activevisionlab / learningtocountanything Goto Github PK

Hello. Thank you for this valuable work.
I want to do some tests on my own images and have the result of counting and heatmap images. Currently, according to the guide, I can only test on the images in the dataset. Please help, thanks

So I've been testing this model for a few days and I am still not able to replicate the results. The density maps overlapped in the pictures for qualitative results doesn't turn out as good as in the paper, nor it doesn't work on all the images.

I trained the model with the example_training config, obtaining multiple 'ckpt' files, which I chose the one I think fitted the best (I don't really have a way to decide the best checkpoint since all of them have similar results with the validation set). Then I run the example_localisation_training with said checkpoint and use the generated ckpt (again, one of the generated checkpoints) to run the example_localisation_vis config, expecting good qualitative results.

Have to say, when I run the example_test with the checkpoint from the training, the results are quite similar to the one on the paper:
'test/loss': 0.3367506265640259,
'test_DDP_MAE': 15.221793174743652,
'test_DDP_RMSE': 53.706050872802734
In the paper the results for the test set are MAE = 14.23 and RMSE = 43.83.

But for some reason using that checkpoint for the visualization generate bad results, compared to the ones on the paper. Even more, not all the images generated have a density/heatmap overlapped. Only half of them generate a density map on them, the rest are in grey-scale. I know this is wrong because when I used the "example weights" provided for FSC-133 (counting.ckpt and localisation.ckpt) all the images turned out well, so even if I trained the model wrong, I don't understand why all the images aren't overlapped with a density map.

This is the result of one of the pictures with the example weights provided:

This is the result for the same image I got after training the model by myself:

And this is one example of many of a picture that did not generate a density map:

So maybe im following the steps wrong? should I consider a different kind of configuration?

When I fully follow the README file, when I run the py file, I report the following error“'VisionTransformer' object has no attribute 'dist_token' ”, how can I solve this.as follows:

Traceback (most recent call last):
File "main.py", line 102, in
main()
File "main.py", line 42, in main
model = CountingAnything(CFG)
File "/home/hao/桌面/LearningToCountAnything/models/CountingAnything.py", line 47, in init
self.backbone = ViTExtractor(vit_config)
File "/home/hao/桌面/LearningToCountAnything/models/backbone_vit.py", line 29, in init
self.create_from_base_model(
File "/home/hao/桌面/LearningToCountAnything/models/backbone_vit.py", line 47, in create_from_base_model
self.dist_token = base_model.dist_token
File "/root/anaconda3/envs/LearningToCountAnything/lib/python3.8/site-packages/torch/nn/modules/module.py", line 947, in getattr
raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'VisionTransformer' object has no attribute 'dist_token'

@victorprad @mahobley
1)could you please help understand why do we have two checkpoitns. localisation & counting.ckpt
2) If we have to train on custom dataset, do we have to use both the ckpts ?

3. during inference which ckpt has to be used localisation or counting ?

4. What is the difference between the self.CFG["resume_counting_head_path"], and self.resume path