G_loss and D_loss are negative numbers when running train_wgangp.py using my own image dataset

[nene211]

G_loss and D_loss are negative numbers when running train_wgangp.py using my own image dataset

[A03ki]

Hi, @nene211.
Thank you for using this repository.
I see you have a few questions, so I'll answer them in order.

When I use my own image dataset to train the WGAN, both of its G_loss and D_loss are negative numbers, I wonder if this situation is right?

Taking negative values is not necessarily a false result. Please check the two losses continue to change stably. (It may be hard work to check the fluctuation of the two losses in this repository.)

And Im not really understand what the class mean, for example, I have 100 normal images and 100 ano images, so I put 100 normal images in the path "your own dataset-dataset-train-class0", and put the 100 ano images in the path"your own dataset-dataset-test-class0", then I run wgangp.py, izif.py, detection.py and save.py one by one?

Yes! You can get the training and test results by executing the following commands in order.

python train_wgangp.py "your own dataset-dataset-train-class0"
python train_encoder_izif.py "your own dataset-dataset-train-class0"
python test_anomaly_detection.py "your own dataset-dataset-train-class0"
python save_compared_images.py "your own dataset-dataset-train-class0" --n_iters 0 --n_grid_lines 10

and the size of my pictures are all 64*64, so...is something wrong in my dataset?

That image size is fine.

I hope that answers your questions.

[nene211]

thank you so much! It runs well on my image dataset and the accuracy is high! Thank you!!!

[A03ki]

Hi, @nene211. Thank you for your question.
That's correct. If your image is not 64×64, it will be automatically resized to the default value (64×64).

[nene211]

OkOk，thank you again for your excellent codes and answers, you really help me a lot!

[nene211]

and the size of my pictures are all 64*64, so...is something wrong in my dataset?

[nene211]

well...i get a high accurary, the only problem is the loss...

[nene211]

Well I have another question now...@A03ki
I want to know that if my image is not 64*64, should i change some parameters of the model? Or i just put my images in the train/test folders and run without changing anyting of the codes?