【pytorch】torch.autograd.Function

# *_*coding:utf-8 *_*
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Function, Variable
import json

class HSwishImplementation(Function):

    @staticmethod
    def symbolic(g, input, bias):
        return g.op("HSwish", input, bias, info_s=json.dumps({
            "kernel_size":3,
            "eps":3e-2,
            "other":"Onnx Plugins"
        }))

    @staticmethod
    def forward(ctx, i, bias):
        ctx.save_for_backward(i)
        return i * F.relu6(i+1) / 6 + bias

class MReLU(nn.Module):
    def __init__(self, shape):
        super(MReLU, self).__init__()
        self.bias = nn.Parameter(torch.zeros(shape))
        self.bias.data.fill_(0.5)
    def forward(self, x):
        return HSwishImplementation.apply(x, self.bias)

class FModel(nn.Module):
    def __init__(self):
        super(FModel, self).__init__()
        self.mrelu = MReLU(1)

    def forward(self, x):
        return self.mrelu(x)


if __name__ == '__main__':
    x = Variable(torch.Tensor([1,3,3,3]))
    print(x)
    net = FModel()
    out = net(x)      
    print(out)


# 输出结果
tensor([1., 3., 3., 3.])
tensor([0.8333, 2.5000, 2.5000, 2.5000],
       grad_fn=<HSwishImplementationBackward>)