Code issue

[sulaimanvesal]

Hi,

I am getting this issue, and I don't know why. The data preprocessing and other steps are done. Is it something related to different timm versions?

  File "C:/scripts/train.py", line 74, in <module>
    trainer.setup()
  File "C:\HanWha\scripts\utils\regression_trainer.py", line 56, in setup
    self.model = Count(args)
  File "C:\scripts\model\model.py", line 40, in __init__
    self.LA_end1 = SAAM(256, 4, 4, 1)
  File "C:\scripts\model\pvt.py", line 311, in __init__
    self._block = GroupBlock(dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
  File "C:\scripts\model\pvt.py", line 73, in __init__
    super(GroupBlock, self).__init__(dim, num_heads, mlp_ratio, qkv_bias, drop, attn_drop,
  File "C:\Users\sulai\anaconda3\envs\pytorch_env\lib\site-packages\timm\models\vision_transformer.py", line 141, in __init__
    self.ls1 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
  File "C:\Users\sulai\anaconda3\envs\pytorch_env\lib\site-packages\timm\models\vision_transformer.py", line 107, in __init__
    self.gamma = nn.Parameter(init_values * torch.ones(dim))
TypeError: unsupported operand type(s) for *: 'type' and 'Tensor'

[cha15yq]

Yes, I think so, the version of package timm has been updated several times since I finished this project. I believe I have left the timm version I used in the readme.

On Wed, 8 Nov 2023 at 21:26, Sulaiman Vesal @.***> wrote:

Hi,

I am getting this issue, and I don't know why. The data preprocessing and other steps are done. Is it something related to different timm versions?

File "C:/scripts/train.py", line 74, in trainer.setup() File "C:\HanWha\scripts\utils\regression_trainer.py", line 56, in setup self.model = Count(args) File "C:\scripts\model\model.py", line 40, in init self.LA_end1 = SAAM(256, 4, 4, 1) File "C:\scripts\model\pvt.py", line 311, in init self._block = GroupBlock(dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., File "C:\scripts\model\pvt.py", line 73, in init super(GroupBlock, self).init(dim, num_heads, mlp_ratio, qkv_bias, drop, attn_drop, File "C:\Users\sulai\anaconda3\envs\pytorch_env\lib\site-packages\timm\models\vision_transformer.py", line 141, in init self.ls1 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity() File "C:\Users\sulai\anaconda3\envs\pytorch_env\lib\site-packages\timm\models\vision_transformer.py", line 107, in init self.gamma = nn.Parameter(init_values torch.ones(dim))TypeError: unsupported operand type(s) for : 'type' and 'Tensor'

— Reply to this email directly, view it on GitHub https://github.com/cha15yq/CUT/issues/6, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANGLTWIRWVB3IPQQX3T5F43YDP2HTAVCNFSM6AAAAAA7DRNYBWVHI2DSMVQWIX3LMV43ASLTON2WKOZRHE4DINBSGQYDMMY . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[sulaimanvesal]

Thanks for the prompt response, just downgraded the timm and it worked well.

However, after 200 epochs starting the validation, getting inf, any experience on this error:

11-08 14:11:59, ----------------------------------------Epoch:200/999----------------------------------------
11-08 14:12:11, Epoch 200 Train, Loss: 0.24, MSE: 20.62, MAE: 7.66, 
level2: ssim: 0.1003 seg: 0.0011 tv:0.3910;
level3: ssim: 0.1111 seg: 0.0011 tv:0.4105;
level4: ssim: 0.1425 seg: 0.0014 tv:0.4735;  Cost: 12.7 sec 
C:\Users\sulai\anaconda3\envs\pytorch_env\lib\site-packages\numpy\core\fromnumeric.py:3432: RuntimeWarning: Mean of empty slice.
  return _methods._mean(a, axis=axis, dtype=dtype,
C:\Users\sulai\anaconda3\envs\pytorch_env\lib\site-packages\numpy\core\_methods.py:190: RuntimeWarning: invalid value encountered in double_scalars
  ret = ret.dtype.type(ret / rcount)
11-08 14:12:13, Epoch 200 Val, MAE: nan, MSE: nan,  Cost 0.0 sec
11-08 14:12:13, ----------------------------------------Epoch:201/999----------------------------------------
Best Result: MAE: inf MSE:inf
11-08 14:12:25, Epoch 201 Train, Loss: 0.22, MSE: 12.60, MAE: 6.40, 
level2: ssim: 0.0941 seg: 0.0011 tv:0.3994;
level3: ssim: 0.1040 seg: 0.0011 tv:0.4190;
level4: ssim: 0.1330 seg: 0.0013 tv:0.4797;  Cost: 12.8 sec 

[cha15yq]

It looks like the test_data_loader contains no samples.

[sulaimanvesal]

Thanks fixed the issue.

I could train CUT on both SHHA and SHHB. However, I couldn't reproduce the results on the paper.

For SSHA, I trained two times, and the lowest MAE and MSE I got was: 58.1, 95.2. For SSHB, MAE and MSE was 8.0 and 12.0.

Any suggestions?

[cha15yq]

You may need to start the evaluation earlier. Maybe you could try train the model with Google Colab, since I did SHA experiment on that. By the time I uploaded the code, I was able to reproduce the result with Google Colab, but since it has been a year, I am not sure if it still gives a similar result.

[sulaimanvesal]

Thanks, I was able to reproduce somewhere close to your results but not exactly. That should be fine.

One last question: why the shape of result during inference for result, _, _, _, _, _ = model(input_data) is 1x1x128x128 when we test and want to plot the density map? and it's not the actual input shape?

[cha15yq]

When I designed this model, I followed a common choice of downsampling ratio in most crowd counting research (which is 8).

[sulaimanvesal]

Hi @cha15yq

You can close this issue, but could you please your preprocessing script for NWPU dataset? I found so many variations for point-map version but not density map. I know, you mentioned that's very similar to qnrf code, but the results I get using that code doesn't match your paper results. Therefore I thought, maybe something is wrong with the data preprocessing for NWPU.

[yangtle]

Hi @cha15yq

You can close this issue, but could you please your preprocessing script for NWPU dataset? I found so many variations for point-map version but not density map. I know, you mentioned that's very similar to qnrf code, but the results I get using that code doesn't match your paper results. Therefore I thought, maybe something is wrong with the data preprocessing for NWPU.

Hello, have you run experiments on the JHU-Crowd and NWPU datasets? Could you please provide the preprocessing code for these two datasets? I would greatly appreciate it!