cjrd / moco

AutoSelf

Example 4-gpu execution/

CUDA_VISIBLE_DEVICES=0,2,5,7 ./4-gpu-end-to-end.sh $PWD/checkpoints/ /path/to/cifar-10/ 'Initial test run'

To not report a run to wandb do

WANDB_MODE=dryrun CUDA_VISIBLE_DEVICES=0,2,5,7 ./4-gpu-end-to-end.sh $PWD/checkpoints/ /path/to/cifar-10/ 'Initial test run'

Open questions:

• What learning rate and temperature should we use for moco?
• What we use the cosine annealing learning scheduler?
• Should our other project heads all have an MLP?
• How can we simulataneously measure the linear classifier performance throughout training? i.e. how should we merge the two scripts?

TODO:

• Save the final/best models to wandb for each run.