| command: | |
| - python3 | |
| - ${program} | |
| - --do_train | |
| - --do_eval | |
| - --use_scan | |
| - --gradient_checkpointing | |
| - --overwrite_output_dir | |
| - --predict_with_generate | |
| - --freeze_encoder | |
| - --streaming | |
| - --use_auth_token | |
| - --compilation_cache | |
| - --return_timestamps | |
| - ${args} | |
| method: grid | |
| metric: | |
| goal: minimize | |
| name: eval/wer | |
| parameters: | |
| model_name_or_path: | |
| value: distil-whisper/large-32-2 | |
| teacher_model_name_or_path: | |
| value: openai/whisper-large-v2 | |
| train_dataset_name: | |
| value: librispeech_asr-timestamped+librispeech_asr-timestamped+librispeech_asr-timestamped+common_voice_13_0-timestamped+voxpopuli-timestamped+ami-ihm-timestamped+ami-sdm-timestamped+peoples_speech-clean-timestamped+tedlium-timestamped+switchboard-data+gigaspeech-l-timestamped+spgispeech-timestamped | |
| train_dataset_config_name: | |
| value: all+all+all+en+en+ihm+sdm+clean+release3+all+l+L | |
| train_split_name: | |
| value: train.clean.100+train.clean.360+train.other.500+train+train+train+train+train+train+train+train+train | |
| train_dataset_samples: | |
| value: 2.9+10.4+14.9+89+18.2+10.9+10.9+288+26.8+371.2+226.6+192.7 | |
| timestamp_probability: | |
| values: | |
| - 0.0 | |
| - 0.2 | |
| - 0.4 | |
| - 0.6 | |
| - 0.8 | |
| - 1.0 | |
| round_timestamps: | |
| values: | |
| - True | |
| - False | |
| eval_dataset_name: | |
| value: "distil-whisper/gigaspeech-l" | |
| eval_dataset_config_name: | |
| value: "l" | |
| cache_dir: | |
| value: /home/sanchitgandhi/.cache | |
| dataset_cache_dir: | |
| value: /home/sanchitgandhi/.cache | |
| output_dir: | |
| value: ./ | |
| per_device_train_batch_size: | |
| value: 64 | |
| dtype: | |
| value: bfloat16 | |
| learning_rate: | |
| value: 1e-4 | |
| lr_scheduler_type: | |
| value: constant_with_warmup | |
| warmup_steps: | |
| value: 50 | |
| max_steps: | |
| value: 2500 | |
| save_steps: | |
| value: 2501 # don't save checkpoints during sweep | |
| dataloader_num_workers: | |
| value: 48 | |
| logging_steps: | |
| value: 25 | |
| wer_threshold: | |
| value: 10 | |
| program: run_distillation.py | |
| project: distil-whisper-sweeps | |