check_non_negative_float,
compute_dtype=_valid_value_checks([mstype.float32, mstype.float16],
"MultiHeadAttention"),
softmax_compute_type=_valid_value_checks([mstype.float32, mstype.float16],
"MultiHeadAttention"),
param_init_type=_valid_value_checks([mstype.float32, mstype.float16],
"MultiHeadAttention"),
parallel_config=_valid_type_checks([OpParallelConfig],
"MultiHeadAttention"),
use_past=Validator.check_bool)
def __init__(self, batch_size,
src_seq_length,
tgt_seq_length,
hidden_size,
num_heads,
hidden_dropout_rate=0.1,
attention_dropout_rate=0.1,
compute_dtype=mstype.float16,
softmax_compute_type=mstype.float32,
param_init_type=mstype.float32,
use_past=False,
parallel_config=default_dpmp_config):
super(MultiHeadAttention, self).__init__()
if _get_parallel_mode() in (ParallelMode.AUTO_PARALLEL,) and _is_sharding_propagation():
_check_config(parallel_config)
self.is_parallel_mode = _get_parallel_mode() in (
ParallelMode.SEMI_AUTO_PARALLEL, ParallelMode.AUTO_PARALLEL)
self.src_seq_length = src_seq_length
self.tgt_seq_length = tgt_seq_length
self.hidden_size = hidden_size
self.batch_size = batch_size
if hidden_dropout_rate < 0 or hidden_dropout_rate >= 1:
raise ValueError("For 'MultiHeadAttention', the class variable 'hidden_dropout_rate' must be "
"in range [0, 1.0), but got the value : {}.".format(hidden_dropout_rate))
if attention_dropout_rate < 0 or attention_dropout_rate >= 1:
raise ValueError("For 'MultiHeadAttention', the class variable 'attention_dropout_rate' must be "
"in range [0, 1.0), but got the value : {}.".format(attention_dropout_rate))
if hidden_size % num_heads != 0:
raise ValueError("For 'MultiHeadAttention', the class variable 'hidden_size' should be a multiple "
"of 'num_heads', but got the hidden_size is {} and the num_heads is {}."
.format(hidden_size, num_heads))
if num_heads % parallel_config.model_parallel != 0:
raise ValueError("For 'MultiHeadAttention', the class variable 'num_heads' must be a multiple of "
"'parallel_config.model_parallel', but got the num_heads is {} "
"and the parallel_config.model_parallel is {}."
.format(num_heads, parallel_config.model_parallel))
if self.is_parallel_mode and batch_size % parallel_config.data_parallel != 0:
raise ValueError("For 'MultiHeadAttention', the class variable 'batch_size' must be a multiple of "
"'parallel_config.data_parallel', but got the batch_size is {} "
"and the parallel_config.data_parallel is {}."
.format(batch_size, parallel_config.data_parallel))
self.is_first_iteration = True
# Output layer
self.projection = _Linear(in_channels=hidden_size,
out_channels=hidden_size,
transpose_b=False,
param_init_type=param_init_type).to_float(compute_dtype)
self.projection.shard(strategy_bias=((parallel_config.data_parallel, 1), (1,)),
strategy_matmul=((parallel_config.data_parallel, parallel_config.model_parallel),
(parallel_config.model_parallel, 1)))
self.projection.bias.parallel_optimizer = False
self.transpose = P.Transpose()
self.merger_head_transpose = P.Transpose()
self.reshape = P.Reshape()
self.n_head = num_heads
# embedding size per head
self.size_per_head = hidden_size // self.n_head
self.concat_k = P.Concat(axis=3)
self.concat_v = P.Concat(axis=2)
self.multiply_data = Tensor([
-10000.0,
], dtype=softmax_compute_type)
self.batch_matmul = P.BatchMatMul()
self.real_div = P.RealDiv()
self.sub = P.Sub()
self.mul = P.Mul()
self.add = P.Add()
# Normalize factor for attention, sqrt(dk) as widely used
self.scale_factor = Tensor(math.sqrt(math.sqrt(self.size_per_head)))
self.use_past = use_past
self.dropout = _Dropout(1 - hidden_dropout_rate)
self.prob_dropout = _Dropout(1 - attention_dropout_rate)
self.softmax = nn.Softmax().to_float(softmax_compute_type)
self.expand_dims = P.ExpandDims()
# Query
self.dense1 = _Linear(hidden_size,
hidden_size,
param_init_type=param_init_type).to_float(compute_dtype)
# Key
self.dense2 = _Linear(hidden_size,
hidden_size,
param_init_type=param_init_type).to_float(compute_dtype)
# Value
self.dense3 = _Linear(hidden_size,
hidden_size,
param_init_type=param_init_type).to_float(compute_dtype)
self.dtype = compute_dtype
self.softmax_dtype = softmax_compute_type
if self.use_past:
# operators used for state reuse
seq_range = np.arange(src_seq_length).reshape(1, 1, -1)
self.range = Tensor(np.tile(seq_range, (batch_size, 1, 1)), mstype.int32)
self.seq_length = src_seq_length
self.attention_mask = Tensor(np.tril(np.ones(shape=(self.seq_length, self.seq_length))), mstype.int32)
self.slice = P.StridedSlice().shard(((1, 1, 1, 1),))
self.not_equal = P.NotEqual().shard(((1, 1, 1, 1), ()))
self.reducesum = P.ReduceSum().shard(((1, 1, 1, 1),))
self.expand_dims = P.ExpandDims().shard(((1, 1, 1),))
self.tensor_le = P.LessEqual().shard(((1, 1, 1), (1, 1, 1)))
self.add = P.Add().shard(((1, 1, 1, 1), (1, 1, 1, 1)))
self.equal = P.Equal().shard(((1, 1, 1), (1, 1, 1)))
self.sub1 = P.Sub().shard(((1,), ()))
self.tile = P.Tile().shard(((1, 1, 1, 1),))
self.less = P.Less().shard(((1, 1, 1), (1, 1, 1)))
self.mul1 = P.Mul().shard(((1, 1, 1, 1), (1, 1, 1, 1)))
else:
_check_config(parallel_config)
self.is_parallel_mode = _get_parallel_mode() in (
ParallelMode.SEMI_AUTO_PARALLEL, ParallelMode.AUTO_PARALLEL)
self.src_seq_length = src_seq_length
self.tgt_seq_length = tgt_seq_length
self.hidden_size = hidden_size
self.batch_size = batch_size
if hidden_dropout_rate < 0 or hidden_dropout_rate >= 1:
raise ValueError("For 'MultiHeadAttention', the class variable 'hidden_dropout_rate' must be "
"in range [0, 1.0), but got the value : {}.".format(hidden_dropout_rate))
if attention_dropout_rate < 0 or attention_dropout_rate >= 1:
raise ValueError("For 'MultiHeadAttention', the class variable 'attention_dropout_rate' must be "
"in range [0, 1.0), but got the value : {}.".format(attention_dropout_rate))
if hidden_size % num_heads != 0:
raise ValueError("For 'MultiHeadAttention', the class variable 'hidden_size' should be a multiple "
"of 'num_heads', but got the hidden_size is {} and the num_heads is {}."
.format(hidden_size, num_heads))
if num_heads % parallel_config.model_parallel != 0:
raise ValueError("For 'MultiHeadAttention', the class variable 'num_heads' must be a multiple of "
"'parallel_config.model_parallel', but got the num_heads is {} "
"and the parallel_config.model_parallel is {}."
.format(num_heads, parallel_config.model_parallel))
if self.is_parallel_mode and batch_size % parallel_config.data_parallel != 0:
raise ValueError("For 'MultiHeadAttention', the class variable 'batch_size' must be a multiple of "
"'parallel_config.data_parallel', but got the batch_size is {} "
"and the parallel_config.data_parallel is {}."
.format(batch_size, parallel_config.data_parallel))
self.is_first_iteration = True
# Output layer
self.projection = _Linear(in_channels=hidden_size,
out_channels=hidden_size,
transpose_b=False,
param_init_type=param_init_type).to_float(compute_dtype)
self.projection.shard(strategy_bias=((parallel_config.data_parallel, 1), (1,)),
strategy_matmul=((parallel_config.data_parallel, parallel_config.model_parallel),
(parallel_config.model_parallel, 1)))
self.projection.bias.parallel_optimizer = False
self.transpose = P.Transpose().shard(
((parallel_config.data_parallel, 1, parallel_config.model_parallel, 1),))
self.merger_head_transpose = P.Transpose().shard(
((parallel_config.data_parallel, parallel_config.model_parallel, 1, 1),))
self.reshape = P.Reshape()
self.n_head = num_heads
# embedding size per head
self.size_per_head = hidden_size // self.n_head
self.concat_k = P.Concat(axis=3)
self.concat_v = P.Concat(axis=2)
self.multiply_data = Tensor([
-10000.0,
], dtype=softmax_compute_type)
self.batch_matmul = P.BatchMatMul().shard(
((parallel_config.data_parallel, parallel_config.model_parallel, 1, 1),
(parallel_config.data_parallel, parallel_config.model_parallel, 1, 1)))
self.real_div = P.RealDiv().shard(
((parallel_config.data_parallel, parallel_config.model_parallel, 1, 1), ()))
self.sub = P.Sub().shard(
((1,), (parallel_config.data_parallel, 1, 1, 1)))
self.mul = P.Mul().shard(
((parallel_config.data_parallel, 1, 1, 1), (1,)))
self.add = P.Add().shard(
((parallel_config.data_parallel, 1, 1, 1),
(parallel_config.data_parallel, parallel_config.model_parallel, 1, 1)))
# Normalize factor for attention, sqrt(dk) as widely used
self.scale_factor = Tensor(math.sqrt(math.sqrt(self.size_per_head)))
self.use_past = use_past
self.dropout = _Dropout(1 - hidden_dropout_rate)
self.dropout.shard(((parallel_config.data_parallel, 1),))
self.prob_dropout = _Dropout(1 - attention_dropout_rate)
self.prob_dropout.shard(
((parallel_config.data_parallel, parallel_config.model_parallel, 1, 1),))
self.softmax = nn.Softmax().to_float(softmax_compute_type)
self.softmax.softmax.shard(((parallel_config.data_parallel, parallel_config.model_parallel, 1),))
self.expand_dims = P.ExpandDims().shard(((parallel_config.data_parallel, 1, 1),))
# Query
self.dense1 = _Linear(hidden_size,
hidden_size,
param_init_type=param_init_type).to_float(compute_dtype)
self.dense1.shard(strategy_matmul=((parallel_config.data_parallel, 1), (parallel_config.model_parallel, 1)),
strategy_bias=((parallel_config.data_parallel, parallel_config.model_parallel),
(parallel_config.model_parallel,)))
# Key
self.dense2 = _Linear(hidden_size,
hidden_size,
param_init_type=param_init_type).to_float(compute_dtype)
self.dense2.shard(strategy_matmul=((parallel_config.data_parallel, 1), (parallel_config.model_parallel, 1)),
strategy_bias=((parallel_config.data_parallel, parallel_config.model_parallel),
(parallel_config.model_parallel,)))
# Value
self.dense3 = _Linear(hidden_size,
hidden_size,
param_init_type=param_init_type).to_float(compute_dtype)
self.dense3.shard(strategy_matmul=((parallel_config.data_parallel, 1), (parallel_config.model_parallel, 1)),
strategy_bias=((parallel_config.data_parallel, parallel_config.model_parallel),
(parallel_config.model_parallel,)))
self.dtype = compute_dtype
self.softmax_dtype = softmax_compute_type
if self.use_past:
# operators used for state reuse
seq_range = np.arange(src_seq_length).reshape(1, 1, -1)
self.range = Tensor(np.tile(seq_range, (batch_size, 1, 1)), mstype.int32)
self.seq_length = src_seq_length
self.attention_mask = Tensor(np.tril(np.ones(shape=(self.seq_length, self.seq_length))), mstype.int32)
self.slice = P.StridedSlice().shard(((1, 1, 1, 1),))
self.not_equal = P.NotEqual().shard(((1, 1, 1, 1), ()))
self.reducesum = P.ReduceSum().shard(((1, 1, 1, 1),))
self.expand_dims = P.ExpandDims().shard(((1, 1, 1),))
self.tensor_le = P.LessEqual().shard(((1, 1, 1), (1, 1, 1)))
self.add = P.Add().shard(((1, 1, 1, 1), (1, 1, 1, 1)))
self.equal = P.Equal().shard(((1, 1, 1), (1, 1, 1)))
self.sub1 = P.Sub().shard(((1,), ()))
self.tile = P.Tile().shard(((1, 1, 1, 1),))
self.less = P.Less().shard(((1, 1, 1), (1, 1, 1)))
self.mul1 = P.Mul().shard(((1, 1, 1, 1), (1, 1, 1, 1)))
def construct(self, query_tensor, key_tensor, value_tensor, attention_mask, key_past=None,
value_past=None, batch_valid_length=None):
self._check_inputs(query_tensor, key_tensor, value_tensor, attention_mask, key_past,
value_past, batch_valid_length)
query_tensor, key_tensor, value_tensor, batch_size, ori_shape = self._convert_to_2d_tensor(query_tensor,
key_tensor,
value_tensor,
attention_mask)
# multi head attention: query, key, value are derived from the same inputs
query = self.dense1(query_tensor)
key = self.dense2(key_tensor)
value = self.dense3(value_tensor)
# the returned shape is [bs, num_heads, seq_length, size_per_head]
query = self.transpose(
F.reshape(
query,
(batch_size, -1, self.n_head, self.size_per_head)),
(0, 2, 1, 3))
# the returned shape is [bs, size_per_head, seq_length, num_heads]
key = self.transpose(
F.reshape(
key, (batch_size, -1, self.n_head, self.size_per_head)),
(0, 2, 3, 1))
# the returned shape is [bs, num_heads, seq_length, size_per_head]
value = self.transpose(
F.reshape(
value,
(batch_size, -1, self.n_head, self.size_per_head)),
(0, 2, 1, 3))
