amir-abdi / disentanglement-pytorch Goto Github PK

Thank you Amir for a great library.

You mentioned i could ask questions by raising them as issues. Do you know how much of the Google-Research disentanglement lab in tensorflow is implemented in your package? I prefer working in pytorch and was hoping to use your library as a substitute.

Thank you very much!

Br,
Johan Ye

When I trying to run scripts/dsprite_FactorVAE.sh
The following error happened:
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [1000, 2]], which is output 0 of TBackward, is at version 2; expected version 1 instead.

Which seems the problem with the discriminator. But even if I changed the inplace flag in the nn.LeakyReLU(0.2,inplace=False), the problem still exsits

Thanks Amir for providing this library.
I was looking at the original b-tcvae github code:
https://github.com/rtqichen/beta-tcvae

And I believe they did not map the encoding to mu and sigma layer, it seems that they only mapped it to mu. But in your code, it seems that you're following the VAE structure (mapping the input to mu and sigma layers). I was wondering if there is any specific reason for including/excluding the sigma layer?

Best,
Rojin

Dependabot couldn't authenticate with https://pypi.python.org/simple/.

You can provide authentication details in your Dependabot dashboard by clicking into the account menu (in the top right) and selecting 'Config variables'.

View the update logs.

error message:
mu, logvar = self.model.encode(x=x_true1, c=label1)
ValueError: too many values to unpack (expected 2)
Solution:
Using xxxGaussianConv64

Hi,

Thank you very much for setting this library ! If I am correct, one can use a custom dataset by creating the corresponding data_loader, and updating accordingly the function:

This will run with the training and evaluation metrics code? There is nothing specific to dSprites and CelebA?

Hi, thanks for posting this very convenient tool to do experiments on disentanglement learning

When I trained the model, I met the following problem:

I use the --evaluate_metric mig sap_score irs factor_vae_metric dci during training the BetaVAE on CelebA dataset.

However, I get

anaconda3/lib/python3.7/site-packages/disentanglement_lib/data/ground_truth/named_data.py", line 65, in get_named_ground_truth_data
raise ValueError("Invalid data set name.")
ValueError: Invalid data set name.
In call to configurable 'dataset' (<function get_named_ground_truth_data at 0x7f8eed2a13b0>)
In call to configurable 'evaluation' (<function evaluate at 0x7f8e6e5558c0>)

I check the 'named_data.py' file and find out 'celebA' is not in the named data list (containing dSprites, 3dshapes, mpi3d, car3d, smallnorb).

Is there any ways that can let the model trained on 'celebA' be evaluated by disentanglement metrics?

Thanks