Tensor size mismatch error when dynamic options are used

[SmnKr]

Hi @muhanzhang,

I've used this model for link prediction on a few datasets. Recently, when I run it on two of my new datasets with dynamic run type options, i.e. --dynamic_train --dynamic_val --dynamic_test, I get the following error at run-time. I should mention that if I don't use the above options, namely if I use the in-memory run type, it runs successfully. Do you know what causes the error and how can I fix it?

File "/nfstmp/.../seal_link_pred.py", line 708, in
model.load_state_dict(torch.load(model_name))
File "/nfstmp/.../python3.7/site-packages/torch/nn/modules/module.py", line 1052, in load_state_dict
self.class.name, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for DGCNN:
size mismatch for lin1.weight: copying a param with shape torch.Size([128, 800]) from checkpoint, the shape in current model is torch.Size([128, 1344]).

[muhanzhang]

Hi! It seems it is loading a previously saved model checkpoint. Did you use "--continue_from" and load a previously saved in-memory trained model in the dynamic mode? The dynamic mode will only use the first 1000 training graphs to estimate the $k$ used in SortPooling of DGCNN (see this line), which might result in different $k$ thus different dense dimension before the first linear layer. You can maually set $k$ to your previously obtained $k$ in in-memory mode.

[SmnKr]

Thanks for your response Muhan.
No, I'm not using "--continue_from" when running the seal_link_pred.py script. These are all options that I'm using for running it in dynamic mode:
--dataset --num_hops --use_feature --use_edge_weight --epochs --num_workers --runs --dynamic_train --dynamic_val --dynamic_test --train_percent --pathdir

[muhanzhang]

Then can you show more lines near line 708 of your seal_link_pred.py file? I'd like to know what line 708's model.load_state_dict(torch.load(model_name)) is doing in order to debug.

[SmnKr]

Yes, sure. These are lines 693-709:

[muhanzhang]

Thanks. The lines show that you are loading a previously saved model checkpoint. Did you train your model using in-memory mode and here load it in dynamic mode? My previous points are still valid, i.e., the gap causes their used $k$ to be different.