# support input shape is [bs, seq, seq] or [bs, heads, seq, seq]
if len(F.shape(attention_mask)) == 3:
# expand attention mask from [bs, seq, seq] -> [bs, 1, seq, seq]
attention_mask = self.expand_dims(attention_mask, 1)
# key and value for current token(s)
key_present = key
value_present = value
if self.use_past:
# The first graph with the input size of (bs, seq_length)
if self.is_first_iteration:
# Get the valid input length without padding
valid_length_vector = F.cast(self.less(self.range, batch_valid_length.view(-1, 1, 1)), self.dtype)
# Cover the key and value numbers corresponding to the padding position
key_present = self.mul1(key, self.expand_dims(valid_length_vector, 2))
value_present = self.mul1(value, self.expand_dims(valid_length_vector, 3))
# The second graph with the inpus size of (bs, 1)
# the shape of query is (bs, num_heads, 1, size_per_head)
# the shape of key is (bs, num_heads, size_per_head, 1)
# the shape of value is (bs, num_heads, 1, size_per_head)
else:
# Get the current token position index
valid_length = self.reducesum(F.cast(self.not_equal(self.slice(key_past, (0, 0, 0, 0),
(F.shape(key_tensor)[0], 1, 1,
self.src_seq_length),
(1, 1, 1, 1)),
0), mstype.float32), (1, 2, 3))
valid_length = F.reshape(valid_length, (-1, 1, 1))
valid_length_vector = F.cast(self.equal(valid_length, self.range), self.dtype)
# Pad the key and value to seq_length with only the position index not zero
current_key = self.mul1(self.tile(key, (1, 1, 1, self.seq_length)),
self.expand_dims(valid_length_vector, 2))
current_value = self.mul1(self.tile(value, (1, 1, self.seq_length, 1)),
self.expand_dims(valid_length_vector, 3))
# Concat the previous saved state and current state
key = self.add(key_past, current_key)
value = self.add(value_past, current_value)
# Update key_present and value_present for state update
key_present = key
value_present = value
attention_mask = F.reshape(self.attention_mask, (self.seq_length, self.seq_length, 1, 1))
layer_present = (key_present, value_present)
# multi head attention considering attention mask
# the return shape is [bs * seq_length, hidden_size]
attention = self._attn(query, key, value, attention_mask)
# Output
output = self.projection(attention)
output = self.dropout(output)
output = F.reshape(output, ori_shape)
return output, layer_present
def _check_inputs(self, query_tensor, key_tensor, value_tensor, attention_mask, key_past=None,
value_past=None, batch_valid_length=None):
r"""Check inputs"""
if not self.use_past or (self.use_past and self.is_first_iteration):
_check_shape_equal(F.shape(query_tensor), "query_tensor", self.cls_name,
[[self.batch_size, self.src_seq_length, self.hidden_size],
[self.batch_size * self.src_seq_length, self.hidden_size]])
_check_shape_equal(F.shape(key_tensor), "key_tensor", self.cls_name,
[[self.batch_size, self.tgt_seq_length, self.hidden_size],
[self.batch_size * self.tgt_seq_length, self.hidden_size]])
_check_shape_equal(F.shape(value_tensor), "value_tensor", self.cls_name,
[[self.batch_size, self.tgt_seq_length, self.hidden_size],
[self.batch_size * self.tgt_seq_length, self.hidden_size]])
_check_shape_equal(F.shape(attention_mask), "attention_mask", self.cls_name,
[self.batch_size, self.src_seq_length, self.tgt_seq_length])
else:
_check_shape_equal(F.shape(query_tensor), "query_tensor", self.cls_name,
[[self.batch_size, 1, self.hidden_size], [self.batch_size, self.hidden_size]])
_check_shape_equal(F.shape(key_tensor), "key_tensor", self.cls_name,
[[self.batch_size, 1, self.hidden_size], [self.batch_size, self.hidden_size]])
_check_shape_equal(F.shape(value_tensor), "value_tensor", self.cls_name,
[[self.batch_size, 1, self.hidden_size], [self.batch_size, self.hidden_size]])
_check_shape_equal(F.shape(attention_mask), "attention_mask", self.cls_name,
[[self.batch_size, 1, self.tgt_seq_length], [self.batch_size, self.hidden_size]])
_check_input_dtype(F.dtype(query_tensor), "query_tensor", [mstype.float32, mstype.float16], self.cls_name)
_check_input_dtype(F.dtype(key_tensor), "key_tensor", [mstype.float32, mstype.float16], self.cls_name)
_check_input_dtype(F.dtype(value_tensor), "value_tensor", [mstype.float32, mstype.float16], self.cls_name)
_check_input_dtype(F.dtype(attention_mask), "attention_mask", [mstype.float32, mstype.float16], self.cls_name)
key_is_tensor = isinstance(key_past, Tensor)
value_is_tensor = isinstance(value_past, Tensor)
batch_valid_length_is_tensor = isinstance(batch_valid_length, Tensor)
key_is_default = key_past is None
value_is_default = value_past is None
batch_is_default = batch_valid_length is None
_check_past_none_input_none(self.use_past, "key_past", self.cls_name, None, key_is_tensor,
key_is_default)
_check_past_none_input_none(self.use_past, "value_past", self.cls_name, None, value_is_tensor,
value_is_default)
_check_past_none_input_none(self.use_past, "batch_valid_length", self.cls_name, None,
batch_valid_length_is_tensor, batch_is_default)
if self.use_past:
_check_shape_equal(F.shape(key_past), "key_past", self.cls_name,
[self.batch_size, self.n_head, self.size_per_head, self.tgt_seq_length])
_check_input_dtype(F.dtype(key_past), "key_past", [mstype.float16], self.cls_name)
_check_shape_equal(F.shape(value_past), "value_past", self.cls_name,
[self.batch_size, self.n_head, self.tgt_seq_length, self.size_per_head])
_check_input_dtype(F.dtype(value_past), "value_past", [mstype.float16], self.cls_name)
_check_shape_equal(F.shape(batch_valid_length), "batch_valid_length", self.cls_name, [self.batch_size])
_check_input_dtype(F.dtype(batch_valid_length), "batch_valid_length", [mstype.int32], self.cls_name)
return True
def _convert_to_2d_tensor(self, query_tensor, key_tensor, value_tensor, attention_mask):
"""convert a nd tensor to a 2d tensor"""
query_shape = F.shape(query_tensor)
query_tensor = F.reshape(query_tensor, (-1, query_shape[-1]))
key_shape = F.shape(key_tensor)
key_tensor = F.reshape(key_tensor, (-1, key_shape[-1]))
value_shape = F.shape(value_tensor)
value_tensor = F.reshape(value_tensor, (-1, value_shape[-1]))
return query_tensor, key_tensor, value_tensor, F.shape(attention_mask)[0], query_shape
def _merge_heads(self, x):
"""
convert a 4d input to a 2d output
Inputs:
x: input tensor
Output:
x_merge: the 2d output
"""
x = self.merger_head_transpose(
x, (0, 2, 1, 3)) # bs, seq_length, head, size_per_head
x_shape = P.Shape()(x)
new_shape = (-1, x_shape[-2] * x_shape[-1])
x_merge = self.reshape(x, new_shape)
return x_merge
def _attn(self, query, key, value, attention_mask):
"""
Get the weighted score along the seq_length
Inputs:
query: the query matrix
key: the key matrix
value: the value matrix
attention_mask: the attention mask matrix with shape (batch_size,
1, seq_length, seq_length)
Outputs:
weighted_values: Tensor, the weighted sum scores
"""
# Normalize query and key before MatMul, default off
# Attention score [bs, num_heads, seq_length, seq_length]
factor = P.Cast()(self.scale_factor, P.DType()(query))
query = self.real_div(query, factor)
key = self.real_div(key, factor)
score = self.batch_matmul(query, key)
ori_dtype = P.DType()(score)
score = P.Cast()(score, self.softmax_dtype)
# for input size of (bs, 1) namely the second graph,
# the shape of attention_mask matrix should be (bs, 1, 1, seq_length)
if self.use_past and not self.is_first_iteration:
# Calculate the current total token
current_index = self.reducesum(F.cast(self.not_equal(self.slice(key, (0, 0, 0, 0),
(F.shape(query)[0], 1, 1, self.seq_length),
(1, 1, 1, 1)),
0), mstype.float32), (1, 2, 3))
# Get the precise position index
index = self.sub1(F.cast(current_index, mstype.int32), 1)
index = F.reshape(index, (-1, 1, 1))
# Calculate the attention_mask matrix via the position index
attention_mask = F.cast(self.tensor_le(self.range, index), mstype.int32)
attention_mask = self.expand_dims(attention_mask, 2)
# Minus 10000 for the position where masked to exclude them from softmax
multiplu_out = self.sub(
P.Cast()(F.tuple_to_array((1.0,)), P.DType()(score)),
P.Cast()(attention_mask, P.DType()(score)))
adder = self.mul(multiplu_out, self.multiply_data)
attention_scores = self.add(adder, score)
shape = F.shape(attention_scores)
# attention probs
attention_probs = self.softmax(
F.reshape(attention_scores,
(shape[0], -1, shape[-1])))
attention_probs = P.Cast()(attention_probs, ori_dtype)
attention_probs = F.reshape(attention_probs, shape)
attention_probs = self.prob_dropout(attention_probs)
# Weighted sum output [bs, num_heads, seq_length, size_per_head]
weighted_values = self.batch_matmul(attention_probs, value)
attention_merge = self._merge_heads(weighted_values)
return attention_merge
[文档]class TransformerEncoderLayer(Cell):
r"""
Transformer Encoder Layer. This is an implementation of the single layer of the transformer
encoder layer, including multihead attention and feedward layer.
Args:
batch_size(int): The batch size of the input tensor.
hidden_size(int): The hidden size of the input.
ffn_hidden_size(int): The hidden size of bottleneck in the feedforward layer.
num_heads(int): The number of the heads.
seq_length(int): The input sequence length.
attention_dropout_rate(float): The dropout rate of the attention scores. Default:0.1.
hidden_dropout_rate(float): The dropout rate of the final output of the layer. Default:0.1.
post_layernorm_residual(bool): Do residuals adds before the layernorm. Default False.
layernorm_compute_type(dtype.Number): The computation type of the layernorm.
Should be dtype.float32 or dtype.float16. Default dtype.float32.
softmax_compute_type(dtype.Number): The computation type of the softmax in the attention.
Should be dtype.float32 or dtype.float16. Default mstype.float32.
param_init_type(dtype.Number): The parameter initialization type of the module.
Should be dtype.float32 or dtype.float16. Default dtype.float32.
hidden_act(str): The activation of the internal feedforward layer. Supports 'relu',
'relu6', 'tanh', 'gelu', 'fast_gelu', 'elu', 'sigmoid', 'prelu', 'leakyrelu', 'hswish',
'hsigmoid', 'logsigmoid' and so on. Default: gelu.
use_past(bool): Use the past state to compute, used for incremental prediction. For example, if we have two
words and want to generate the ten more words. We just need to compute the two words' state only once,
and generate the next word one by one. When use_past is True, there are two steps to run the prediction.
In the first step, set the is_first_iteration to be True by
`model.add_flags_recursive(is_first_iteration=True)`, and pass the full inputs. Then, set the
is_first_iteration to be False by `model.add_flags_recursive(is_first_iteration=False)`.
At this moment, pass the single step's input tensor, and loop it. Default False.
moe_config(MoEConfig): The configuration of MoE (Mixture of Expert). Default is an instance of MoEConfig
with default values. Please see `MoEConfig`.
parallel_config(OpParallelConfig, MoEParallelConfig): The parallel configure. When MoE is applied,
MoEParallelConfig is effective, otherwise OpParallelConfig is effective. Default `default_dpmp_config`,
an instance of `OpParallelConfig` with default args.
Inputs:
- **x** (Tensor) - Float Tensor, shape should be [batch_size, seq_length, hidden_size] or
[batch_size * seq_length, hidden_size], if the use_past is False or is_first_iteration=True. Otherwise,
should be [batch_size, 1, hidden_size]
- **input_mask** (Tensor) - Float Tensor, attention mask with shape [batch_size, seq_length, seq_length],
if the use_past is False or is_first_iteration=True. Otherwise, should be [batch_size, 1, hidden_size]
- **init_reset** (Tensor) - A bool tensor with shape [1], used to clear the past key parameter and
past value parameter used in the incremental prediction. Only valid when use_past is True. Default True.
- **batch_valid_length** (Tensor) - Int32 tensor with shape [batch_size] the past calculated the index.
Used for incremental prediction when the use_past is True. Default None.
Outputs:
Tuple, a tuple contains(`output`, `layer_present`).
