List(model.parameters()) is []

vision
1 
class Mymodel(nn.Module):
    def __init__(self, 
                 base_encoder, 
                 num_classes=21, 
                 moving_average_decay = 0.99,
                 augment_fn = None,
                 augment_fn2 = None,
                 use_momentum=True):
        super(Mymodel, self).__init__()

        self.m = moving_average_decay

        self.online_encoder = base_encoder
        self.target_encoder = copy.deepcopy(self.online_encoder)

        for param_q, param_k in zip(self.online_encoder.parameters(), self.target_encoder.parameters()):
            param_k.data.copy_(param_q.data)  # initialize
            param_k.requires_grad = False  # not update by gradient
            print("????")

    @torch.no_grad()
    def _momentum_update_key_encoder(self):
        """
        Momentum update of the key encoder
        """
        for param_q, param_k in zip(self.online_encoder.parameters(), self.target_encoder.parameters()):
            param_k.data = param_k.data * self.m + param_q.data * (1. - self.m)
            print(f"---{param_k.data}---{param_q.data}")


    def forward(self, x, rotate_type=0):

        assert rotate_type in [0, 1, 2, 3]

        # for name, i in self.online_encoder.named_parameters():
        #     print(name)


        online_pred = self.online_encoder(TF.rotate(x, rotate_type * 90))

        with torch.no_grad():
            self._momentum_update_key_encoder()
            target_pred = TF.rotate(self.target_encoder(x), rotate_type * 90)

        return online_pred, target_pred

When I tried to write code like this.
I can’t set the parameters in self.online_encoder, and I found that list(self.online_encoder.parameters()) is empty list. (in forward and _momentum_update_key_encoder)

2

maybe yoou need the line:

self.online_encoder = base_encoder()

instead of

self.online_encoder = base_encoder

This depends on what your input base_encoder is

3

Hi, my base_encoder is set before. Such as unet model, defined like: net = UNet(in_channels=3, num_classes=10)

4

Could you show the implement of base_encoder or a snippest code of base_encoder?

5

Just like the model here.

6 
class Mymodel(nn.Module):
    def __init__(self, 
                 base_encoder, 
                 num_classes=21, 
                 moving_average_decay = 0.99,
                 augment_fn = None,
                 augment_fn2 = None,
                 use_momentum=True):
        super(Mymodel, self).__init__()

        self.m = moving_average_decay

        self.online_encoder = base_encoder
        print(type(base_encoder))  # try to debug

try to debug and show the code how you instantiate the Mymodel

7 
    model_map = {
        'deeplabv3_resnet50': network.deeplabv3_resnet50,
        'deeplabv3plus_resnet50': network.deeplabv3plus_resnet50,
        'deeplabv3_resnet101': network.deeplabv3_resnet101,
        'deeplabv3plus_resnet101': network.deeplabv3plus_resnet101,
        'deeplabv3_mobilenet': network.deeplabv3_mobilenet,
        'deeplabv3plus_mobilenet': network.deeplabv3plus_mobilenet
    }

    net = model_map[opts.model](num_classes=opts.num_classes, output_stride=opts.output_stride)
    if opts.separable_conv and 'plus' in opts.model:
        network.convert_to_separable_conv(net.classifier)
    utils.set_bn_momentum(net.backbone, momentum=0.01)
    
    # model = network.SiamSeg(net, 
    #                         image_size=(opts.crop_size, opts.crop_size),
    #                         projection_size=opts.num_classes,
    #                         use_momentum=False,
    #                 )
    model = Mymodel(
        net,
        num_classes=opts.num_classes, 
        moving_average_decay=opts.momentum,
        augment_fn = None,
        augment_fn2 = None,
        use_momentum=True,
    )

Like this

8

Solved.

I used nn.Dataparallel to train my model on multi-gpu machine, it calculated the gradient on the main gpu, so my function in class to update the parameters is incorrect.

When I choose to use on gpu or DDP, it works well.