Reproducing cifar10 results

[diane-wagner]

Hi Fábio,

I merged my branch with the Master and I have successfully run the following tests:

CIFAR10 Pretraining

python -m metassl.train_simsiam --config "metassl/default_metassl_config_cifar10.yaml" --use_fixed_args --data.dataset_percentage_usage 25 --train.epochs 5 --expt.warmup_epochs 0 --expt.seed 0 --expt.save_model_frequency 50 --expt.is_non_grad_based --expt.multiprocessing_distributed --model.arch "baseline_resnet" --simsiam.use_baselines_loss --expt.expt_name testing_cifar10_pretraining_1

CIFAR10 Finetuning

python -m metassl.train_linear_classifier_simsiam --config "metassl/default_metassl_config_cifar10.yaml" --use_fixed_args --data.dataset_percentage_usage 25 --finetuning.epochs 5 --expt.warmup_epochs 0 --expt.seed 0 --expt.save_model_frequency 10 --expt.is_non_grad_based --expt.multiprocessing_distributed --model.arch baseline_resnet --expt.expt_name testing_cifar10_finetuning_1 --expt.target_model_checkpoint_path "/work/dlclarge2/wagnerd-metassl-experiments/metassl/CIFAR10/testing_cifar10_pretraining_1/checkpoint_0004.pth.tar"

ImageNet Pretraining

python -m metassl.train_simsiam --config "metassl/default_metassl_config_imagenet.yaml" --use_fixed_args --data.dataset_percentage_usage 0.1 --train.epochs 2 --train.batch_size 4 --expt.warmup_epochs 0 --expt.seed 0 --expt.save_model_frequency 50 --expt.is_non_grad_based --expt.multiprocessing_distributed --model.arch "baseline_resnet" --simsiam.use_baselines_loss --expt.expt_name testing_imagenet_pretraining_1

ImageNet Finetuning

python -m metassl.train_linear_classifier_simsiam --config "metassl/default_metassl_config_imagenet.yaml" --use_fixed_args --data.dataset_percentage_usage 0.1 --finetuning.epochs 2 --finetuning.batch_size 4 --expt.warmup_epochs 0 --expt.seed 0 --expt.save_model_frequency 10 --expt.is_non_grad_based --expt.multiprocessing_distributed --model.arch baseline_resnet --expt.expt_name testing_imagenet_finetuning_1 --expt.ssl_model_checkpoint_path "/work/dlclarge2/wagnerd-metassl-experiments/metassl/ImageNet/testing_imagenet_pretraining_1/checkpoint_0001.pth.tar"

CIFAR10 Alternating

python -m metassl.train_alternating_simsiam --config "metassl/default_metassl_config_cifar10.yaml" --use_fixed_args --data.dataset_percentage_usage 25 --train.epochs 5 --finetuning.epochs 5 --expt.expt_name testing_cifar10_alternating_2

ImageNet Alternating

python -m metassl.train_alternating_simsiam --config "metassl/default_metassl_config_imagenet.yaml" --use_fixed_args --data.dataset_percentage_usage 0.1 --train.batch_size 4 --finetuning.batch_size 2 --train.epochs 2 --finetuning.epochs 2 --expt.expt_name testing_imagenet_alternating_1

Notes:

• In order to run 'CIFAR10 Alternating' I needed to comment out 'get_image_data_to_plot' in train_alternating_simsiam.py > Please check commit: 852cd2b1bbb278a4b5deeb9867927248aad3f4f5
• SimSiam CIFAR10 results get reproduced with the Pretraining and Finetuning scripts :D
• The changes to get reproducible CIFAR10 results are currently not integrated into the alternating script. I use another model that is not yet adapted to run it in an alternating manner.
• I integrated NEPS ('Danny's new BOHB')
• NPES currently requires Python 3.7 (so I needed to adapt the code for it)
• The code urgently needs some cleaning/refactoring! I will do this on a separate branch.

Warm regards

Diane