- **output** (Tensor) - The float tensor of the output of the layer with
shape (batch_size, seq_length, hidden_size) or (batch_size * seq_length, hidden_size), if the use_past is
False or is_first_iteration=True. Otherwise, it will be (batch_size, 1, hidden_size)
- **layer_present** (Tuple) - A tuple of the Tensor of the projected key and value vector with
((batch_size, num_heads, size_per_head, seq_length),
(batch_size, num_heads, seq_length, size_per_head)).
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> import numpy as np
>>> from mindspore import dtype as mstype
>>> from mindspore.nn.transformer import TransformerEncoderLayer
>>> from mindspore import Tensor
>>> model = TransformerEncoderLayer(batch_size=2, hidden_size=8, ffn_hidden_size=64, seq_length=16,
... num_heads=2)
>>> encoder_input_value = Tensor(np.ones((2, 16, 8)), mstype.float32)
>>> encoder_input_mask = Tensor(np.ones((2, 16, 16)), mstype.float16)
>>> output, past = model(encoder_input_value, encoder_input_mask)
>>> print(output.shape)
(2, 16, 8)
>>> print(past[0].shape)
(2, 2, 4, 16)
>>> print(past[1].shape)
(2, 2, 16, 4)
>>> # When use use_past=True, it includes two steps to implement the incremental prediction.
>>> # Step 1: set is_first_iteration=True, and input the full sequence length's state.
>>> batch_valid_length = Tensor(np.ones((2,)), mstype.int32)
>>> init_reset = Tensor([True], mstype.bool_)
>>> # Set is_first_iteration=True to generate the full memory states
>>> model = TransformerEncoderLayer(batch_size=2, hidden_size=8, ffn_hidden_size=64, seq_length=16,
... num_heads=2, use_past=True)
>>> model.add_flags_recursive(is_first_iteration=True)
>>> hidden, past = model(encoder_input_value, encoder_input_mask, init_reset, batch_valid_length)
>>> print(hidden.shape)
(2, 16, 8)
>>> print(past[0].shape)
(2, 2, 4, 16)
>>> print(past[1].shape)
(2, 2, 16, 4)
>>> encoder_input_value = Tensor(np.ones((2, 1, 8)), mstype.float32)
>>> encoder_input_mask = Tensor(np.ones((2, 1, 16)), mstype.float16)
>>> init_reset = Tensor([False], mstype.bool_)
>>> # Step 2: set is_first_iteration=False, and pass the single word to run the prediction rather than
>>> # the full sequence.
>>> model.add_flags_recursive(is_first_iteration=False)
>>> hidden, past = model(encoder_input_value, encoder_input_mask, init_reset, batch_valid_length)
>>> print(hidden.shape)
(2, 1, 8)
>>> print(past[0].shape)
(2, 2, 4, 16)
>>> print(past[1].shape)
(2, 2, 16, 4)
"""
@_args_type_validator_check(batch_size=Validator.check_positive_int,
hidden_size=Validator.check_positive_int,
num_heads=Validator.check_positive_int,
ffn_hidden_size=Validator.check_positive_int,
seq_length=Validator.check_positive_int,
attention_dropout_rate=Validator.check_non_negative_float,
hidden_dropout_rate=Validator.check_non_negative_float,
hidden_act=_valid_type_checks([str], "TransformerEncoderLayer"),
post_layernorm_residual=Validator.check_bool,
layernorm_compute_type=_valid_value_checks([mstype.float32, mstype.float16],
"TransformerEncoderLayer"),
softmax_compute_type=_valid_value_checks([mstype.float32, mstype.float16],
"TransformerEncoderLayer"),
param_init_type=_valid_value_checks([mstype.float32, mstype.float16],
"TransformerEncoderLayer"),
parallel_config=_valid_type_checks([OpParallelConfig, MoEParallelConfig],
"TransformerEncoderLayer"),
use_past=Validator.check_bool)
def __init__(self,
batch_size,
hidden_size,
ffn_hidden_size,
num_heads,
seq_length,
attention_dropout_rate=0.1,
hidden_dropout_rate=0.1,
post_layernorm_residual=False,
layernorm_compute_type=mstype.float32,
softmax_compute_type=mstype.float32,
param_init_type=mstype.float32,
hidden_act='gelu',
use_past=False,
moe_config=default_moe_config,
parallel_config=default_dpmp_config):
super(TransformerEncoderLayer, self).__init__()
if _get_parallel_mode() in (ParallelMode.AUTO_PARALLEL,) and _is_sharding_propagation():
_check_config(parallel_config)
if num_heads % parallel_config.model_parallel != 0:
raise ValueError(
"For 'TransformerEncoderLayer', the class variable 'num_heads' must be divisibled by the "
"'parallel_config.model_parallel', but got the num_heads is {} and "
"parallel_config.model_parallel is {}.".format(num_heads, parallel_config.model_parallel))
if hidden_size % parallel_config.model_parallel != 0:
raise ValueError(
"For 'TransformerEncoderLayer', the class variable 'hidden_size' must be divisibled by "
"the 'parallel_config.model_parallel', but got the hidden_size is {} and parallel_config."
" model_parallel is {}.".format(hidden_size, parallel_config.model_parallel))
if ffn_hidden_size % parallel_config.model_parallel != 0:
raise ValueError(
"For 'TransformerEncoderLayer', the class variable 'ffn_hidden_size' must be divisibled "
"by the 'parallel_config.model_parallel', but got the ffn_hidden_size is {} "
"and parallel_config. model_parallel is {}."
.format(ffn_hidden_size, parallel_config.model_parallel))
_check_moe_config(moe_config, parallel_config)
self.use_moe = (moe_config.expert_num > 1)
self.use_past = use_past
self.seq_length = seq_length
self.hidden_size = hidden_size
self.batch_size = batch_size
self.layernorm1 = _LayerNorm((hidden_size,)).to_float(layernorm_compute_type)
self.layernorm2 = _LayerNorm((hidden_size,)).to_float(layernorm_compute_type)
self.attention = MultiHeadAttention(batch_size=batch_size,
src_seq_length=seq_length,
tgt_seq_length=seq_length,
hidden_size=hidden_size,
num_heads=num_heads,
hidden_dropout_rate=hidden_dropout_rate,
attention_dropout_rate=attention_dropout_rate,
softmax_compute_type=softmax_compute_type,
param_init_type=param_init_type,
use_past=use_past,
parallel_config=parallel_config.dpmp if self.use_moe
else parallel_config)
if self.use_moe:
self.output = MoE(hidden_size=hidden_size,
dropout_rate=hidden_dropout_rate,
ffn_hidden_size=ffn_hidden_size,
param_init_type=param_init_type,
hidden_act=hidden_act,
moe_config=moe_config,
parallel_config=parallel_config)
else:
# Feed Forward Network, FFN
self.output = FeedForward(hidden_size=hidden_size,
dropout_rate=hidden_dropout_rate,
ffn_hidden_size=ffn_hidden_size,
param_init_type=param_init_type,
hidden_act=hidden_act,
parallel_config=parallel_config)
self.post_layernorm_residual = post_layernorm_residual
self.add = P.Add().shard(((parallel_config.data_parallel, 1), (parallel_config.data_parallel, 1)))
self.add_3d = P.Add().shard(((parallel_config.data_parallel, 1, 1), (parallel_config.data_parallel, 1, 1)))
self.dtype = mstype.float16
self.key_past = None
self.value_past = None
if self.use_past:
# operator used for state reuse
self.reducesum = P.ReduceSum().shard(((1, 1, 1, 1),))
self.not_equal = P.NotEqual().shard(((1, 1, 1, 1), ()))
self.slice = P.StridedSlice().shard(((1, 1, 1, 1),))
size_per_head = int(hidden_size / num_heads)
self.key_shape = (batch_size, num_heads, size_per_head, seq_length)
self.value_shape = (batch_size, num_heads, seq_length, size_per_head)
# parameters saving key and value states
self.key_past = Parameter(Tensor(np.zeros(shape=self.key_shape), self.dtype), name="key_past")
self.value_past = Parameter(Tensor(np.zeros(shape=self.value_shape), self.dtype), name="value_past")
self.tile = P.Tile().shard(((1, 1),))
self.mul = P.Mul().shard(((1, 1, 1, 1), (1,)))
self.assign = P.Assign().shard(((1, 1, 1, 1), (1, 1, 1, 1)))
elif _get_parallel_mode() not in (ParallelMode.AUTO_PARALLEL,):
_check_config(parallel_config)
if num_heads % parallel_config.model_parallel != 0:
raise ValueError(
"For 'TransformerEncoderLayer', the class variable 'num_heads' must be divisibled by the "
"'parallel_config.model_parallel', but got the num_heads is {} and "
"parallel_config.model_parallel is {}.".format(num_heads, parallel_config.model_parallel))
if hidden_size % parallel_config.model_parallel != 0:
raise ValueError(
"For 'TransformerEncoderLayer', the class variable 'hidden_size' must be divisibled by "
"the 'parallel_config.model_parallel', but got the hidden_size is {} and parallel_config."
" model_parallel is {}.".format(hidden_size, parallel_config.model_parallel))
if ffn_hidden_size % parallel_config.model_parallel != 0:
raise ValueError(
"For 'TransformerEncoderLayer', the class variable 'ffn_hidden_size' must be divisibled "
"by the 'parallel_config.model_parallel', but got the ffn_hidden_size is {} "
"and parallel_config. model_parallel is {}."
.format(ffn_hidden_size, parallel_config.model_parallel))
_check_moe_config(moe_config, parallel_config)
self.use_moe = (moe_config.expert_num > 1)
self.use_past = use_past
self.seq_length = seq_length
self.hidden_size = hidden_size
self.batch_size = batch_size
self.layernorm1 = _LayerNorm((hidden_size,)).to_float(layernorm_compute_type)
self.layernorm1.shard(((parallel_config.data_parallel, 1),))
self.layernorm2 = _LayerNorm((hidden_size,)).to_float(layernorm_compute_type)
self.layernorm2.shard(((parallel_config.data_parallel, 1),))
self.attention = MultiHeadAttention(batch_size=batch_size,
src_seq_length=seq_length,
tgt_seq_length=seq_length,
hidden_size=hidden_size,
num_heads=num_heads,
hidden_dropout_rate=hidden_dropout_rate,
attention_dropout_rate=attention_dropout_rate,
softmax_compute_type=softmax_compute_type,
param_init_type=param_init_type,
use_past=use_past,
parallel_config=parallel_config.dpmp if self.use_moe
else parallel_config)
if self.use_moe:
self.output = MoE(hidden_size=hidden_size,
dropout_rate=hidden_dropout_rate,
ffn_hidden_size=ffn_hidden_size,
param_init_type=param_init_type,
hidden_act=hidden_act,
moe_config=moe_config,
parallel_config=parallel_config)
else:
# Feed Forward Network, FFN
self.output = FeedForward(hidden_size=hidden_size,
dropout_rate=hidden_dropout_rate,
ffn_hidden_size=ffn_hidden_size,
param_init_type=param_init_type,
hidden_act=hidden_act,
parallel_config=parallel_config)
self.post_layernorm_residual = post_layernorm_residual
self.add = P.Add().shard(((parallel_config.data_parallel, 1), (parallel_config.data_parallel, 1)))
self.add_3d = P.Add().shard(((parallel_config.data_parallel, 1, 1), (parallel_config.data_parallel, 1, 1)))
self.dtype = mstype.float16
self.key_past = None
self.value_past = None
if self.use_past:
# operator used for state reuse
self.reducesum = P.ReduceSum().shard(((1, 1, 1, 1),))
self.not_equal = P.NotEqual().shard(((1, 1, 1, 1), ()))
self.slice = P.StridedSlice().shard(((1, 1, 1, 1),))
size_per_head = int(hidden_size / num_heads)
self.key_shape = (batch_size, num_heads, size_per_head, seq_length)
self.value_shape = (batch_size, num_heads, seq_length, size_per_head)
# parameters saving key and value states
self.key_past = Parameter(Tensor(np.zeros(shape=self.key_shape), self.dtype), name="key_past")
self.value_past = Parameter(Tensor(np.zeros(shape=self.value_shape), self.dtype), name="value_past")
self.tile = P.Tile().shard(((1, 1),))
self.mul = P.Mul().shard(((1, 1, 1, 1), (1,)))
self.assign = P.Assign().shard(((1, 1, 1, 1), (1, 1, 1, 1)))
else:
raise RuntimeError(f"The {self.cls_name} only support sharding propagation or "
f"semi-auto parallel mode now.")
