模型权重初始化

?

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

def weight_init(m):  # 初始化权重     # print(m)     if isinstance(m, torch.nn.Conv3d):         n = m.kernel_size[0] * m.kernel_size[1] * m.kernel_size[2] * m.out_channels         m.weight.data = torch.randint_like(m.weight.data, low=-128, high=127)         # m.bias.data.zero_()         if m.bias!=None:             m.bias.data = torch.randint_like(m.bias.data, low=-128, high=127)     elif isinstance(m, torch.nn.Conv2d):         n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels         # data = np.load("weight.npy")         # m.weight.data = torch.tensor(data)         m.weight.data = torch.randint_like(m.weight.data, low=-128, high=127)         # print("weight",m.weight.data.shape)         # print(m.weight.data)         # print(m.weight.data)         # m=torch.nn.Conv2d(in_channels=m.in_channels, out_channels=m.out_channels, kernel_size=m.kernel_size, bias=True, stride=m.stride, padding=m.padding)         if m.bias!=None:             m.bias.data = torch.randint_like(m.bias.data, low=-128, high=127)     elif isinstance(m, torch.nn.BatchNorm3d):         m.weight.data.fill_(1)         m.bias.data.zero_()     elif isinstance(m, torch.nn.Linear):         m.weight.data=torch.randint_like(m.weight.data, low=-128, high=127)         if m.bias is not None:             m.bias.data.zero_()                              # 将模型权重初始化为int8  model.apply(weight_init)