search

openai / requests-for-research

A living collection of deep learning problems
https://openai.com/requests-for-research
1.69k stars 609 forks source link

I added a link to my version of Inverse DRAW #28

Closed dojoteef closed 7 years ago

dojoteef commented 7 years ago

Hi! Thanks for providing an interesting list of research requests. It is nice to have a set of research ideas from prominent researchers to explore. I have included a link to an implementation of an Inverse DRAW model. The model was implemented in TensorFlow v0.12 and can be trained running:

python glas/train.py --defaults_file config/defaults.json --config_file config/mnist_binarized.json --log_dir train/ --datasubset train

The model can be evaluated using either the validation or test data by issuing the following commands:

python glas/eval.py --defaults_file config/defaults.json --config_file config/mnist_binarized.json --checkpoint_path train/ --log_dir eval/ --datasubset validate
python glas/eval.py --defaults_file config/defaults.json --config_file config/mnist_binarized.json --checkpoint_path train/ --log_dir eval/ --datasubset test --once

Images can be generated from the model like so:

python glas/generate.py --defaults_file config/defaults.json --config_file config/mnist_binarized.json --checkpoint_path train/ --log_dir generate/

In order to view the model images and training metrics you will need to run TensorBoard specifying the appropriate log directory:

tensorboard --logdir=<train|eval|generate>

Additional configuration can be specified on the command line. Command line options can be listed as:

python glas/train.py --help
python glas/eval.py --help
python glas/generate.py --help

If you have any questions, just let me know!