def construct(self, x, input_mask, init_reset=True, batch_valid_length=None):
self._check_input(x, input_mask, init_reset, batch_valid_length)
x_shape = F.shape(x)
x = F.reshape(x, (-1, x_shape[-1]))
input_x = self.layernorm1(x)
input_x = F.cast(input_x, self.dtype)
# indicate whether reset saved states
key_reset = None
value_reset = None
if self.use_past:
# reset states, init_reset True for reuse and False for reset
key_reset = self.assign(self.key_past, self.mul(self.key_past, F.cast(init_reset, self.dtype)))
value_reset = self.assign(self.value_past, self.mul(self.value_past, F.cast(init_reset, self.dtype)))
# add dependency for desired execution order
input_x = F.depend(input_x, key_reset)
input_x = F.depend(input_x, value_reset)
attention, layer_present = self.attention(input_x, input_x, input_x, input_mask,
self.key_past, self.value_past, batch_valid_length)
# For post-layernorm the inputs for residual path are output of self-attention and output of layernorm
if self.post_layernorm_residual:
x = self.add(input_x, attention)
# For pre-layernorm the inputs for residual path are output of self-attention and input of this layer
else:
x = self.add(x, attention)
output_x = self.layernorm2(x)
output_x = F.cast(output_x, self.dtype)
aux_loss = None
if self.use_moe:
mlp_logit, aux_loss = self.output(output_x)
else:
mlp_logit = self.output(output_x)
value_update = None
key_update = None
if self.use_past:
# current key and value
key_present, value_present = layer_present
# update key and value calculated this step
key_update = self.assign(self.key_past, key_present)
value_update = self.assign(self.value_past, value_present)
# add dependency for desired execution order
key_update = F.depend(key_update, key_reset)
value_update = F.depend(value_update, value_reset)
# add dependency for desired execution order
mlp_logit = F.depend(mlp_logit, value_update)
mlp_logit = F.depend(mlp_logit, key_update)
# if shape is 3d, we reshape the inputs of the add
if len(x_shape) == 3:
output_x = P.Reshape()(output_x, x_shape)
mlp_logit = P.Reshape()(mlp_logit, x_shape)
x = P.Reshape()(x, x_shape)
if self.post_layernorm_residual:
output = self.add_3d(output_x, mlp_logit)
else:
output = self.add_3d(x, mlp_logit)
else:
if self.post_layernorm_residual:
output = self.add(output_x, mlp_logit)
else:
output = self.add(x, mlp_logit)
output = F.reshape(output, x_shape)
if self.use_moe:
return output, layer_present, aux_loss
return output, layer_present
def _check_input(self, x, input_mask, init_reset, batch_valid_length):
r"""Check inputs"""
if not self.use_past or (self.use_past and self.is_first_iteration):
_check_shape_equal(F.shape(x), "x", self.cls_name,
[[self.batch_size, self.seq_length, self.hidden_size],
[self.batch_size * self.seq_length, self.hidden_size]])
_check_shape_equal(F.shape(input_mask), "input_mask", self.cls_name,
[self.batch_size, self.seq_length, self.seq_length])
else:
_check_shape_equal(F.shape(x), "x", self.cls_name, [self.batch_size, 1, self.hidden_size])
_check_shape_equal(F.shape(input_mask), "input_mask", self.cls_name,
[self.batch_size, 1, self.seq_length])
_check_input_dtype(F.dtype(x), "x", [mstype.float32, mstype.float16], self.cls_name)
_check_input_dtype(F.dtype(input_mask), "input_mask", [mstype.float32, mstype.float16], self.cls_name)
init_reset_is_tensor = isinstance(init_reset, Tensor)
init_reset_is_default = init_reset is True
batch_valid_length_is_tensor = isinstance(batch_valid_length, Tensor)
batch_is_default = batch_valid_length is None
_check_past_none_input_none(self.use_past, "init_reset", self.cls_name, True, init_reset_is_tensor,
init_reset_is_default)
_check_past_none_input_none(self.use_past, "batch_valid_length", self.cls_name, None,
batch_valid_length_is_tensor, batch_is_default)
if self.use_past:
_check_shape_equal(F.shape(init_reset), "init_reset", self.cls_name, [1])
_check_input_dtype(F.dtype(init_reset), "init_reset", [mstype.bool_], self.cls_name)
_check_shape_equal(F.shape(batch_valid_length), "batch_valid_length", self.cls_name, [self.batch_size])
_check_input_dtype(F.dtype(batch_valid_length), "batch_valid_length", [mstype.int32], self.cls_name)
return True
[文档]class TransformerDecoderLayer(Cell):
r"""
Transformer Decoder Layer. This is an implementation of the single layer of the transformer
decoder layer, including self-attention, cross attention and feedward layer. When the encoder_output is None,
the cross attention will not be effective.
Args:
hidden_size(int): The hidden size of the input.
ffn_hidden_size(int): The hidden size of bottleneck in the feedforward layer.
num_heads(int): The number of the heads.
batch_size(int): The batch size of the input tensor.
src_seq_length(int): The input source sequence length.
tgt_seq_length(int): The input target sequence length.
attention_dropout_rate(float): The dropout rate of the attention scores. Default:0.1.
hidden_dropout_rate(float): The dropout rate of the final output of the layer. Default:0.1.
post_layernorm_residual(bool): Do residuals adds before the layernorm. Default False.
use_past(bool): Use the past state to compute, used for incremental prediction. Default False.
layernorm_compute_type(dtype.Number): The computation type of the layernorm.
Should be dtype.float32 or dtype.float16. Default dtype.float32.
softmax_compute_type(dtype.Number): The computation type of the softmax in the attention.
Should be dtype.float32 or dtype.float16. Default mstype.float32.
param_init_type(dtype.Number): The parameter initialization type of the module.
Should be dtype.float32 or dtype.float16. Default dtype.float32.
hidden_act(str): The activation of the internal feedforward layer. Supports 'relu',
'relu6', 'tanh', 'gelu', 'fast_gelu', 'elu', 'sigmoid', 'prelu', 'leakyrelu', 'hswish',
'hsigmoid', 'logsigmoid' and so on. Default: gelu.
moe_config(MoEConfig): The configuration of MoE (Mixture of Expert). Default is an instance of MoEConfig
with default values. Please see `MoEConfig`.
parallel_config(OpParallelConfig, MoEParallelConfig): The parallel configure. When MoE is applied,
MoEParallelConfig is effective, otherwise OpParallelConfig is effective. Default `default_dpmp_config`,
an instance of `OpParallelConfig` with default args.
Inputs:
- **hidden_stats** (Tensor) - The input tensor with shape [batch_size, tgt_seq_length, hidden_size] or
[batch_size * tgt_seq_length, hidden_size].
- **decoder_mask** (Tensor) - The attention mask for decoder with shape [batch_size, src_seq_length,
seq_length].
- **encoder_output** (Tensor) - The output of the encoder with shape [batch_size, seq_length, hidden_size]
or [batch_size * seq_length, hidden_size].
Note this args can not be passed by None when the net is in outermost layer. Default None.
- **memory_mask** (Tensor) - The memory mask of the cross attention with shape [batch, tgt_seq_length,
src_seq_length] where tgt_seq_length is the length of the decoder. Note this args can not be passed by
None when the net is in outermost layer. Default None.
- **init_reset** (Tensor) - A bool tensor with shape [1], used to clear the past key parameter and
past value parameter used in the incremental prediction. Only valid when use_past is True. Default True.
- **batch_valid_length** (Tensor) - Int32 tensor with shape [batch_size] the past calculated the index.
Used for incremental prediction when the use_past is True. Default None.
Outputs:
Tuple, a tuple contains(`output`, `layer_present`)
- **output** (Tensor) - The output logit of this layer. The shape is [batch, seq_length, hidden_size] or
[batch * seq_length, hidden_size].
- **layer_present** (Tuple) - A tuple, where each tuple is the tensor of the projected key and value
vector in self attention with shape ((batch_size, num_heads, size_per_head, tgt_seq_length),
(batch_size, num_heads, tgt_seq_length, size_per_head), and of the projected key and value vector
in cross attention with shape (batch_size, num_heads, size_per_head, src_seq_length),
(batch_size, num_heads, src_seq_length, size_per_head)).
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> import numpy as np
>>> from mindspore import dtype as mstype
>>> from mindspore.nn.transformer import TransformerDecoderLayer
>>> from mindspore import Tensor
>>> model = TransformerDecoderLayer(batch_size=2, hidden_size=64, ffn_hidden_size=64, num_heads=2,
... src_seq_length=20, tgt_seq_length=10)
>>> encoder_input_value = Tensor(np.ones((2, 20, 64)), mstype.float32)
>>> decoder_input_value = Tensor(np.ones((2, 10, 64)), mstype.float32)
>>> decoder_input_mask = Tensor(np.ones((2, 10, 10)), mstype.float16)
>>> memory_mask = Tensor(np.ones((2, 10, 20)), mstype.float16)
>>> output, past = model(decoder_input_value, decoder_input_mask, encoder_input_value, memory_mask)
>>> print(output.shape)
(2, 10, 64)
>>> print(past[0].shape)
(2, 2, 32, 10)
>>> print(past[1].shape)
(2, 2, 10, 32)
>>> print(past[2].shape)
(2, 2, 32, 20)
>>> print(past[3].shape)
(2, 2, 20, 32)
"""
@_args_type_validator_check(batch_size=Validator.check_positive_int,
hidden_size=Validator.check_positive_int,
num_heads=Validator.check_positive_int,
ffn_hidden_size=Validator.check_positive_int,
src_seq_length=Validator.check_positive_int,
tgt_seq_length=Validator.check_positive_int,
attention_dropout_rate=Validator.check_non_negative_float,
hidden_dropout_rate=Validator.check_non_negative_float,
hidden_act=_valid_type_checks([str], "TransformerDecoderLayer"),
post_layernorm_residual=Validator.check_bool,
layernorm_compute_type=_valid_value_checks([mstype.float32, mstype.float16],
"TransformerDecoderLayer"),
softmax_compute_type=_valid_value_checks([mstype.float32, mstype.float16],
"TransformerDecoderLayer"),
param_init_type=_valid_value_checks([mstype.float32, mstype.float16],
"TransformerDecoderLayer"),
parallel_config=_valid_type_checks([OpParallelConfig, MoEParallelConfig],
"TransformerDecoderLayer"),
use_past=Validator.check_bool)
def __init__(self, hidden_size,
ffn_hidden_size,
num_heads,
batch_size,
src_seq_length,
tgt_seq_length,
attention_dropout_rate=0.1,
hidden_dropout_rate=0.1,
post_layernorm_residual=False,
use_past=False,
layernorm_compute_type=mstype.float32,
softmax_compute_type=mstype.float32,
param_init_type=mstype.float32,
hidden_act='gelu',
moe_config=default_moe_config,
parallel_config=default_dpmp_config):
super(TransformerDecoderLayer, self).__init__()
if _get_parallel_mode() in (ParallelMode.AUTO_PARALLEL,) and _is_sharding_propagation():
_check_config(parallel_config)
if num_heads % parallel_config.model_parallel != 0:
raise ValueError("For 'TransformerDecoderLayer', the class variable 'num_heads' must be divisibled by "
"'parallel_config.model_parallel', but got the num_heads is {} and "
"parallel_config.model_parallel is {}.".format(num_heads,
parallel_config.model_parallel))
if hidden_size % parallel_config.model_parallel != 0:
raise ValueError(
"For 'TransformerDecoderLayer', the class variable 'hidden_size' must be divisibled by "
"'parallel_config.model_parallel', but got the hidden_size is {} and "
"parallel_config.model_parallel is {}."
.format(hidden_size, parallel_config.model_parallel))
if ffn_hidden_size % parallel_config.model_parallel != 0:
raise ValueError("For 'TransformerDecoderLayer', the class variable 'ffn_hidden_size' must be "
"divisibled by 'parallel_config.model_parallel', but got the ffn_hidden_size is {} "
"and parallel_config.model_parallel is {}."
.format(ffn_hidden_size, parallel_config.model_parallel))
_check_moe_config(moe_config, parallel_config)
self.use_moe = (moe_config.expert_num > 1)
if use_past:
raise ValueError(f"The {self.cls_name} does not support use_past=True.")
self.batch_size = batch_size
self.use_past = use_past
self.softmax_compute_type = softmax_compute_type
self.src_seq_length = src_seq_length
self.tgt_seq_length = tgt_seq_length
self.use_past = use_past
self.hidden_size = hidden_size
self.layernorm1 = _LayerNorm((hidden_size,)).to_float(layernorm_compute_type)
self.layernorm2 = _LayerNorm((hidden_size,)).to_float(layernorm_compute_type)
self.attention = MultiHeadAttention(hidden_size=hidden_size,
num_heads=num_heads,
batch_size=batch_size,
src_seq_length=tgt_seq_length,
tgt_seq_length=tgt_seq_length,
hidden_dropout_rate=hidden_dropout_rate,
attention_dropout_rate=attention_dropout_rate,
use_past=use_past,
softmax_compute_type=softmax_compute_type,
param_init_type=param_init_type,
parallel_config=parallel_config.dpmp if self.use_moe
else parallel_config)
# Cross attention with the output of encoder as memory tensor
self.cross_attention = MultiHeadAttention(hidden_size=hidden_size,
num_heads=num_heads,
batch_size=batch_size,
src_seq_length=tgt_seq_length,
tgt_seq_length=src_seq_length,
hidden_dropout_rate=hidden_dropout_rate,
attention_dropout_rate=attention_dropout_rate,
softmax_compute_type=softmax_compute_type,
use_past=use_past,
param_init_type=param_init_type,
parallel_config=parallel_config.dpmp
if self.use_moe else parallel_config)
self.cross_attention_layernorm = _LayerNorm((hidden_size,)).to_float(
layernorm_compute_type)
if self.use_moe:
self.output = MoE(hidden_size=hidden_size,
dropout_rate=hidden_dropout_rate,
ffn_hidden_size=ffn_hidden_size,
param_init_type=param_init_type,
hidden_act=hidden_act,
moe_config=moe_config,
parallel_config=parallel_config)
else:
# Feed Forward Network, FFN
self.output = FeedForward(hidden_size=hidden_size,
dropout_rate=hidden_dropout_rate,
ffn_hidden_size=ffn_hidden_size,
hidden_act=hidden_act,
param_init_type=param_init_type,
parallel_config=parallel_config)
self.post_layernorm_residual = post_layernorm_residual
self.add = P.Add()
self.add_3d = P.Add()
self.dtype = mstype.float16
self.key_past = None
self.value_past = None
if self.use_past:
# operator used for state reuse
self.reducesum = P.ReduceSum().shard(((1, 1, 1, 1),))
self.not_equal = P.NotEqual().shard(((1, 1, 1, 1), ()))
self.slice = P.StridedSlice().shard(((1, 1, 1, 1),))
size_per_head = int(hidden_size / num_heads)
self.key_shape = (batch_size, num_heads, size_per_head, tgt_seq_length)
self.value_shape = (batch_size, num_heads, tgt_seq_length, size_per_head)
# parameters saving key and value states
self.key_past = Parameter(Tensor(np.zeros(shape=self.key_shape), self.dtype), name="key_past")
self.value_past = Parameter(Tensor(np.zeros(shape=self.value_shape), self.dtype), name="value_past")
self.tile = P.Tile().shard(((1, 1),))
self.mul = P.Mul().shard(((1, 1, 1, 1), (1,)))
self.assign = P.Assign().shard(((1, 1, 1, 1), (1, 1, 1, 1)))
elif _get_parallel_mode() not in (ParallelMode.AUTO_PARALLEL,):
_check_config(parallel_config)
if num_heads % parallel_config.model_parallel != 0:
raise ValueError("For 'TransformerDecoderLayer', the class variable 'num_heads' must be divisibled by "
"'parallel_config.model_parallel', but got the num_heads is {} and "
"parallel_config.model_parallel is {}.".format(num_heads,
parallel_config.model_parallel))
if hidden_size % parallel_config.model_parallel != 0:
raise ValueError(
"For 'TransformerDecoderLayer', the class variable 'hidden_size' must be divisibled by "
"'parallel_config.model_parallel', but got the hidden_size is {} and "
"parallel_config.model_parallel is {}."
