This is a code repository for work in progress paper Visualizing Neural netwroks Imagination

Launch on cloud

chmod +x launch_cloud.sh
./launch_cloud.sh gol3ts.yaml job_name

You might have to change the --region flag in launch_cloud to get it to schedule. Sometimes jobs stay Queued for 10 hours. See here for regions with p100s or whatever gpu type is in cloud_config.yaml

Train local

 python3 -m train.train --job_dir something

The models and flagfile are saved after regular training and before adversarial training.

Tensorboard

Replace these paths with the ones you want to see

python3 -m tensorboard.main --logdir_spec 0:gs://visualize-nn-imagination-runs/full_hp_gpu_603721,1:gs://visualize-nn-imagination-runs/full_hp_gpu_p100_605262

Get images

This will add the images to a img directory in gs://visualize-nn-imagination-runs/full_hp_gpu_p100_605262

python3 -m train.save_model_result_imgs --is_hp_serach_root --job_dir gs://visualize-nn-imagination-runs/cc_count_3ts_5x5board_3875040

Download them with

gsutil cp -r gs://visualize-nn-imagination-runs/cc_count2ts_dec_enc_fix_8335585/imgs/* /Users/wichersn/visualize_nn_runs/cc_count2ts_dec_enc_fix_8335585

Disclaimer

This is not an officially supported Google product