gpu support for train_generator.py

[CW-Huang]

@51N84D can you assess the need for this? for the smaller datasets cpu will be enough, and it's fairly fast. I'm not sure for bigger datasets (e.g. LBIDD) whether it will take too much time.

[51N84D]

Yeah, I'll check it out

[51N84D]

@CW-Huang I'm getting NaN losses trying to run an experiment with lbidd:

2020-08-30 23:56:24.873470 / Iteration 200: nan nan 2020-08-30 23:56:25.062086 / Iteration 200 valid loss nan

Here is the script I'm using: python train_generator.py --dataroot=/Users/Sunsmeister/Desktop/Research/Causal/causal-benchmark/datasets --data=lbidd_quadratic_50k --dist='LogLogistic' --dim_h=256 --n_hidden_layers=2 --batch_size=128 --lr=0.0005

Have you been able to successfully run an experiment on lbidd? If so, could you share your script / params?

[CW-Huang]

that's probably because log-logistic is supported on R_>0, and lbidd has data <=0.

[51N84D]

Good call, that was it. I'm able to run an exp in ~8 min on CPU with this script:

python train_generator.py --dataroot=/Users/Sunsmeister/Desktop/Research/Causal/causal-benchmark/datasets --data=lbidd_quadratic_50k --dist='FactorialGaussian' --dim_h=256 --n_hidden_layers=2 --batch_size=128 --lr=0.0005

[51N84D]

Is the code currently set up to train on GPU? If not, I can work on that