.format(hidden_size, parallel_config.model_parallel))
if ffn_hidden_size % parallel_config.model_parallel != 0:
raise ValueError("For 'TransformerDecoderLayer', the class variable 'ffn_hidden_size' must be "
"divisibled by 'parallel_config.model_parallel', but got the ffn_hidden_size is {} "
"and parallel_config.model_parallel is {}."
.format(ffn_hidden_size, parallel_config.model_parallel))
_check_moe_config(moe_config, parallel_config)
self.use_moe = (moe_config.expert_num > 1)
if use_past:
raise ValueError(f"The {self.cls_name} does not support use_past=True.")
self.batch_size = batch_size
self.use_past = use_past
self.softmax_compute_type = softmax_compute_type
self.src_seq_length = src_seq_length
self.tgt_seq_length = tgt_seq_length
self.use_past = use_past
self.hidden_size = hidden_size
self.layernorm1 = _LayerNorm((hidden_size,)).to_float(layernorm_compute_type)
self.layernorm1.shard(((parallel_config.data_parallel, 1),))
self.layernorm2 = _LayerNorm((hidden_size,)).to_float(layernorm_compute_type)
self.layernorm2.shard(((parallel_config.data_parallel, 1),))
self.attention = MultiHeadAttention(hidden_size=hidden_size,
num_heads=num_heads,
batch_size=batch_size,
src_seq_length=tgt_seq_length,
tgt_seq_length=tgt_seq_length,
hidden_dropout_rate=hidden_dropout_rate,
attention_dropout_rate=attention_dropout_rate,
use_past=use_past,
softmax_compute_type=softmax_compute_type,
param_init_type=param_init_type,
parallel_config=parallel_config.dpmp if self.use_moe
else parallel_config)
# Cross attention with the output of encoder as memory tensor
self.cross_attention = MultiHeadAttention(hidden_size=hidden_size,
num_heads=num_heads,
batch_size=batch_size,
src_seq_length=tgt_seq_length,
tgt_seq_length=src_seq_length,
hidden_dropout_rate=hidden_dropout_rate,
attention_dropout_rate=attention_dropout_rate,
softmax_compute_type=softmax_compute_type,
use_past=use_past,
param_init_type=param_init_type,
parallel_config=parallel_config.dpmp
if self.use_moe else parallel_config)
self.cross_attention_layernorm = _LayerNorm((hidden_size,)).to_float(
layernorm_compute_type)
self.cross_attention_layernorm.shard(((parallel_config.data_parallel, 1),))
if self.use_moe:
self.output = MoE(hidden_size=hidden_size,
dropout_rate=hidden_dropout_rate,
ffn_hidden_size=ffn_hidden_size,
param_init_type=param_init_type,
hidden_act=hidden_act,
moe_config=moe_config,
parallel_config=parallel_config)
else:
# Feed Forward Network, FFN
self.output = FeedForward(hidden_size=hidden_size,
dropout_rate=hidden_dropout_rate,
ffn_hidden_size=ffn_hidden_size,
hidden_act=hidden_act,
param_init_type=param_init_type,
parallel_config=parallel_config)
self.post_layernorm_residual = post_layernorm_residual
self.add = P.Add().shard(((parallel_config.data_parallel, 1), (parallel_config.data_parallel, 1)))
self.add_3d = P.Add().shard(((parallel_config.data_parallel, 1, 1), (parallel_config.data_parallel, 1, 1)))
self.dtype = mstype.float16
self.key_past = None
self.value_past = None
if self.use_past:
# operator used for state reuse
self.reducesum = P.ReduceSum().shard(((1, 1, 1, 1),))
self.not_equal = P.NotEqual().shard(((1, 1, 1, 1), ()))
self.slice = P.StridedSlice().shard(((1, 1, 1, 1),))
size_per_head = int(hidden_size / num_heads)
self.key_shape = (batch_size, num_heads, size_per_head, tgt_seq_length)
self.value_shape = (batch_size, num_heads, tgt_seq_length, size_per_head)
# parameters saving key and value states
self.key_past = Parameter(Tensor(np.zeros(shape=self.key_shape), self.dtype), name="key_past")
self.value_past = Parameter(Tensor(np.zeros(shape=self.value_shape), self.dtype), name="value_past")
self.tile = P.Tile().shard(((1, 1),))
self.mul = P.Mul().shard(((1, 1, 1, 1), (1,)))
self.assign = P.Assign().shard(((1, 1, 1, 1), (1, 1, 1, 1)))
else:
raise RuntimeError(f"The {self.cls_name} only support sharding propagation or "
f"semi-auto parallel mode now.")
def construct(self, hidden_stats,
decoder_mask,
encoder_output=None,
memory_mask=None,
init_reset=True, batch_valid_length=None):
self._check_input(hidden_stats, decoder_mask, encoder_output, memory_mask, init_reset, batch_valid_length)
# the returned shape is [bs, seq_length, embedding_size] or [bs * seq_length, embedding_size]
hidden_shape = F.shape(hidden_stats)
hidden_stats = F.reshape(hidden_stats, (-1, hidden_shape[-1]))
input_x = self.layernorm1(hidden_stats)
input_x = F.cast(input_x, self.dtype)
# indicate whether reset saved states
key_reset = None
value_reset = None
if self.use_past:
# reset states, init_reset True for reuse and False for reset
key_reset = self.assign(self.key_past, self.mul(self.key_past, F.cast(init_reset, self.dtype)))
value_reset = self.assign(self.value_past, self.mul(self.value_past, F.cast(init_reset, self.dtype)))
# add dependency for desired execution order
input_x = F.depend(input_x, key_reset)
input_x = F.depend(input_x, value_reset)
attention, layer_present = self.attention(input_x, input_x, input_x, decoder_mask, self.key_past,
self.value_past, batch_valid_length)
# For post-layernorm the inputs for residual path are output of self-attention and output of layernorm
if self.post_layernorm_residual:
x = self.add(input_x, attention)
# For pre-layernorm the inputs for residual path are output of self-attention and input of this layer
else:
x = self.add(hidden_stats, attention)
middle_output = None
if encoder_output is not None:
middle_output = self.cross_attention_layernorm(x)
middle_output = F.cast(middle_output, self.dtype)
cross_attn_output, cross_layer_present = self.cross_attention(middle_output, encoder_output,
encoder_output,
memory_mask, self.key_past,
self.value_past, batch_valid_length)
layer_present += cross_layer_present
if self.post_layernorm_residual:
x = self.add(middle_output, cross_attn_output)
else:
x = self.add(x, cross_attn_output)
output_x = self.layernorm2(x)
output_x = F.cast(output_x, self.dtype)
aux_loss = None
if self.use_moe:
mlp_logit, aux_loss = self.output(output_x)
else:
mlp_logit = self.output(output_x)
value_update = None
key_update = None
if self.use_past:
# current key and value
key_present, value_present = layer_present
# update key and value calculated this step
key_update = self.assign(self.key_past, key_present)
value_update = self.assign(self.value_past, value_present)
# add dependency for desired execution order
key_update = F.depend(key_update, key_reset)
value_update = F.depend(value_update, value_reset)
# add dependency for desired execution order
mlp_logit = F.depend(mlp_logit, value_update)
mlp_logit = F.depend(mlp_logit, key_update)
# if shape is 3d, we reshape the inputs of the add
if len(hidden_shape) == 3:
output_x = P.Reshape()(output_x, hidden_shape)
mlp_logit = P.Reshape()(mlp_logit, hidden_shape)
x = P.Reshape()(x, hidden_shape)
if self.post_layernorm_residual:
output = self.add_3d(output_x, mlp_logit)
else:
output = self.add_3d(x, mlp_logit)
else:
if self.post_layernorm_residual:
output = self.add(output_x, mlp_logit)
else:
output = self.add(x, mlp_logit)
output = F.reshape(output, hidden_shape)
if self.use_moe:
return output, layer_present, aux_loss
return output, layer_present
def _check_input(self, hidden_states, attention_mask, encoder_output, memory_mask, init_reset, batch_valid_length):
r"""Check inputs"""
if not self.use_past or (self.use_past and self.is_first_iteration):
_check_shape_equal(F.shape(hidden_states), "hidden_states", self.cls_name,
[[self.batch_size, self.tgt_seq_length, self.hidden_size],
[self.batch_size * self.tgt_seq_length, self.hidden_size]])
_check_shape_equal(F.shape(attention_mask), "attention_mask", self.cls_name,
[self.batch_size, self.tgt_seq_length, self.tgt_seq_length])
else:
_check_shape_equal(F.shape(hidden_states), "hidden_states", self.cls_name,
[self.batch_size, 1, self.hidden_size])
_check_shape_equal(F.shape(attention_mask), "attention_mask", self.cls_name,
[self.batch_size, 1, self.tgt_seq_length])
_check_input_dtype(F.dtype(hidden_states), "hidden_states", [mstype.float32, mstype.float16], self.cls_name)
_check_input_dtype(F.dtype(attention_mask), "attention_mask", [mstype.float32, mstype.float16], self.cls_name)
if encoder_output is not None:
_check_shape_equal(F.shape(encoder_output), "encoder_output", self.cls_name,
[[self.batch_size, self.src_seq_length, self.hidden_size],
[self.batch_size * self.src_seq_length, self.hidden_size]])
_check_input_dtype(F.dtype(encoder_output), "encoder_output",
[mstype.float32, mstype.float16], self.cls_name)
if memory_mask is not None:
_check_shape_equal(F.shape(memory_mask), "memory_mask", self.cls_name,
[self.batch_size, self.tgt_seq_length, self.src_seq_length])
_check_input_dtype(F.dtype(memory_mask), "memory_mask",
[mstype.float32, mstype.float16], self.cls_name)
init_reset_is_tensor = isinstance(init_reset, Tensor)
init_reset_is_default = init_reset is True
batch_valid_length_is_tensor = isinstance(batch_valid_length, Tensor)
batch_is_default = batch_valid_length is None
_check_past_none_input_none(self.use_past, "init_reset", self.cls_name, True, init_reset_is_tensor,
init_reset_is_default)
_check_past_none_input_none(self.use_past, "batch_valid_length", self.cls_name, None,
batch_valid_length_is_tensor, batch_is_default)
if self.use_past:
_check_shape_equal(F.shape(init_reset), "init_reset", self.cls_name, [1])
_check_input_dtype(F.dtype(init_reset), "init_reset", [mstype.bool_], self.cls_name)
_check_shape_equal(F.shape(batch_valid_length), "batch_valid_length", self.cls_name, [self.batch_size])
_check_input_dtype(F.dtype(batch_valid_length), "batch_valid_length", [mstype.int32], self.cls_name)
return True
def _get_lambda_func(total_layer=None):
r"""
A wrapper function of specifying pipeline stage and gradient aggregation fusion. If the total layer
is not None, for example, set in the transformer model, the pipeline stage setting function will be
`(layer_id + 0) // (total_layers / parallel_config.pipeline_stage)` for the encoder and,
`(layer_id + offset) //
(total_layers / parallel_config.pipeline_stage)` for the decoder, where `offset` is the layers in the encoder.
"""
def _set_parallel_configure_for_layer(network, layer_id, offset, parallel_config, layers):
r"""
Default setting for the pipeline is: `(layer_id + offset) // (layers / pipeline_stage)`.
Args:
network(Cell) - Represents the transformer block
layer_id(int) - Means the layer index for the current module, counts from zero.
offset(int) - Means the layer_index needs an offset, if there are other modules in the net.
layers(int) - The total layers used for the model.
