Add resume_config.json
Browse files- resume_config.json +25 -0
resume_config.json
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_name": "openai/whisper-tiny",
|
| 3 |
+
"language": "urdu",
|
| 4 |
+
"task": "transcribe",
|
| 5 |
+
"batch_size": 8,
|
| 6 |
+
"grad_accum_steps": 2,
|
| 7 |
+
"effective_batch": 16,
|
| 8 |
+
"learning_rate": 3e-06,
|
| 9 |
+
"warmup_steps": 200,
|
| 10 |
+
"num_epochs": 20,
|
| 11 |
+
"early_stopping_patience": 3,
|
| 12 |
+
"eval_every_steps": 500,
|
| 13 |
+
"save_every_steps": 500,
|
| 14 |
+
"target_sr": 16000,
|
| 15 |
+
"max_audio_length_s": 30,
|
| 16 |
+
"fp16": true,
|
| 17 |
+
"frozen_layers": [
|
| 18 |
+
"model.encoder.conv1",
|
| 19 |
+
"model.encoder.conv2"
|
| 20 |
+
],
|
| 21 |
+
"optimizer": "AdamW (via Seq2SeqTrainer)",
|
| 22 |
+
"scheduler": "linear with warmup",
|
| 23 |
+
"generation_max_length": 225,
|
| 24 |
+
"notes": "To resume fine-tuning, load this model with WhisperForConditionalGeneration.from_pretrained(repo_id) and use the same Seq2SeqTrainer config. Re-freeze conv1/conv2 if continuing on the same dataset; unfreeze them if adapting to a new domain."
|
| 25 |
+
}
|