RuntimeError with customized loss while running loss.backward()

[nowangry]

I wrote a simple loss function in {eval_rcnn.py → eval_one_epoch_joint() → for loop (for data in dataloader: ...)}. I have enabled all gradient propagation by replacing "torch.no_grad()" with "torch.set_grad_enabled(True)" and set "inputs.requires_grad = True". Here is the use of my loss:

    Z_rcnn_t = rcnn_cls.float()
    loss_adv_cls = - Z_rcnn_t.sum()
    model.zero_grad()
    loss_adv_cls.backward()

The ".backward()" end up with a Runtime Error:

File "/home/jqwu/Codes/PointRCNN/tools/simple_attack.py", line 974, in eval_single_ckpt(root_result_dir) File "/home/jqwu/Codes/PointRCNN/tools/simple_attack.py", line 832, in eval_single_ckpt eval_one_epoch(model, test_loader, epoch_id, root_result_dir, logger) File "/home/jqwu/Codes/PointRCNN/tools/simple_attack.py", line 759, in eval_one_epoch ret_dict = eval_one_epoch_joint(model, dataloader, epoch_id, result_dir, logger) File "/home/jqwu/Codes/PointRCNN/tools/simple_attack.py", line 580, in eval_one_epoch_joint loss_adv_cls.backward() File "/home/jqwu/.conda/envs/PointRCNN-py37/lib/python3.7/site-packages/torch/tensor.py", line 195, in backward torch.autograd.backward(self, gradient, retain_graph, create_graph) File "/home/jqwu/.conda/envs/PointRCNN-py37/lib/python3.7/site-packages/torch/autograd/init.py", line 99, in backward allow_unreachable=True) # allow_unreachable flag RuntimeError: Expected isFloatingType(grads[i].type().scalarType()) to be true, but got false. (Could this error message be improved? If so, please report an enhancement request to PyTorch.)

I have tried the following solutions, but they didn't work in my case:

https://github.com/qiqihaer/votenet-kitti/issues/5 https://github.com/HaozheQi/P2B/issues/8

I am using CUDA 10.1, torch 1.4, python 3.7. I have spent a lot of time on this. Anyone has solved a similar problem? 😿

[zqy411470859]

Have you solved this problem？解决了么

[nowangry]

Have you solved this problem？解决了么

Not yet....