"""
# override the layers
if total_layer:
layers = total_layer
# Used for the pipeline's stages setting
if layers < parallel_config.pipeline_stage:
raise ValueError(f"layers {layers} must be larger than pipeline stage {parallel_config.pipeline_stage}")
pp_dis = max(int(layers / parallel_config.pipeline_stage), 1)
# the pipeline stage must be in [0, parallel_config.pipeline_stage - 1]
pp_id = min((layer_id + offset) // pp_dis, parallel_config.pipeline_stage - 1)
network.pipeline_stage = pp_id
logger.info(f"pipeline stage id is {pp_id}")
# Used for optimizer's fusion tag
dis = max(int(layers / parallel_config.gradient_aggregation_group), 1)
network.set_comm_fusion(int((layer_id + offset) / dis) + 1)
# Used for enabling recomputation of the block
if isinstance(parallel_config.recompute, bool):
if parallel_config.recompute:
network.recompute()
else:
if parallel_config.recompute.recompute:
paralel_op_comm_compute = parallel_config.recompute.parallel_optimizer_comm_recompute
network.recompute(parallel_optimizer_comm_recompute=paralel_op_comm_compute,
mp_comm_recompute=parallel_config.recompute.mp_comm_recompute,
recompute_slice_activation=parallel_config.recompute.recompute_slice_activation)
return _set_parallel_configure_for_layer
[文档]class TransformerEncoder(Cell):
r"""
Transformer Encoder module with multi-layer stacked of `TransformerEncoderLayer`, including multihead self
attention and feedforward layer.
Args:
batch_size(int): The batch size of the input tensor.
num_layers(int): The layers of the `TransformerEncoderLayer`
hidden_size(int): The hidden size of the input.
ffn_hidden_size(int): The hidden size of bottleneck in the feedforward layer.
seq_length(int): The seq_length of the input tensor.
num_heads(int): The number of the heads.
attention_dropout_rate(float): The dropout rate of the attention scores. Default:0.1.
hidden_dropout_rate(float): The dropout rate of the final output of the layer. Default:0.1.
hidden_act(str): The activation of the internal feedforward layer. Supports 'relu',
'relu6', 'tanh', 'gelu', 'fast_gelu', 'elu', 'sigmoid', 'prelu', 'leakyrelu', 'hswish',
'hsigmoid', 'logsigmoid' and so on. Default: gelu.
post_layernorm_residual(bool): Do residuals adds before the layernorm. Default False.
layernorm_compute_type(dtype.Number): The computation type of the layernorm.
Should be dtype.float32 or dtype.float16. Default dtype.float32.
softmax_compute_type(dtype.Number): The computation type of the softmax in the attention.
Should be dtype.float32 or dtype.float16. Default mstype.float32.
param_init_type(dtype.Number): The parameter initialization type of the module.
Should be dtype.float32 or dtype.float16. Default dtype.float32.
lambda_func: A function can determine the fusion index, pipeline stages and recompute attribute. If the
user wants to determine the pipeline stage and gradient aggregation fusion, the user can pass a
function that accepts `network`, `layer_id`, `offset`, `parallel_config`, `layers`. The `network(Cell)`
represents the transformer block, `layer_id(int)` means the layer index for the current module, counts
from zero, `offset(int)` means the layer_index needs an offset, if there are other modules in the net.
The default setting for the pipeline is: `(layer_id + offset) // (layers / pipeline_stage)`.
Default None.
offset(int): The initial layer index for the `encoder`. Used for setting the fusion id and stage id, to not
overlap with the encoder layer. Default 0.
use_past(bool): Use the past state to compute, used for incremental prediction. For example, if we have two
words and want to generate the ten more words. We just need to compute the two words' state only once,
and generate the next word one by one. When use_past is True, there are two steps to run the prediction.
In the first step, set the is_first_iteration to be True by
`model.add_flags_recursive(is_first_iteration=True)`, and pass the full inputs. Then, set the
is_first_iteration to be False by `model.add_flags_recursive(is_first_iteration=False)`. At this moment,
pass the single step's input tensor, and loop it. Default False.
moe_config(MoEConfig): The configuration of MoE (Mixture of Expert). Default is an instance of MoEConfig
with default values. Please see `MoEConfig`.
parallel_config(TransformerOpParallelConfig): The parallel configure. Default `default_transformer_config`,
an instance of `TransformerOpParallelConfig` with default args.
Inputs:
- **hidden_states** (Tensor) - Tensor, shape should be [batch_size, seq_length, hidden_size] or
[batch_size * seq_length, hidden_size], if the use_past is False or is_first_iteration=True. Otherwise,
should be [batch_size, 1, hidden_size].
- **attention_mask** (Tensor) - Tensor, attention mask with shape [batch_size, seq_length, seq_length]
- **init_reset** (Tensor) - A bool tensor with shape [1], used to clear the past key parameter and
past value parameter used in the incremental prediction. Only valid when use_past is True. Default True.
- **batch_valid_length** (Tensor) - Int32 tensor with shape [batch_size] the past calculated the index.
Used for incremental prediction when the use_past is True. Default None.
Outputs:
Tuple, a tuple contains(`output`, `layer_present`)
- **output** (Tensor) - The float tensor of the output of the layer with
shape (batch_size, seq_length, hidden_size) or (batch_size * seq_length, hidden_size), if the use_past is
False or is_first_iteration=True. Otherwise, it will be (batch_size, 1, hidden_size).
- **layer_present** (Tuple) - A tuple with size of num_layers, where each tuple contains the Tensor the
projected key and value vector with shape ((batch_size, num_heads, size_per_head, seq_length),
and (batch_size, num_heads, seq_length, size_per_head)).
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> import numpy as np
>>> from mindspore import dtype as mstype
>>> from mindspore.nn.transformer import TransformerEncoder
>>> from mindspore import Tensor
>>> model = TransformerEncoder(batch_size=2, num_layers=2, hidden_size=8, ffn_hidden_size=64,
... seq_length=16, num_heads=2)
>>> encoder_input_value = Tensor(np.ones((2, 16, 8)), mstype.float32)
>>> encoder_input_mask = Tensor(np.ones((2, 16, 16)), mstype.float16)
>>> output, past = model(encoder_input_value, encoder_input_mask)
>>> print(output.shape)
(2, 16, 8)
>>> print(len(past))
2
>>> print(past[0][0].shape)
(2, 2, 4, 16)
>>> print(past[0][1].shape)
(2, 2, 16, 4)
>>> # When use use_past=True, it includes two steps to implement the incremental prediction.
>>> # Step 1: set is_first_iteration=True, and input the full sequence length's state.
>>> batch_valid_length = Tensor(np.ones((2,)), mstype.int32)
>>> init_reset = Tensor([True], mstype.bool_)
>>> # Set is_first_iteration=True to generate the full memory states
>>> model = TransformerEncoder(batch_size=2, hidden_size=8, ffn_hidden_size=64, seq_length=16,
... num_heads=2, num_layers=2, use_past=True)
>>> model.add_flags_recursive(is_first_iteration=True)
>>> hidden, past = model(encoder_input_value, encoder_input_mask, init_reset, batch_valid_length)
>>> print(hidden.shape)
(2, 16, 8)
>>> print(past[0][0].shape)
(2, 2, 4, 16)
>>> print(past[0][1].shape)
(2, 2, 16, 4)
>>> encoder_input_value = Tensor(np.ones((2, 1, 8)), mstype.float32)
>>> encoder_input_mask = Tensor(np.ones((2, 1, 16)), mstype.float16)
>>> init_reset = Tensor([False], mstype.bool_)
>>> # Step 2: set is_first_iteration=False, and pass the single word to run the prediction rather than
>>> # the full sequence.
>>> model.add_flags_recursive(is_first_iteration=False)
>>> hidden, past = model(encoder_input_value, encoder_input_mask, init_reset, batch_valid_length)
>>> print(hidden.shape)
(2, 1, 8)
>>> print(past[0][0].shape)
(2, 2, 4, 16)
>>> print(past[0][1].shape)
(2, 2, 16, 4)
"""
@_args_type_validator_check(batch_size=Validator.check_positive_int,
hidden_size=Validator.check_positive_int,
num_heads=Validator.check_positive_int,
ffn_hidden_size=Validator.check_positive_int,
seq_length=Validator.check_positive_int,
num_layers=Validator.check_positive_int,
offset=Validator.check_non_negative_int,
attention_dropout_rate=Validator.check_non_negative_float,
hidden_dropout_rate=Validator.check_non_negative_float,
hidden_act=_valid_type_checks([str], "TransformerEncoder"),
post_layernorm_residual=Validator.check_bool,
layernorm_compute_type=_valid_value_checks([mstype.float32, mstype.float16],
"TransformerEncoder"),
softmax_compute_type=_valid_value_checks([mstype.float32, mstype.float16],
"TransformerEncoder"),
param_init_type=_valid_value_checks([mstype.float32, mstype.float16],
"TransformerEncoder"),
parallel_config=_valid_type_checks([TransformerOpParallelConfig],
"TransformerEncoder"),
use_past=Validator.check_bool)
def __init__(self,
batch_size,
num_layers,
hidden_size,
ffn_hidden_size,
seq_length,
num_heads,
attention_dropout_rate=0.1,
hidden_dropout_rate=0.1,
hidden_act='gelu',
post_layernorm_residual=False,
layernorm_compute_type=mstype.float32,
softmax_compute_type=mstype.float32,
param_init_type=mstype.float32,
lambda_func=None,
offset=0,
use_past=False,
moe_config=default_moe_config,
parallel_config=default_transformer_config):
super(TransformerEncoder, self).__init__()
if _get_parallel_mode() in (ParallelMode.AUTO_PARALLEL,) and _is_sharding_propagation():
_check_config(parallel_config)
_check_moe_config(moe_config, parallel_config)
self.use_moe = (moe_config.expert_num > 1)
self.add = P.Add()
self.aux_loss = Tensor(0.0, mstype.float32)
self.num_layers = num_layers
self.blocks = nn.CellList()
for i in range(num_layers):
block = TransformerEncoderLayer(hidden_size=hidden_size,
batch_size=batch_size,
ffn_hidden_size=ffn_hidden_size,
seq_length=seq_length,
attention_dropout_rate=attention_dropout_rate,
hidden_dropout_rate=hidden_dropout_rate,
layernorm_compute_type=layernorm_compute_type,
softmax_compute_type=softmax_compute_type,
num_heads=num_heads,
hidden_act=hidden_act,
post_layernorm_residual=post_layernorm_residual,
param_init_type=param_init_type,
use_past=use_past,
moe_config=moe_config,
parallel_config=parallel_config.moe_parallel_config if self.use_moe
else parallel_config.dp_mp_config)
# If the user doesn't pass the fusion function, use the default one
if not lambda_func:
lambda_func = _get_lambda_func()
lambda_func(block, layer_id=i, layers=num_layers,
offset=offset, parallel_config=parallel_config)
self.blocks.append(block)
elif _get_parallel_mode() not in (ParallelMode.AUTO_PARALLEL,):
_check_config(parallel_config)
_check_moe_config(moe_config, parallel_config)
self.use_moe = (moe_config.expert_num > 1)
self.add = P.Add().shard(((), ()))
self.aux_loss = Tensor(0.0, mstype.float32)
logger.warning("For parallel mode, sharding propagation is recommended, you can use it by setting "
"'set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL, "
"search_mode=\"sharding_propagation\")' and "
"'set_algo_parameters(elementwise_op_strategy_follow=False, fully_use_devices=False)'")
self.num_layers = num_layers
self.blocks = nn.CellList()
for i in range(num_layers):
block = TransformerEncoderLayer(hidden_size=hidden_size,
batch_size=batch_size,
ffn_hidden_size=ffn_hidden_size,
seq_length=seq_length,
attention_dropout_rate=attention_dropout_rate,
hidden_dropout_rate=hidden_dropout_rate,
layernorm_compute_type=layernorm_compute_type,
softmax_compute_type=softmax_compute_type,
num_heads=num_heads,
hidden_act=hidden_act,
post_layernorm_residual=post_layernorm_residual,
param_init_type=param_init_type,
use_past=use_past,
moe_config=moe_config,
parallel_config=parallel_config.moe_parallel_config if self.use_moe
else parallel_config.dp_mp_config)
# If the user doesn't pass the fusion function, use the default one
if not lambda_func:
lambda_func = _get_lambda_func()
lambda_func(block, layer_id=i, layers=num_layers,
offset=offset, parallel_config=parallel_config)
self.blocks.append(block)
else:
raise RuntimeError(f"The {self.cls_name} only support sharding propagation or "
f"semi-auto parallel mode now.")
def construct(self, hidden_states, attention_mask, init_reset=True, batch_valid_length=None):
present_layer = ()
if self.use_moe:
accum_loss = self.aux_loss
for i in range(self.num_layers):
hidden_states, present, aux_loss = self.blocks[i](hidden_states,
attention_mask,
init_reset,
batch_valid_length)
present_layer = present_layer + (present,)
accum_loss = self.add(accum_loss, aux_loss)
return hidden_states, present_layer, accum_loss
for i in range(self.num_layers):
hidden_states, present = self.blocks[i](hidden_states,
attention_mask,
init_reset,
batch_valid_length)
present_layer = present_layer + (present,)
return hidden_states, present_layer
[文档]class TransformerDecoder(Cell):
r"""
Transformer Decoder module with multi-layer stacked of `TransformerDecoderLayer`, including multihead self
attention, cross attention and feedforward layer.
