设置随即种子
使用pytorch_lightning下的seed_everything方法。
若调用GPU,有时还不够,还需排除cudnn加速的随机性。
from pytorch_lightning import seed_everything
seed_everything(0)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
网络初始化
可以用如下代码进行网络参数初始化
class My_CNN(nn.Module):
def __init__(self):
super(My_CNN, self).__init__()
self.tmp1 = nn.Linear(4 * 4 * 64, 512)
self.tmp2 = nn.Conv2d(32, 32, 3)
self.tmp3 = nn.BatchNorm2d(64)
...
for m in self.children():
if isinstance(m, nn.Linear):
print("init linear.")
nn.init.normal_(m.weight, 0, 0.02)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Conv2d):
print("init conv.")
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
print("init batchnorm2d.")
nn.init.normal_(m.weight, 1, 0.02)
nn.init.constant_(m.bias, 0)
...
但套在nn.Sequential()内的网络参数不会被初始化,如下所示:
self.conv1 = nn.Sequential(
nn.Conv2d(1, 32, 3),
nn.BatchNorm2d(32),
nn.ReLU()
)
这种需要手动对Sequential()内的网络进行初始化。
参考资料:
- https://blog.csdn.net/qq_41917697/article/details/115042465
- https://zhuanlan.zhihu.com/p/97651850