1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
|
import torch
import torch.ao.nn.quantized.dynamic as nnqd
import torch.ao.nn.intrinsic as nni
class LinearReLU(nnqd.Linear):
r"""
A LinearReLU module fused from Linear and ReLU modules that can be used
for dynamic quantization.
Supports both, FP16 and INT8 quantization.
We adopt the same interface as :class:`torch.ao.nn.quantized.dynamic.Linear`.
Attributes:
Same as torch.ao.nn.quantized.dynamic.Linear
Examples::
>>> m = nn.intrinsic.quantized.dynamic.LinearReLU(20, 30)
>>> input = torch.randn(128, 20)
>>> # xdoctest: +SKIP
>>> output = m(input)
>>> print(output.size())
torch.Size([128, 30])
"""
_FLOAT_MODULE = nni.LinearReLU # type: ignore[assignment]
def __init__(self, in_features, out_features, bias=True, dtype=torch.qint8):
super().__init__(in_features, out_features, bias, dtype)
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self._packed_params.dtype == torch.qint8:
# TODO check if we should set reduce_rage = True by default here
Y = torch.ops.quantized.linear_relu_dynamic(
x, self._packed_params._packed_params, reduce_range=True)
elif self._packed_params.dtype == torch.float16:
Y = torch.ops.quantized.linear_relu_dynamic_fp16(
x, self._packed_params._packed_params)
else:
raise RuntimeError('Unsupported dtype on dynamic quantized linear relu!')
return Y.to(x.dtype)
def _get_name(self):
return 'DynamicQuantizedLinearReLU'
@classmethod
def from_float(cls, mod):
return super(LinearReLU, cls).from_float(mod)
@classmethod
def from_reference(cls, ref_qlinear_relu):
return super().from_reference(ref_qlinear_relu[0])
|
