Skip to content

Latest commit

 

History

History
52 lines (43 loc) · 1.28 KB

LOG.md

File metadata and controls

52 lines (43 loc) · 1.28 KB
Raw

Experiment Logs

Benchmark on Standard Models (Table 2)

We use the standard ResNet50 pretrained on ImageNet.

# Baselines
wandb sweep sweeps/continual_IN.yaml
# MECTA
wandb sweep sweeps/continual_IN_mecta_std_batch.yaml

Cache Constrained Benchmarks (Table 1)

wandb sweep sweeps/continual_IN_mecta.yaml
wandb sweep sweeps/continual_cifar10_mecta.yaml
wandb sweep sweeps/continual_cifar100_mecta.yaml

Ablation study on beta

Ablation on the forget gate with different CTA methods.

wandb sweep sweeps/continual_IN_mecta_ablate_gate.yaml
# => [03/03] https://wandb.ai/jyhong/MECTA_release/sweeps/7jlh4hm7
#    @GPU9 ablate gate on three methods again.

Ablation studies.

  • (B) Abalation on the batch size
wandb sweep sweeps/continual_IN_ablation_batch.yaml
# => [03/03] https://wandb.ai/jyhong/MECTA_release/sweeps/5o2wcgi5
#    @GPU9 ablation on batch size.
  • (L) Abalation on beta_thre
wandb sweep sweeps/continual_IN_ablation_beta_thre.yaml
# => [03/03] https://wandb.ai/jyhong/MECTA_release/sweeps/hy3fkd0j
#    @GPU9 ablation on beta_thre
  • (C) Prune channels
wandb sweep sweeps/continual_IN_ablation_prune.yaml
# => [03/03] https://wandb.ai/jyhong/MECTA_release/sweeps/fwbdlkoh
#    @GPU9

Pair-eval