Skip to content

ablations.md

Latest commit

 

History

History
113 lines (89 loc) · 4.03 KB

ablations.md

File metadata and controls

113 lines (89 loc) · 4.03 KB

Ablations for Barkopedia + ECAPA-TDNN

Common Configuration

COMMON="writer.logger.mode=online sample_rate=8000 trainer.compile.enabled=true trainer.compile.call._target_=torch.compile trainer.compile.call.mode=default trainer.compile.call.fullgraph=false trainer.compile.call.dynamic=false trainer.compile.call.backend=cudagraphs trainer.amp.enabled=false -cn=x_vector"

1) Ablation #1: Loss Function Parameters (margin m, scale s)

m=0.1, s=30

MPLBACKEND=Agg uv run -- python train.py loss_function.margin=0.1 loss_function.scale=30 trainer.save_dir=saved/ablation/loss_m0.1_s30 writer.logger.run_name="loss_m0.1_s30" $COMMON

m=0.1, s=40

MPLBACKEND=Agg uv run -- python train.py loss_function.margin=0.1 loss_function.scale=40 trainer.save_dir=saved/ablation/loss_m0.1_s40 writer.logger.run_name="loss_m0.1_s40" $COMMON

m=0.1, s=50

MPLBACKEND=Agg uv run -- python train.py loss_function.margin=0.1 loss_function.scale=50 trainer.save_dir=saved/ablation/loss_m0.1_s50 writer.logger.run_name="loss_m0.1_s50" $COMMON

m=0.2, s=30

MPLBACKEND=Agg uv run -- python train.py loss_function.margin=0.2 loss_function.scale=30 trainer.save_dir=saved/ablation/loss_m0.2_s30 writer.logger.run_name="loss_m0.2_s30" $COMMON

m=0.2, s=40

MPLBACKEND=Agg uv run -- python train.py loss_function.margin=0.2 loss_function.scale=40 trainer.save_dir=saved/ablation/loss_m0.2_s40 writer.logger.run_name="loss_m0.2_s40" $COMMON

m=0.2, s=50

MPLBACKEND=Agg uv run -- python train.py loss_function.margin=0.2 loss_function.scale=50 trainer.save_dir=saved/ablation/loss_m0.2_s50 writer.logger.run_name="loss_m0.2_s50" $COMMON

m=0.3, s=30

MPLBACKEND=Agg uv run -- python train.py loss_function.margin=0.3 loss_function.scale=30 trainer.save_dir=saved/ablation/loss_m0.3_s30 writer.logger.run_name="loss_m0.3_s30" $COMMON

m=0.3, s=50

MPLBACKEND=Agg uv run -- python train.py loss_function.margin=0.3 loss_function.scale=50 trainer.save_dir=saved/ablation/loss_m0.3_s50 writer.logger.run_name="loss_m0.3_s50" $COMMON

2) Ablation #2: Batch Sampler On/Off

Batch Sampler On

MPLBACKEND=Agg uv run -- python train.py batch_sampler=hpm trainer.save_dir=saved/ablation/batch_sampler_on writer.logger.run_name="batch_sampler_on" $COMMON

Batch Sampler Off

MPLBACKEND=Agg uv run -- python train.py ~batch_sampler trainer.save_dir=saved/ablation/batch_sampler_off writer.logger.run_name="batch_sampler_off" $COMMON

3) Ablation #3: Backend PLDA On

MPLBACKEND=Agg uv run -- python train.py backends=plda trainer.save_dir=saved/ablation/plda_on writer.logger.run_name="plda_on" $COMMON

4) Ablation #4: Augmentation Types

No Augmentation

MPLBACKEND=Agg uv run -- python train.py transforms.instance_transforms=basic trainer.save_dir=saved/ablation/aug_none writer.logger.run_name="aug_none" $COMMON

MFCC2 Augmentation (Musan + RIR)

MPLBACKEND=Agg uv run -- python train.py transforms.instance_transforms=mfcc_2 trainer.save_dir=saved/ablation/aug_mfcc2 writer.logger.run_name="aug_mfcc2" $COMMON

Full Augmentation

MPLBACKEND=Agg uv run -- python train.py transforms.instance_transforms=augmented trainer.save_dir=saved/ablation/aug_augmented writer.logger.run_name="aug_augmented" $COMMON

5) Final Run: ECAPA-TDNN with Best Hyperparameters + AMP

MPLBACKEND=Agg uv run -- python train.py --config-name ecappa_tdnn \
  sample_rate=8000 \
  batch_sampler=hpm \
  backends=plda \
  loss_function.margin=0.2 \
  loss_function.scale=40 \
  transforms.instance_transforms=augmented \
  trainer.compile.enabled=true \
  trainer.compile.call._target_=torch.compile \
  trainer.compile.call.mode=default \
  trainer.compile.call.fullgraph=false \
  trainer.compile.call.dynamic=false \
  trainer.compile.call.backend=inductor \
  trainer.amp.enabled=true \
  trainer.amp.dtype=float16 \
  trainer.save_dir=saved/best_ecappa_tdnn \
  writer.logger.run_name="best_ecappa_tdnn"