# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# 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.
#
"""
Linear module.
"""
from aitemplate.compiler import ops
from aitemplate.frontend.nn.module import Module
from aitemplate.frontend.nn.parameter import Parameter
from aitemplate.testing import detect_target
[docs]class Linear(Module):
r"""Applies a linear transformation to the incoming data: :math:`y = xA^T + b`
Args:
in_channels: size of each input sample
out_channels: size of each output sample
bias: If set to ``False``, the layer will not learn an additive bias.
Default: ``True``
specialization: elementwise operation to add after the linear operation,
Default: ``None``
dtype: data type, default: ``float16``
Shape:
- Input: :math:`(*, H_{in})` where :math:`*` means any number of
dimensions including none and :math:`H_{in} = \text{in_channels}`.
- Output: :math:`(*, H_{out})` where all but the last dimension
are the same shape as the input and :math:`H_{out} = \text{out_channels}`.
Attributes:
weight: the learnable weights of the module of shape
:math:`(\text{out_channels}, \text{in_channels})`. The values are
initialized from :math:`\mathcal{U}(-\sqrt{k}, \sqrt{k})`, where
:math:`k = \frac{1}{\text{in\_channels}}`
bias: the learnable bias of the module of shape :math:`(\text{out_channels})`.
If :attr:`bias` is ``True``, the values are initialized from
:math:`\mathcal{U}(-\sqrt{k}, \sqrt{k})` where
:math:`k = \frac{1}{\text{in_channels}}`
Examples::
>>> m = nn.Linear(20, 30)
>>> input = Tensor(shape=[128, 20])
>>> output = m(input)
Tensor(shape=[128, 30])
"""
USE_CUDA = None
def __init__(
self,
in_channels,
out_channels,
bias=True,
specialization=None,
dtype="float16",
**kwargs,
):
super().__init__()
if Linear.USE_CUDA is None:
Linear.USE_CUDA = detect_target().name() == "cuda"
self.weight = Parameter(shape=[out_channels, in_channels], dtype=dtype)
op_name = "gemm_rcr_bias" if bias else "gemm_rcr"
if specialization is not None:
op_name += "_" + specialization
if bias:
self.bias = Parameter(shape=[out_channels], dtype=dtype)
op_func = getattr(ops, op_name)
self._op_name = op_name
self.op = op_func(**kwargs)
self.use_bias = bias
self.in_channels = in_channels
[docs] def forward(self, *args):
assert len(args) >= 1
x = args[0]
if not self.USE_CUDA and len(x._attrs["shape"]) != 2:
x = ops.reshape()(x, [-1, self.in_channels])
inputs = [x, self.weight.tensor()]
if self.use_bias:
inputs.append(self.bias.tensor())
if len(args) == 2:
inputs.append(args[1])
output = self.op(*inputs)
return output