Spaces:
Running
on
Zero
Running
on
Zero
File size: 2,294 Bytes
53a077e |
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 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 |
""" Activations (jit)
A collection of jit-scripted activations fn and modules with a common interface so that they can
easily be swapped. All have an `inplace` arg even if not used.
All jit scripted activations are lacking in-place variations on purpose, scripted kernel fusion does not
currently work across in-place op boundaries, thus performance is equal to or less than the non-scripted
versions if they contain in-place ops.
Copyright 2020 Ross Wightman
"""
import torch
from torch import nn as nn
from torch.nn import functional as F
__all__ = ['swish_jit', 'SwishJit', 'mish_jit', 'MishJit',
'hard_sigmoid_jit', 'HardSigmoidJit', 'hard_swish_jit', 'HardSwishJit']
@torch.jit.script
def swish_jit(x, inplace: bool = False):
"""Swish - Described originally as SiLU (https://arxiv.org/abs/1702.03118v3)
and also as Swish (https://arxiv.org/abs/1710.05941).
TODO Rename to SiLU with addition to PyTorch
"""
return x.mul(x.sigmoid())
@torch.jit.script
def mish_jit(x, _inplace: bool = False):
"""Mish: A Self Regularized Non-Monotonic Neural Activation Function - https://arxiv.org/abs/1908.08681
"""
return x.mul(F.softplus(x).tanh())
class SwishJit(nn.Module):
def __init__(self, inplace: bool = False):
super(SwishJit, self).__init__()
def forward(self, x):
return swish_jit(x)
class MishJit(nn.Module):
def __init__(self, inplace: bool = False):
super(MishJit, self).__init__()
def forward(self, x):
return mish_jit(x)
@torch.jit.script
def hard_sigmoid_jit(x, inplace: bool = False):
# return F.relu6(x + 3.) / 6.
return (x + 3).clamp(min=0, max=6).div(6.) # clamp seems ever so slightly faster?
class HardSigmoidJit(nn.Module):
def __init__(self, inplace: bool = False):
super(HardSigmoidJit, self).__init__()
def forward(self, x):
return hard_sigmoid_jit(x)
@torch.jit.script
def hard_swish_jit(x, inplace: bool = False):
# return x * (F.relu6(x + 3.) / 6)
return x * (x + 3).clamp(min=0, max=6).div(6.) # clamp seems ever so slightly faster?
class HardSwishJit(nn.Module):
def __init__(self, inplace: bool = False):
super(HardSwishJit, self).__init__()
def forward(self, x):
return hard_swish_jit(x)
|