Train on New Dataset

[LokeSiewWen]

Hi, may I know how to train on new Dataset? I am facing this error when trying to input the new dataset.

Traceback (most recent call last):
File "/content/CounTR/FSC_finetune_cross.py", line 456, in
main(args)
File "/content/CounTR/FSC_finetune_cross.py", line 287, in main
output = model(samples, boxes, shot_num)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1511, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1520, in _call_impl
return forward_call(*args, **kwargs)
File "/content/CounTR/models_mae_cross.py", line 206, in forward
pred = self.forward_decoder(latent, boxes, shot_num) # [N, 384, 384]
File "/content/CounTR/models_mae_cross.py", line 174, in forward_decoder
y = torch.cat(y1,dim=0).reshape(shot_num,N,C).to(x.device)
RuntimeError: shape '[2, 1, 512]' is invalid for input of size 512
Traceback (most recent call last):
File "/content/CounTR/FSC_finetune_cross.py", line 456, in
main(args)
File "/content/CounTR/FSC_finetune_cross.py", line 287, in main
output = model(samples, boxes, shot_num)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1511, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1520, in _call_impl
return forward_call(*args, **kwargs)
File "/content/CounTR/models_mae_cross.py", line 206, in forward
pred = self.forward_decoder(latent, boxes, shot_num) # [N, 384, 384]
File "/content/CounTR/models_mae_cross.py", line 174, in forward_decoder
y = torch.cat(y1,dim=0).reshape(shot_num,N,C).to(x.device)
RuntimeError: shape '[2, 1, 512]' is invalid for input of size 512
[08:42:33.922301] [08:42:33.922375] [08:42:33.922399] [08:42:33.922411] [08:42:33.922425] [08:42:33.922438] [08:42:33.922451] [08:42:33.922477]

Thanks for your help in advance!

[Verg-Avesta]

This seems to be a shape mismatch problem. So if you can ensure that the form and size of the new input dataset are the same as FSC147, there should be no problem.

[LokeSiewWen]

Hi, Thanks! Is working now. But I do have a few questions and would like some clarifications from you.

1. Do we need to manually generate the density map?
2. Are they necessary during training?

Thanks for your help in advance! :)

[Verg-Avesta]

Yes, I believe it is essential to manually generate density maps, as Gaussian density maps could lead to smaller errors compared to dot annotations.

[famersalltheday]

Hi, Thanks! Is working now. But I do have a few questions and would like some clarifications from you.

1. Do we need to manually generate the density map?
2. Are they necessary during training?

Thanks for your help in advance! :)

Hi, I was wondering about the process for training new data sets. Is a new dataset file created, or is it a supplement to the original FSC147 file? Because I found out that the annotation.json in FSC147 is read-only, I can't edit it.