Before running training please adjust checkpoint_dir
variable.
- SA-MIRI's network:
sbatch job_cte_nostrategy.sh --architecture sa_miri --epochs 20 --batch_size 128
sbatch job_cte_mirroredstrategy.sh --architecture sa_miri --epochs 20 --batch_size 128
sbatch job_cte_multiworkermirroredstrategy.sh --architecture sa_miri --epochs 20 --batch_size 128
- MobileNet:
sbatch job_cte_nostrategy.sh --architecture mobilenet --epochs 5 --batch_size 128
sbatch job_cte_mirroredstrategy.sh --architecture mobilenet --epochs 5 --batch_size 128
sbatch job_cte_multiworkermirroredstrategy.sh --architecture mobilenet --epochs 5 --batch_size 128
- Resnet101:
sbatch job_cte_nostrategy.sh --architecture resnet101 --epochs 5 --batch_size 64
sbatch job_cte_mirroredstrategy.sh --architecture resnet101 --epochs 5 --batch_size 64
sbatch job_cte_multiworkermirroredstrategy.sh --architecture resnet101 --epochs 5 --batch_size 64
Evaluations made intra-node (using 2-4 GPUs) were using NCCLAllReduce
, evaluations made inter-node (using 8+ GPUs) were using RING
collective communication.
It would be interesting to measure speedup for MultiWorkerMirroredStrategy
with NCCL
set as collective communication mechanism.