train siamrpn_alex failed

When I train siamrpn_alex in experiments using "python -m torch.distributed.launch --nproc_per_node=4 --master_port=2341 …/…/tools/train.py --cfg config.yaml ", it give me this error.

It seems like network structure error?

result = self.forward(*input, **kwargs)
  File "/root/ly18/pysot/pysot/utils/distributed.py", line 43, in forward
    return self.module(*args, **kwargs)
  File "/opt/conda/envs/pysot/lib/python3.7/site-packages/torch/nn/modules/module.py", line 477, in __call__
    result = self.forward(*input, **kwargs)
  File "/root/ly18/pysot/pysot/models/model_builder.py", line 105, in forward
    loc_loss = weight_l1_loss(loc, label_loc, label_loc_weight)
  File "/root/ly18/pysot/pysot/models/loss.py", line 45, in weight_l1_loss
    diff = (pred_loc - label_loc).abs()
RuntimeError: The size of tensor a (17) must match the size of tensor b (25) at non-singleton dimension 4

I have trained siamrpn_mobilev2 already.

[lb1100]

siamrpn_alex_dwxcorr is only for inference, there is no training parameters in the config. If you want to fine-tune alexnet, you can use siamrpn_alex_dwxcorr_16gpu as config, and load the pretrained network.

[yustaub]

@MaxLin86 @StrangerZhang @lb1100 if I want to fine-tune from the pretrained models on my dataset, change the BACKBONE:PRETRAINED:‘pretrained_models/alexnet-bn.pth’ with the provides model in the model zoo?thanks for your reply!