search

sw-gong / MoNet

Pytorch reproduction of the paper "Gaussian Mixture Model Convolutional Networks" (CVPR 17)
MIT License
55 stars 9 forks source link

RuntimeError: The size of tensor a (4096) must match the size of tensor b (25) at non-singleton dimension 1 #3

Open sure7018 opened 3 years ago

sure7018 commented 3 years ago

The code is: def message(self, x_j, pseudo): (E, D), K = pseudo.size(), self.mu.size(0) print("E:") print(E) print("D:") print(D) print("K:") print(K) gaussian = -0.5 * (pseudo.view(E, 1, D) - self.mu.view(1, K, D))2 print("######################") print(gaussian.shape) a = EPS + self.sigma.view(1, K, D)2 print("!!!!!!!!!!!!!!!!!!!!!") print(a.shape) gaussian = gaussian / (EPS + self.sigma.view(1, K, D)*2) print("@@@@@@@@@@@@@@@@@@@@@@@@@@") print(gaussian.shape) gaussian = torch.exp(gaussian.sum(dim=-1, keepdim=True)) # [E, K, 1] print("$$$$$$$$$$$$$$$$$$$$$$$$$$") print(gaussian.shape) print("x_j:") print(x_j.shape) return (x_j gaussian).sum(dim=1)

And the error is:

Loading model check point from checkpoints/SGRN/res101_faster_rcnn_iter_1200000.pth load_state_dict in network_gcn Loaded. Loaded. test_net, num_images=4961 fc7 = torch.Size([128, 2048]) E: 4096 D: 2 K: 25 ###################### torch.Size([4096, 25, 2]) !!!!!!!!!!!!!!!!!!!!! torch.Size([1, 25, 2]) @@@@@@@@@@@@@@@@@@@@@@@@@@ torch.Size([4096, 25, 2]) $$$$$$$$$$$$$$$$$$$$$$$$$$ torch.Size([4096, 25, 1]) x_j: torch.Size([128, 4096, 2048]) Traceback (most recent call last): File "tools/test_net.py", line 159, in test_net(net, imdb, filename, max_per_image=100) #args.max_per_image) File "/home1/lws/SGRN/tools/../lib/model/test.py", line 145, in test_net scores, boxes = im_detect(net, im) File "/home1/lws/SGRN/tools/../lib/model/test.py", line 107, in im_detect blobs['im_info']) File "/home1/lws/SGRN/tools/../lib/nets/network_gcn.py", line 551, in test_image self.forward(image, im_info, None, mode='TEST') File "/home1/lws/SGRN/tools/../lib/nets/network_gcn.py", line 503, in forward rois, cls_prob, bbox_pred = self._predict() File "/home1/lws/SGRN/tools/../lib/nets/network_gcn.py", line 474, in _predict f = self.gaussian(represent, relation, U) File "/home/user/anaconda3/envs/sgrn/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(input, kwargs) File "/home1/lws/SGRN/tools/../lib/conv/gmm_conv.py", line 47, in forward out = self.propagate(edge_index, x=out, pseudo=pseudo) File "/home/user/anaconda3/envs/sgrn/lib/python3.7/site-packages/torch_geometric/nn/conv/message_passing.py", line 269, in propagate out = self.message(msg_kwargs) File "/home1/lws/SGRN/tools/../lib/conv/gmm_conv.py", line 76, in message return (x_j gaussian).sum(dim=1) RuntimeError: The size of tensor a (4096) must match the size of tensor b (25) at non-singleton dimension 1

vox-1 commented 2 years ago

I met the same problem, did you solve it?

huyen-spec commented 2 years ago

Your x_j should have size torch.Size ([4096, K, in_channels]) with K the kernel size (25 in the above example). in_channels represents the dimension of each input nodes (I think 2048 in your example?).

DeltaMarine101 commented 2 years ago

I met the same problem, did you solve it?

I suggest instead using the more up-to-date version found here built into PyTorch geometric: https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html#torch_geometric.nn.conv.GMMConv