Args:
num_layers(int): The layers of the `TransformerDecoderLayer`.
batch_size(int): The batch size of the input tensor.
hidden_size(int): The hidden size of the input.
ffn_hidden_size(int): The hidden size of bottleneck in the feedforward layer.
src_seq_length(int): The input source sequence length.
tgt_seq_length(int): The input target sequence length.
num_heads(int): The number of the heads.
attention_dropout_rate(float): The dropout rate of the attention scores. Default:0.1.
hidden_dropout_rate(float): The dropout rate of the final output of the layer. Default:0.1.
post_layernorm_residual(bool): Do residuals adds before the layernorm. Default False.
layernorm_compute_type(dtype.Number): The computation type of the layernorm.
Should be dtype.float32 or dtype.float16. Default dtype.float32.
softmax_compute_type(dtype.Number): The computation type of the softmax in the attention.
Should be dtype.float32 or dtype.float16. Default mstype.float32.
param_init_type(dtype.Number): The parameter initialization type of the module.
Should be dtype.float32 or dtype.float16. Default dtype.float32.
hidden_act(str): The activation of the internal feedforward layer. Supports 'relu',
'relu6', 'tanh', 'gelu', 'fast_gelu', 'elu', 'sigmoid', 'prelu', 'leakyrelu', 'hswish',
'hsigmoid', 'logsigmoid' and so on. Default: gelu.
lambda_func: A function can determine the fusion index, pipeline stages and recompute attribute. If the
user wants to determine the pipeline stage and gradient aggregation fusion, the user can pass a
function that accepts `network`, `layer_id`, `offset`, `parallel_config`, `layers`. The `network(Cell)`
represents the transformer block, `layer_id(int)` means the layer index for the current module, counts
from zero, `offset(int)` means the layer_index needs an offset, if there are other modules in the net.
The default setting for the pipeline is: `(layer_id + offset) // (layers / pipeline_stage)`.
Default: None.
use_past(bool): Use the past state to compute, used for incremental prediction. Default False.
offset(int): The initial layer index for the `decoder`. Used for setting the fusion id and stage id, to not
overlap with the encoder layer. Default 0.
moe_config(MoEConfig): The configuration of MoE (Mixture of Expert). Default is an instance of MoEConfig
with default values. Please see `MoEConfig`.
parallel_config(TransformerOpParallelConfig): The parallel configure. Default `default_transformer_config`,
an instance of `TransformerOpParallelConfig` with default args.
Inputs:
- **hidden_stats** (Tensor) - The input tensor with shape [batch_size, seq_length, hidden_size] or
[batch_size * seq_length, hidden_size]
- **attention_mask** (Tensor) - The attention mask for decoder with shape
[batch_size, seq_length, seq_length]
- **encoder_output** (Tensor) - The output of the encoder with shape [batch_size, seq_length, hidden_size]
or [batch_size * seq_length, hidden_size]. Note this args can not be passed by None when the net is in
outermost layer. Default None.
- **memory_mask** (Tensor) - The memory mask of the cross attention with shape [batch, tgt_seq_length,
src_seq_length] where tgt_seq_length is the length of the decoder. Note this args can not be passed by
None when the net is in outermost layer. Default None.
- **init_reset** (Tensor) - A bool tensor with shape [1], used to clear the past key parameter and
past value parameter used in the incremental prediction. Only valid when use_past is True. Default True.
- **batch_valid_length** (Tensor) - Int32 tensor with shape [batch_size] the past calculated the index.
Used for incremental prediction when the use_past is True. Default None.
Outputs:
Tuple, a tuple contains(`output`, `layer_present`)
- **output** (Tensor) - The output logit of this layer. The shape is [batch, tgt_seq_length, hidden_size] or
[batch * tgt_seq_length, hidden_size]
- **layer_present** (Tuple) - A tuple with size of num_layers, where each tuple is the tensor of the
projected key and value vector in self attention with shape ((batch_size, num_heads, size_per_head,
tgt_seq_length), (batch_size, num_heads, tgt_seq_length, size_per_head), and of the projected key
and value vector in cross attention with shape (batch_size, num_heads, size_per_head, src_seq_length),
(batch_size, num_heads, src_seq_length, size_per_head)).
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> import numpy as np
>>> from mindspore import dtype as mstype
>>> from mindspore.nn.transformer import TransformerDecoder
>>> from mindspore import Tensor
>>> model = TransformerDecoder(batch_size=2, num_layers=1, hidden_size=64, ffn_hidden_size=64,
... num_heads=2, src_seq_length=20, tgt_seq_length=10)
>>> encoder_input_value = Tensor(np.ones((2, 20, 64)), mstype.float32)
>>> decoder_input_value = Tensor(np.ones((2, 10, 64)), mstype.float32)
>>> decoder_input_mask = Tensor(np.ones((2, 10, 10)), mstype.float16)
>>> memory_mask = Tensor(np.ones((2, 10, 20)), mstype.float16)
>>> output, past = model(decoder_input_value, decoder_input_mask, encoder_input_value, memory_mask)
>>> print(output.shape)
(2, 10, 64)
>>> print(len(past))
1
>>> print(past[0][0].shape)
(2, 2, 32, 10)
>>> print(past[0][1].shape)
(2, 2, 10, 32)
>>> print(past[0][2].shape)
(2, 2, 32, 20)
>>> print(past[0][3].shape)
(2, 2, 20, 32)
"""
@_args_type_validator_check(batch_size=Validator.check_positive_int,
hidden_size=Validator.check_positive_int,
num_heads=Validator.check_positive_int,
ffn_hidden_size=Validator.check_positive_int,
src_seq_length=Validator.check_positive_int,
num_layers=Validator.check_positive_int,
tgt_seq_length=Validator.check_positive_int,
offset=Validator.check_non_negative_int,
attention_dropout_rate=Validator.check_non_negative_float,
hidden_dropout_rate=Validator.check_non_negative_float,
hidden_act=_valid_type_checks([str], "TransformerDecoder"),
post_layernorm_residual=Validator.check_bool,
layernorm_compute_type=_valid_value_checks([mstype.float32, mstype.float16],
"TransformerDecoder"),
softmax_compute_type=_valid_value_checks([mstype.float32, mstype.float16],
"TransformerDecoder"),
param_init_type=_valid_value_checks([mstype.float32, mstype.float16],
"TransformerDecoder"),
parallel_config=_valid_type_checks([TransformerOpParallelConfig],
"TransformerDecoder"),
use_past=Validator.check_bool)
def __init__(self,
num_layers,
batch_size,
hidden_size,
ffn_hidden_size,
src_seq_length,
tgt_seq_length,
num_heads,
attention_dropout_rate=0.1,
hidden_dropout_rate=0.1,
post_layernorm_residual=False,
layernorm_compute_type=mstype.float32,
softmax_compute_type=mstype.float32,
param_init_type=mstype.float32,
hidden_act='gelu',
lambda_func=None,
use_past=False,
offset=0,
moe_config=default_moe_config,
parallel_config=default_transformer_config):
super(TransformerDecoder, self).__init__()
if _get_parallel_mode() in (ParallelMode.AUTO_PARALLEL,) and _is_sharding_propagation():
_check_config(parallel_config)
self.add = P.Add()
self.aux_loss = Tensor(0.0, mstype.float32)
self.num_layers = num_layers
self.blocks = nn.CellList()
_check_moe_config(moe_config, parallel_config)
self.use_moe = (moe_config.expert_num > 1)
for i in range(num_layers):
block = TransformerDecoderLayer(hidden_size=hidden_size,
batch_size=batch_size,
ffn_hidden_size=ffn_hidden_size,
src_seq_length=src_seq_length,
tgt_seq_length=tgt_seq_length,
attention_dropout_rate=attention_dropout_rate,
hidden_dropout_rate=hidden_dropout_rate,
num_heads=num_heads,
layernorm_compute_type=layernorm_compute_type,
softmax_compute_type=softmax_compute_type,
hidden_act=hidden_act,
use_past=use_past,
param_init_type=param_init_type,
post_layernorm_residual=post_layernorm_residual,
moe_config=moe_config,
parallel_config=parallel_config.moe_parallel_config if self.use_moe
else parallel_config.dp_mp_config)
# If the user doesn't pass the fusion function, use the default one
if not lambda_func:
lambda_func = _get_lambda_func()
lambda_func(block, layer_id=i, layers=num_layers,
offset=offset, parallel_config=parallel_config)
self.blocks.append(block)
elif _get_parallel_mode() not in (ParallelMode.AUTO_PARALLEL,):
_check_config(parallel_config)
self.add = P.Add().shard(((), ()))
self.aux_loss = Tensor(0.0, mstype.float32)
logger.warning("For parallel mode, sharding propagation is recommended, you can use it by setting "
"'set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL, "
"search_mode=\"sharding_propagation\")' and "
"'set_algo_parameters(elementwise_op_strategy_follow=False, fully_use_devices=False)'")
self.num_layers = num_layers
self.blocks = nn.CellList()
_check_moe_config(moe_config, parallel_config)
self.use_moe = (moe_config.expert_num > 1)
for i in range(num_layers):
block = TransformerDecoderLayer(hidden_size=hidden_size,
batch_size=batch_size,
ffn_hidden_size=ffn_hidden_size,
src_seq_length=src_seq_length,
tgt_seq_length=tgt_seq_length,
attention_dropout_rate=attention_dropout_rate,
hidden_dropout_rate=hidden_dropout_rate,
num_heads=num_heads,
layernorm_compute_type=layernorm_compute_type,
softmax_compute_type=softmax_compute_type,
hidden_act=hidden_act,
use_past=use_past,
param_init_type=param_init_type,
post_layernorm_residual=post_layernorm_residual,
moe_config=moe_config,
parallel_config=parallel_config.moe_parallel_config if self.use_moe
else parallel_config.dp_mp_config)
# If the user doesn't pass the fusion function, use the default one
if not lambda_func:
lambda_func = _get_lambda_func()
lambda_func(block, layer_id=i, layers=num_layers,
offset=offset, parallel_config=parallel_config)
self.blocks.append(block)
else:
raise RuntimeError(f"The {self.cls_name} only support sharding propagation or "
f"semi-auto parallel mode now.")
def construct(self, hidden_states, attention_mask, encoder_output=None, memory_mask=None,
init_reset=True, batch_valid_length=None):
present_layer = ()
if self.use_moe:
accum_loss = self.aux_loss
for i in range(self.num_layers):
hidden_states, present, aux_loss = self.blocks[i](hidden_states,
attention_mask,
encoder_output,
memory_mask,
init_reset,
batch_valid_length)
present_layer = present_layer + (present,)
accum_loss = self.add(accum_loss, aux_loss)
return hidden_states, present_layer, accum_loss
# Loop through each self-attention layer
for i in range(self.num_layers):
hidden_states, present = self.blocks[i](hidden_states,
attention_mask,
encoder_output,
memory_mask,
init_reset,
batch_valid_length)
present_layer = present_layer + (present,)
return hidden_states, present_layer
[文档]class Transformer(Cell):
r"""
Transformer module including encoder and decoder. The difference with the original implements is the module use
the residual addition before the layer normalization. And the default hidden act is `gelu`.
The details can be found in `Attention is all you need <https://arxiv.org/pdf/1706.03762v5.pdf>`_.
Note:
This is an experimental interface that is subject to change or deletion.
Args:
hidden_size(int): The hidden size of the input.
batch_size(int): The batch size of the input tensor.
ffn_hidden_size(int): The hidden size of bottleneck in the feedforward layer.
src_seq_length(int): The seq_length of the encoder's input tensor.
tgt_seq_length(int): The seq_length of the decoder's input tensor.
encoder_layers(int): The layers of the `TransformerEncoderLayer`. Default 3.
decoder_layers(int): The layers of the `TransformerDecoderLayer`. Default 3.
num_heads(int): The number of the heads. Default: 2.
attention_dropout_rate(float): The dropout rate of the attention scores. Default:0.1.
hidden_dropout_rate(float): The dropout rate of the final output of the layer. Default:0.1.
hidden_act(str): The activation of the internal feedforward layer. Supports 'relu',
'relu6', 'tanh', 'gelu', 'fast_gelu', 'elu', 'sigmoid', 'prelu', 'leakyrelu', 'hswish',
'hsigmoid', 'logsigmoid' and so on. Default: gelu.
post_layernorm_residual(bool): Do residuals adds before the layernorm. Default False.
layernorm_compute_type(dtype.Number): The computation type of the layernorm.
Should be dtype.float32 or dtype.float16. Default dtype.float32.
softmax_compute_type(dtype.Number): The computation type of the softmax in the attention.
Should be dtype.float32 or dtype.float16. Default mstype.float32.
param_init_type(dtype.Number): The parameter initialization type of the module.
Should be dtype.float32 or dtype.float16. Default dtype.float32.
lambda_func: A function can determine the fusion index, pipeline stages and recompute attribute. If the user
wants to determine the pipeline stage and gradient aggregation fusion, the user can pass a function
that accepts `network`, `layer_id`, `offset`, `parallel_config`, `layers`. The `network(Cell)`
represents the transformer block, `layer_id(int)` means the layer index for the current module, counts
from zero, `offset(int)` means the layer_index needs an offset, if there are other modules in the net.
