Skip to content

Latest commit

 

History

History
51 lines (29 loc) · 5.24 KB

commands.md

File metadata and controls

51 lines (29 loc) · 5.24 KB
Raw

Commands to run the experiments from the paper

This file includes a list of example commands to train and evaluate our models to help reproduce the experiments from the paper:

  • Original game setting, 

    python commgame.py train --dataset STL10 --output stl10-original-model-lgame --loss HingeLoss --sigma2 5e-4 --nlines 20 --learning-rate 0.0001 --dataset-root ./data --imagenet-weights --freeze-vgg --imagenet-norm --epochs 250 --invert --feature-layer relu5_3  --batch-size 100 --data-seed 0

python commgame.py eval --dataset STL10 --output stl10-original-model-lgame-eval --loss HingeLoss --sigma2 5e-4 --nlines 20 --dataset-root ./data --imagenet-weights --freeze-vgg --imagenet-norm --invert --feature-layer relu5_3  --batch-size 100 --data-seed 0 --weights stl10-original-model-lgame/model_final.pt

  • Original game setting, 

    python commgame.py train --dataset STL10 --output stl10-original-model-lpercep --loss HingeLoss --sigma2 5e-4 --nlines 20 --learning-rate 0.0001 --dataset-root ./data --imagenet-weights --freeze-vgg --imagenet-norm --epochs 250 --invert --feature-layer relu5_3 --sort-of-perceptual-loss --batch-size 100 --data-seed 0

python commgame.py eval --dataset STL10 --output stl10-original-model-lpercep-eval --loss HingeLoss --sigma2 5e-4 --nlines 20 --dataset-root ./data --imagenet-weights --freeze-vgg --imagenet-norm --invert --feature-layer relu5_3 --sort-of-perceptual-loss --batch-size 100 --data-seed 0 --weights stl10-original-model-lpercep/model_final.pt

  • Object-oriented same, 

    python commgame.py train --dataset STL10 --output stl10-oosame-model-lgame --loss HingeLoss --sigma2 5e-4 --nlines 20 --learning-rate 0.0001 --dataset-root ./data --imagenet-weights --freeze-vgg --imagenet-norm --epochs 250 --invert --feature-layer relu5_3 --object-oriented same --batch-size 100 --data-seed 0

python commgame.py eval --dataset STL10  --output stl10-oosame-model-lgame-eval --loss HingeLoss --sigma2 5e-4 --nlines 20 --dataset-root ./data --imagenet-weights --freeze-vgg --imagenet-norm --invert --feature-layer relu5_3 --object-oriented same --batch-size 100 --data-seed 0 --weights stl10-oosame-model-lgame/model_final.pt

  • Object-oriented different, 

    python commgame.py train --dataset STL10 --output stl10-oodiff-model-lgame --loss HingeLoss --sigma2 5e-4 --nlines 20 --learning-rate 0.0001 --dataset-root ./data --imagenet-weights --freeze-vgg --imagenet-norm --epochs 250 --invert --feature-layer relu5_3 --object-oriented different --batch-size 100 --data-seed 0

python commgame.py eval --dataset STL10  --output stl10-oodiff-model-lgame-eval --loss HingeLoss --sigma2 5e-4 --nlines 20 --dataset-root ./data --imagenet-weights --freeze-vgg --imagenet-norm --invert --feature-layer relu5_3 --object-oriented different --batch-size 100 --data-seed 0 --weights stl10-oodiff-model-lgame/model_final.pt

For the object-oriented games with the objective , just add the --sort-of-perceptual-loss argument to the train/eval commands above.

  • Weighting the feature maps that contribute to the perceptual loss - example of using only the features out of the first convolution block:

    python commgame.py train --dataset STL10 --output stl10-original-sopweights10000 --loss HingeLoss --sigma2 5e-4 --nlines 20 --learning-rate 0.0001 --dataset-root ./data --imagenet-weights --freeze-vgg --imagenet-norm --epochs 250 --invert --feature-layer relu5_3 --sort-of-perceptual-loss --data-seed 0 --batch-size 100 --sop-weights 1 0 0 0 0

  • Varying model capacity, example on Caltech-101 128px images:

    python commgame.py train --dataset Caltech101 --output caltech101-oodiff-model-wide --loss HingeLoss --sigma2 5e-4 --nlines 20 --learning-rate 0.0001 --dataset-root ./data --imagenet-weights --freeze-vgg --imagenet-norm --epochs 250 --invert --feature-layer relu5_3 --object-oriented different --sort-of-perceptual-loss --latent-size 1024 --rx-hidden 1024 --decoder-hidden 1024 --decoder-hidden2 1024 --weight-decay 0.0001 --batch-size 100 --data-seed 0

  • Inducing shape bias into the model by using the weights of a VGG-16 pretrained on Stylized ImageNet (instead of the --imagenet, use --sin-weights):

    python commgame.py train --dataset CelebA --output celebA-model-original-sinweights --loss HingeLoss --sigma2 5e-4 --nlines 20 --learning-rate 0.0001 --dataset-root ./data --sin-weights --freeze-vgg --imagenet-norm --epochs 250 --invert --feature-layer relu5_3 --sort-of-perceptual-loss --data-seed 0 --batch-size 100 --image-size 112

Some general arguments you can control:

[--sort-of-perceptual-loss] to use the additional perceptual loss
[--sop-weights] to specify sort-of-perceptual-loss weights for the different feature maps, default is [1, 1, 1, 1, 1]
[--sop-coef] to control the coeffiecient of the perceptual loss
[--device]
[--data-seed]
[--snapshot-interval]
[--num-reconstructions]