Faster-rcnn.pytorch: RuntimeError: reciprocal is not implemented for type torch.cuda.LongTensor

[ArchieGu]

I saw someone posted a similar problem #116

Does this have something to do with Pytorch version???

Anyone with solution, plz contact me, thank you.

[ArchieGu]

OK, it seems that Pytorch 0.3.0 can solve the problem. Maybe it is the version caused the problem.

[LiuHao-THU]

I also met this problem when I try to compile with pytorch 0.4.0? how to solve this problem? while i do not want to change the version of my pytorch...

[xylcbd]

modify faster-rcnn.pytorch\lib\model\rpn\anchor_target_layer.py +156, from

num_examples = torch.sum(labels[i] >= 0)

to

num_examples = torch.sum(labels[i] >= 0).item()

[xylcbd]

another issue can be resoved with modify faster-rcnn.pytorch\lib\model\rpn\proposal_target_layer_cascade.py +133, from

labels = gt_boxes[:,:,4].contiguous().view(-1).index(offset.view(-1))\ .view(batch_size, -1)

to

labels = gt_boxes[:,:,4].contiguous().view(-1).index((offset.view(-1), ))\ .view(batch_size, -1)

[D-X-Y]

https://github.com/pytorch/pytorch/issues/2772

[yjump]

Hi @xylcbd, thank you for your solutions. But I meet with anothor issue caused by:

Traceback (most recent call last):
File "trainval_net.py", line 332, in
clip_gradient(fasterRCNN, 10.)
File "/home/yanjp/faster-rcnn.pytorch/lib/model/utils/net_utils.py", line 52, in clip_gradient
p.grad.mul_(norm)
RuntimeError: Expected object of type torch.cuda.FloatTensor but found type torch.FloatTensor for argument #3 'other'
Exception NameError: "global name 'FileNotFoundError' is not defined" in > ignored

I try to add norm.cuda() to solve it, but failed as:

Traceback (most recent call last):
File "trainval_net.py", line 332, in
clip_gradient(fasterRCNN, 10.)
File "/home/yanjp/faster-rcnn.pytorch/lib/model/utils/net_utils.py", line 51, in clip_gradient
norm = norm.cuda()
AttributeError: 'float' object has no attribute 'cuda'
Exception NameError: "global name 'FileNotFoundError' is not defined" in > ignored

could you kindly offer some advices? Thanks a lot!

[xylcbd]

@jwyang, maybe you should check the input parameter "clip_norm", is it hosted by CUDA?

[yjump]

Thanks for your opinion@xylcbd I have checked clip_norm, it is a float constant. At last I solve my problem by adding

add

        norm = torch.tensor([norm],device='cuda')

before
faster-rcnn.pytorch/lib/model/utils/net_utils.py", line 51,
And now it seems work

[wtliao]

@xylcbd thanks so much for your solution. It works for me now

[isalirezag]

@yjump @wtliao where did you put it exactly? can you please see if i put it in the right position?:

    norm = clip_norm / max(totalnorm, clip_norm)
    for p in model.parameters():
        if p.requires_grad:
            p.grad.mul_(norm)
            norm = torch.tensor([norm],device='cuda')

[yjump]

@isalirezag put it before p.grad.mul_(norm) as gradiants are calculated on GPU

[babyjie57]

@yjump @wtliao Thank you so much for your solution!

[JingXiaolun]

Hi @xylcbd, thank you for your solutions. But I meet with anothor issue caused by:
File "/media/csu/新加卷/AI Competition/Baidu Competition/intermediary_contest/faster-rcnn.pytorch-master/trainval_net.py", line 316, in
loss.backward()
File "/home/csu/anaconda3/lib/python3.6/site-packages/torch/tensor.py", line 93, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph)
File "/home/csu/anaconda3/lib/python3.6/site-packages/torch/autograd/__init__.py", line 89, in backward
allow_unreachable=True) # allow_unreachable flag
File "/media/csu/新加卷/AI Competition/Baidu Competition/intermediary_contest/faster-rcnn.pytorch-master/lib/model/roi_align/functions/roi_align.py", line 38, in backward
assert(self.feature_size is not None and grad_output.is_cuda)
AssertionError
Could you offer me some advice?Thanks a lot

[JingXiaolun]

@kurosaki-fish ,no,have you solved the problem?

[changjo]

For me, changing

totalnorm = np.sqrt(totalnorm) 

to

totalnorm = torch.sqrt(totalnorm)

Solved this problem RuntimeError: Expected object of type torch.cuda.FloatTensor but found type torch.FloatTensor...

[Daniellllzzh]

command:python trainval_net.py --cuda
Solved the problem: AssertionError

[jwyang]

the current master supports pytorch 0.4.0 now, feel free to use it!

[ssli23]

I also met this problem when I try to compile with pytorch 0.4.0? how to solve this problem? while i do not want to change the version of my pytorch
Have you solved your problem?