fpcasale / gppvae Goto Github PK

Hi,

I am trying to reproduce the rotating MNIST results in your GPPVAE paper. I was wondering if you somehow saved the trained networks. More specifically, I would like to give the same image as input to your model as well as mine, and visually compare both reconstructions. Is there a way of doing this except training the model from scratch?

Hello,

I am trying to reproduce your results. I trained VAE for 10000 iterations (using train_vae.py with default parameters). Then, I tried passing VAE weights from 1000, 5000 and 9000 iteration checkpoints as an input for the train_gppvae.py script.

Am I correct in saying that callback.py does the following: it prints a grid of three rows of images, where the first row is ground truth, the second row is VAE reconstruction and the third row is GPPVAE reconstruction?

If my interpretation is correct, then each time GPPVAE yields worse results than VAE and I don't understand why. Could you, please, mention in the readme the exact number of iterations you used for training the VAE? And the exact parameters you were passing to both scripts (train_vae.py and train_gppvae.py) if those differ from the defaults. Also, the kernel heatmaps look different from those presented in your paper, that's what also concerns me.

Thank you in advance!

python preprocess_data.py --datadir /home/ali/data
python: can't open file 'preprocess_data.py': [Errno 2] No such file or directory

Thanks for your great contribution to NIPS paper, GPPVA.

I have a question to run your code on missing data. Consider 30 degree left rotation is missed for one person in our dataset and I want to predict the missing data using available rotations.
Which available rotation should I use for prediction? Or which available data is preferred for the prediction of unavailable one?

Hi can you please add the file
Or tell me how to make one of my own?
@fpcasale

The link to download the dataset is down, it leads to a 404 page.