Testing the trained models

[AbdulMoqeet]

Hello,

Thank you for sharing the code.

Once the network is trained using train.py, how to test the model?

Regards,
AM

[carolchenyx]

@AbdulMoqeet Dear AbdulMoqeet, I would like to ask a question about train.py. When I ran it , this line of the code(
torch.autograd.grad(dis_validation_loss, discriminator.parameters())) will raise the error({RuntimeError}invalid gradient at index 0 - got [1] but expected shape compatible with [1, 1]). Do you know how to fix it? Thank you.

[AbdulMoqeet]

@carolchenyx I checked my code but couldn't get hint. I was facing similar errors but don't remember how I fixed them though.

Did you make any changes in the code? How's your data directory look like?

[carolchenyx]

@AbdulMoqeet I changed a little bit about the data process. The data directory is like (/media/hkuit104/24d4ed16-ee67-4121-8359-66a09cede5e7/AbnormalDetection/DATASET/SHT/testing/frames/).

[AbdulMoqeet]

It may be a problem of data loading.. My directory looks like videos\frames\08\08_003 where 08 is the Task number and '08_003' is video number. I am giving 'videos\frames' as an input.

[carolchenyx]

@AbdulMoqeet Oh, I see. Let me try it first. Super thanks for you quick reply.

[carolchenyx]

@AbdulMoqeet Dear AM, it seems will raise the same problem. I pushed the code I used in (https://github.com/carolchenyx/Few-shot-Scene-adaptive-Anomaly-Detection). I will be very grateful if you can help to check it. Thank you.
Traceback (most recent call last):
File "/media/hkuit104/24d4ed16-ee67-4121-8359-66a09cede5e7/AbnormalDetection/Few-shot-Scene-adaptive-Anomaly-Detection/train.py", line 269, in
main(k_shots, num_tasks, adam_betas, gen_lr, dis_lr, total_epochs, model_folder_path)
File "/media/hkuit104/24d4ed16-ee67-4121-8359-66a09cede5e7/AbnormalDetection/Few-shot-Scene-adaptive-Anomaly-Detection/train.py", line 199, in main
gen_grads = torch.autograd.grad(gen_validation_loss, generator.parameters())
File "/home/hkuit104/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/autograd/init.py", line 225, in grad
inputs, allow_unused, accumulate_grad=False)
RuntimeError: invalid gradient at index 0 - got [1] but expected shape compatible with [1, 1]

[AbdulMoqeet]

@carolchenyx Added a pull request for working code. Changes are made in your code. Have fun!

[carolchenyx]

@AbdulMoqeet Great! Super thanks again.

[abhishekaich27]

@AbdulMoqeet @carolchenyx Can you share the folder structure of training videos? I have extracted the frames under folder name of videos. Below is the structure I have. Is this correct?

|- Training
    |-frames
         |- 01_001
         |- 01_002
         |- 01_003

This is the structure of the training videos (which by default in Shanghai-Tech, doesn't provide anomalies in training set.)

[AbdulMoqeet]

@abhishekaich27 Yes, but that default structure didn't work for me. I had to change the directories structure.

My directory looks like \Training\frames\08\08_003*.jpg where 08 is the Task number, '08_003' is video number and *.jpg represent the video frames.

Let me know if the problem still persists.

[abhishekaich27]

Exactly @AbdulMoqeet. Does the Task number has some relation with Video number? e.g. 08 of 08_003? These first two digits have some particular meaning?

[AbdulMoqeet]

Yes, all of the videos started from same numbers (e.g. 08 in 08_003) should be in the same folder named as 08.

Those particular two digits refer to task/scene number.

For Task 8:
\Training\frames\08\08_001\*.jpg
\Training\frames\08\08_002\*.jpg

For Task 7:
\Training\frames\07\07_001\*.jpg
\Training\frames\07\07_002\*.jpg

[abhishekaich27]

Thank you! I will restructure the folders this way then.