The default setting for the pipeline is: `(layer_id + offset) // ((encoder_layers + decoder_layers)
/ pipeline_stage)`. Default None.
use_past(bool): Use the past state to compute, used for incremental prediction. Default False.
moe_config(MoEConfig): The configuration of MoE (Mixture of Expert). Default is an instance of MoEConfig
with default values. Please see `MoEConfig`.
parallel_config(TransformerOpParallelConfig): The parallel configure. Default `default_transformer_config`,
an instance of `TransformerOpParallelConfig` with default args.
Inputs:
- **encoder_inputs** (Tensor) - The input tensor with shape [batch_size, seq_length, hidden_size] or
[batch_size * seq_length, hidden_size].
- **encoder_masks** (Tensor) - The attention mask for decoder with shape
[batch_size, seq_length, seq_length].
- **decoder_inputs** (Tensor) - The output of the encoder with shape [batch_size, seq_length, hidden_size]
or [batch_size * seq_length, hidden_size], this should be none if the decoder layer is 0.
- **decoder_masks** (Tensor) - The attention mask for decoder with shape
[batch_size, seq_length, seq_length]
- **memory_mask** (Tensor) - The memory mask of the cross attention with shape [batch, tgt_seq_length,
src_seq_length]
where tgt_seq_length is the length of the decoder. The output of the encoder with shape [batch_size,
seq_length, hidden_size], this should be none if the decoder layer is 0.
- **init_reset** (Tensor) - A bool tensor with shape [1], used to clear the past key parameter and
past value parameter used in the incremental prediction. Only valid when use_past is True. Default True.
- **batch_valid_length** (Tensor) - Int32 tensor with shape [batch_size] the past calculated the index.
Used for incremental prediction when the use_past is True. Default None.
Outputs:
Tuple, a tuple contains(`output`, `encoder_layer_present`, `decoder_layer_present`, `accum_loss`)
- **output** (Tensor) - If there is only encoder, the output logit of the encoder layer. The shape is
[batch, src_seq_length, hidden_size] or [batch * src_seq_length, hidden_size], if there are encoder and
decoders, the output is from the decoder layer. The shape is [batch, tgt_seq_length, hidden_size] or
[batch * tgt_seq_length, hidden_size].
- **encoder_layer_present** (Tuple) - A tuple with size of num_layers, where each tuple is the tensor the
projected key and value vector in self attention with shape ((batch_size, num_heads, size_per_head,
src_seq_length), (batch_size, num_heads, src_seq_length, size_per_head)).
- **decoder_layer_present** (Tuple) - A tuple with size of num_layers, where each tuple is the tensor
of the projected key and value vector in self attention with shape ((batch_size, num_heads, size_per_head,
tgt_seq_length), (batch_size, num_heads, tgt_seq_length, size_per_head)), and the
projected key and value vector in cross attention with shape
(batch_size, num_heads, size_per_head, src_seq_length),
(batch_size, num_heads, src_seq_length, size_per_head)). If the decoder is not set, the
returned value will be None.
- **accum_loss** (Tensor) - A Tensor indicates an auxiliary loss to minimize the mean square of the data
part routed to each expert, and only returned if the number of experts is greater than 1.
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> import numpy as np
>>> from mindspore import dtype as mstype
>>> from mindspore.nn.transformer import Transformer
>>> from mindspore import Tensor
>>> model = Transformer(batch_size=2, encoder_layers=1, decoder_layers=2, hidden_size=64,
... ffn_hidden_size=64, src_seq_length=20, tgt_seq_length=10)
>>> encoder_input_value = Tensor(np.ones((2, 20, 64)), mstype.float32)
>>> encoder_input_mask = Tensor(np.ones((2, 20, 20)), mstype.float16)
>>> decoder_input_value = Tensor(np.ones((2, 10, 64)), mstype.float32)
>>> decoder_input_mask = Tensor(np.ones((2, 10, 10)), mstype.float16)
>>> memory_mask = Tensor(np.ones((2, 10, 20)), mstype.float16)
>>> output, en_past, de_past = model(encoder_input_value, encoder_input_mask, decoder_input_value,
... decoder_input_mask, memory_mask)
>>> print(output.shape)
(2, 10, 64)
>>> print(len(en_past))
1
>>> print(len(de_past))
2
>>> print(en_past[0][0].shape)
(2, 2, 32, 20)
>>> print(en_past[0][1].shape)
(2, 2, 20, 32)
>>> print(de_past[0][0].shape)
(2, 2, 32, 10)
>>> print(de_past[0][1].shape)
(2, 2, 10, 32)
>>> print(de_past[0][2].shape)
(2, 2, 32, 20)
>>> print(de_past[0][3].shape)
(2, 2, 20, 32)
"""
@_args_type_validator_check(batch_size=Validator.check_positive_int,
hidden_size=Validator.check_positive_int,
num_heads=Validator.check_positive_int,
ffn_hidden_size=Validator.check_positive_int,
src_seq_length=Validator.check_positive_int,
encoder_layers=Validator.check_positive_int,
decoder_layers=Validator.check_non_negative_int,
tgt_seq_length=Validator.check_positive_int,
attention_dropout_rate=Validator.check_non_negative_float,
hidden_dropout_rate=Validator.check_non_negative_float,
hidden_act=_valid_type_checks([str], "Transformer"),
post_layernorm_residual=Validator.check_bool,
layernorm_compute_type=_valid_value_checks([mstype.float32, mstype.float16],
"Transformer"),
softmax_compute_type=_valid_value_checks([mstype.float32, mstype.float16],
"Transformer"),
param_init_type=_valid_value_checks([mstype.float32, mstype.float16], "Transformer"),
parallel_config=_valid_type_checks([TransformerOpParallelConfig], "Transformer"),
use_past=Validator.check_bool)
def __init__(self,
hidden_size,
batch_size,
ffn_hidden_size,
src_seq_length,
tgt_seq_length,
encoder_layers=3,
decoder_layers=3,
num_heads=2,
attention_dropout_rate=0.1,
hidden_dropout_rate=0.1,
hidden_act='gelu',
post_layernorm_residual=False,
layernorm_compute_type=mstype.float32,
softmax_compute_type=mstype.float32,
param_init_type=mstype.float32,
lambda_func=None,
use_past=False,
moe_config=default_moe_config,
parallel_config=default_transformer_config):
super(Transformer, self).__init__()
if _get_parallel_mode() in (ParallelMode.AUTO_PARALLEL,) and _is_sharding_propagation():
_check_config(parallel_config)
self.batch_size = batch_size
self.hidden_size = hidden_size
self.src_seq_length = src_seq_length
self.tgt_seq_length = tgt_seq_length
self.use_past = use_past
if encoder_layers <= 0 < decoder_layers:
raise ValueError(f"Transformer doest support encoder layer {encoder_layers} and decoder"
f"layer {decoder_layers}, please use TransformerDecoder")
if encoder_layers > 0 and decoder_layers > 0 and use_past:
raise ValueError(f"The {self.cls_name} with encoder and decoder does not support use_past=True.")
# The shard setting of Transformer is set within the TransformerEncoderLayer
if not lambda_func:
lambda_func = _get_lambda_func(total_layer=encoder_layers + decoder_layers)
_check_moe_config(moe_config, parallel_config)
self.use_moe = (moe_config.expert_num > 1)
self.add = P.Add()
self.aux_loss = Tensor(0.0, mstype.float32)
if encoder_layers > 0:
self.encoder = TransformerEncoder(num_layers=encoder_layers,
batch_size=batch_size,
hidden_size=hidden_size,
ffn_hidden_size=ffn_hidden_size,
num_heads=num_heads,
seq_length=src_seq_length,
attention_dropout_rate=attention_dropout_rate,
hidden_dropout_rate=hidden_dropout_rate,
hidden_act=hidden_act,
layernorm_compute_type=layernorm_compute_type,
softmax_compute_type=softmax_compute_type,
post_layernorm_residual=post_layernorm_residual,
param_init_type=param_init_type,
lambda_func=lambda_func,
use_past=use_past,
moe_config=moe_config,
parallel_config=parallel_config)
else:
self.encoder = None
# Offset is needed as the encoder has consumed some flags.
# so the decoder need to increase the flags based on the encoder layer
self.decoder = None
if decoder_layers > 0:
self.decoder = TransformerDecoder(num_layers=decoder_layers,
batch_size=batch_size,
hidden_size=hidden_size,
ffn_hidden_size=ffn_hidden_size,
num_heads=num_heads,
src_seq_length=src_seq_length,
tgt_seq_length=tgt_seq_length,
attention_dropout_rate=attention_dropout_rate,
hidden_dropout_rate=hidden_dropout_rate,
hidden_act=hidden_act,
post_layernorm_residual=post_layernorm_residual,
layernorm_compute_type=layernorm_compute_type,
softmax_compute_type=softmax_compute_type,
lambda_func=lambda_func,
use_past=use_past,
param_init_type=param_init_type,
offset=encoder_layers,
moe_config=moe_config,
parallel_config=parallel_config)
elif _get_parallel_mode() not in (ParallelMode.AUTO_PARALLEL,):
_check_config(parallel_config)
self.batch_size = batch_size
self.hidden_size = hidden_size
self.src_seq_length = src_seq_length
self.tgt_seq_length = tgt_seq_length
self.use_past = use_past
if encoder_layers <= 0 < decoder_layers:
raise ValueError(f"Transformer doest support encoder layer {encoder_layers} and decoder"
f"layer {decoder_layers}, please use TransformerDecoder")
if encoder_layers > 0 and decoder_layers > 0 and use_past:
raise ValueError(f"The {self.cls_name} with encoder and decoder does not support use_past=True.")
logger.warning("For parallel mode, sharding propagation is recommended, you can use it by setting "
"'set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL, "
"search_mode=\"sharding_propagation\")' and "
"'set_algo_parameters(elementwise_op_strategy_follow=False, fully_use_devices=False)'")
# The shard setting of Transformer is set within the TransformerEncoderLayer
if not lambda_func:
lambda_func = _get_lambda_func(total_layer=encoder_layers + decoder_layers)
_check_moe_config(moe_config, parallel_config)
self.use_moe = (moe_config.expert_num > 1)
self.add = P.Add().shard(((), ()))
self.aux_loss = Tensor(0.0, mstype.float32)
if encoder_layers > 0:
self.encoder = TransformerEncoder(num_layers=encoder_layers,
batch_size=batch_size,
hidden_size=hidden_size,
ffn_hidden_size=ffn_hidden_size,
num_heads=num_heads,
seq_length=src_seq_length,
attention_dropout_rate=attention_dropout_rate,
hidden_dropout_rate=hidden_dropout_rate,
hidden_act=hidden_act,
layernorm_compute_type=layernorm_compute_type,
softmax_compute_type=softmax_compute_type,
post_layernorm_residual=post_layernorm_residual,
param_init_type=param_init_type,
lambda_func=lambda_func,
use_past=use_past,
moe_config=moe_config,
parallel_config=parallel_config)
else:
self.encoder = None
# Offset is needed as the encoder has consumed some flags.
# so the decoder need to increase the flags based on the encoder layer
self.decoder = None
if decoder_layers > 0:
self.decoder = TransformerDecoder(num_layers=decoder_layers,
batch_size=batch_size,
hidden_size=hidden_size,
ffn_hidden_size=ffn_hidden_size,
num_heads=num_heads,
src_seq_length=src_seq_length,
tgt_seq_length=tgt_seq_length,
attention_dropout_rate=attention_dropout_rate,
hidden_dropout_rate=hidden_dropout_rate,
hidden_act=hidden_act,
post_layernorm_residual=post_layernorm_residual,
layernorm_compute_type=layernorm_compute_type,
softmax_compute_type=softmax_compute_type,
lambda_func=lambda_func,
use_past=use_past,
param_init_type=param_init_type,
offset=encoder_layers,
moe_config=moe_config,
parallel_config=parallel_config)
else:
raise RuntimeError(f"The {self.cls_name} only support sharding propagation or "
f"semi-auto parallel mode now.")
def construct(self, encoder_inputs,
encoder_masks,
decoder_inputs=None,
decoder_masks=None,
memory_mask=None,
init_reset=True,
batch_valid_length=None):
encoder_output = None
output = None
encoder_layer_present = None
decoder_layer_present = None
accum_loss = self.aux_loss
if self.encoder is not None:
if self.use_moe:
encoder_output, encoder_layer_present, encoder_aux_loss = self.encoder(encoder_inputs, encoder_masks,
init_reset, batch_valid_length)
accum_loss = self.add(accum_loss, encoder_aux_loss)
else:
encoder_output, encoder_layer_present = self.encoder(encoder_inputs, encoder_masks, init_reset,
batch_valid_length)
output = encoder_output
if self.decoder is not None:
# decoder mask should be created outside of the model
if self.use_moe:
decoder_output, decoder_layer_present, decoder_aux_loss = self.decoder(decoder_inputs, decoder_masks,
encoder_output, memory_mask,
init_reset, batch_valid_length)
accum_loss = self.add(accum_loss, decoder_aux_loss)
else:
decoder_output, decoder_layer_present = self.decoder(decoder_inputs,
decoder_masks,
encoder_output,
memory_mask, init_reset,
batch_valid_length)
output = decoder_output
if self.use_moe:
return output, encoder_layer_present, decoder_layer_present, accum_loss
return output, encoder_layer_present, decoder_layer_present