couly not understand the code in train?

[rochesterlmg]

hi, sir
class resnet34:

def __init__(self, num_classes):
    self.pre = [
        conv_layer(3, 64, 7, 7, stride=2, shift=False),#input image is 3 channels
        bn_layer(64),
        relu(),
        max_pooling(3,3,2,same=True)
    ]
    self.layer1 = self.stack_ResBlock(64, 64, 3, 1)
    self.layer2 = self.stack_ResBlock(64, 128, 4, 2)
    self.layer3 = self.stack_ResBlock(128, 256, 6, 2)
    self.layer4 = self.stack_ResBlock(256, 512, 3, 2)
    self.avg = global_average_pooling()
    self.fc = fc_sigmoid(512, num_classes)

def train(self):
    self.pre[1].train()
  
    for l in self.layer1:
        l.train()
    for l in self.layer2:
        l.train()
    for l in self.layer3:
        l.train()
    for l in self.layer4:
        l.train()
pre[1] pointed the bn_layer,why not   run self.pre[0].train() to back propogation the conv_layer in self.pre? I found the code in train.py including ResBlock just bp the bn_layer ,no bp operation in conv layer?

[lyzustc]

Very sorry for my late reponse! This is because train() does not mean doing back propagation( It is done in backward() function). What train() does here is to set the batch normalization layers to "train" state (I think you know batch normalization behaves differently during training or evaluating). The name "train" seems ambiguous here, because I just name functions like in Pytorch~

[rochesterlmg]

thanks,could you give me some suggestion to accelerate your codes for training by some method like as numba,cupy and so on?

[lyzustc]

I have not tried any numpy-like libraries with GPU support. So I am sorry that I have no idea whether they will accelerate computing or there are other possible speeding methods :( But I think you can have a try : )