Do you think this is the new train-test split authors propose in the paper? Were you able to reproduce the results of the paper for the Shanghai-Tech dataset?

[AbdulMoqeet]

I am not sure. I had to do reverse engineering to make this code work. So, by tracing the errors and codes, I came to that directory structure.

I never meant to reproduce the quantitative results. However, I check some of qualitative results, there were positive results that ensure that code is working fine. I had to use this code for some other industrial application (for experimental purpose) that I am working on.

[abhishekaich27]

Alright. I have emailed the authors regarding this dataset's training/testing split and one of the authors has replied to get back to me in a while. I will update here if I get anything concrete.

Thank you again for your kind help.

[abhishekaich27]

@AbdulMoqeet The author of the repo confirmed that the folder structure you have is correct.

[AbdulMoqeet]

@abhishekaich27 Thanks a lot for the confirmation.

[BenedictGreen]

@AbdulMoqeet Hi, AbdulMoqeet. Thanks a lot for your conversation, my 'train.py' code is already running. Have you solved the problem of how to test the model?

[zugexiaodui]

Same question

[accountyangaaa]

@AbdulMoqeet Dear AbdulMoqeet, I would like to ask a question about train.py. When I ran it , this line of the code(
main(k_shots, num_tasks, adam_betas, gen_lr, dis_lr, total_epochs, model_folder_path) will raise the error(TypeError: add(): argument 'other' (position 1) must be Tensor, not BCELoss). Do you know how to fix it? Thank you.
Can you share your environment requirement？Thank you.

[AbdulMoqeet]

@accountyangaaa Could you please run with default arguments? I have also shared a working code above. Check my answer to @carolchenyx Please share a screenshot.

[accountyangaaa]

@AbdulMoqeet Dear AbdulMoqeet, thanks a lot for your reply. I just ran the working code above. There is another error.
RuntimeError: CUDA out of memory. Tried to allocate 32.00 MiB (GPU 0; 4.00 GiB total capacity; 3.19 GiB already allocated; 0 bytes free; 28.04 MiB cached)
Do you know how to fix it? Thank you.

[carolchenyx]

Hi, maybe you can try to change to a smaller batch size. accountyangaaa ***@***.***> 于2021年11月22日周一 下午4:00写道：

…

@AbdulMoqeet <https://github.com/AbdulMoqeet> Thanks a lot for your reply.I just ran the working code above.There is another problem.Do you know how to fix it? Thank you. The screenshot is as follows. [image: image] <https://user-images.githubusercontent.com/56595279/142824356-368ea4ea-adf0-4d31-8b43-634650b1af81.png> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#9 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AL2PSUEP3NGCL54UORY7ZLDUNH2DBANCNFSM447O6GHA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

[accountyangaaa]

@carolchenyx Thanks a lot for your reply. Can I see your dataset structure

[accountyangaaa]

@carolchenyx Hi，below is the structure I have. Is this correct?

[carolchenyx]

@carolchenyx Hi，below is the structure I have. Is this correct?

Hi, I think it's correct.

[dekathomas]

Hello @AbdulMoqeet @carolchenyx, where do you know the version list of the dependencies? Have you got information to test the trained model? Thank you

[AbdulMoqeet]

@dekathomas I did by trial & error.

For testing, I made changes in the given code by following the paper. You can modify the train code accordingly. I tested on my private dataset at that time and, the results were unsatisfactory.

[dekathomas]

Can you share to me the testing code you have made?

[AbdulMoqeet]

Sorry, I am not allowed to do so :-(

[dekathomas]

Owh oke, no problem

[lzyuan168]

@AbdulMoqeet Hi are you able to provide a skeleton code for the test.py file without the sensitive information. Thanks

[AbdulMoqeet]

@lzyuan168 Hi, I am sorry, I do not have access to those files anymore.

As it's been long time so I do not even remember about them. If I remember correctly, I used train file/conditions to set conditions in test file

[lzyuan168]

@AbdulMoqeet alright thank you