#!/bin/bash
# Phase 3: scale-up. Warm-start from P1 (weights-only init), train on ~337h
# (IndicVoices-hi 307h + keywords + svarah + numbers) with x3 speed perturbation.
# bf16, pure CTC, causal streaming, fixed 103-char vocab.
set -e
source /workspace/venvs/icefall/bin/activate
export PYTHONPATH=/root/icefall:$PYTHONPATH
cd /root/icefall/egs/hindi/ASR
ROOT=/workspace/hindi_ft/asr_ctc
python zipformer/train.py \
  --world-size 1 \
  --num-epochs 15 \
  --start-epoch 1 \
  --init-ckpt $ROOT/exp_p1/epoch-30.pt \
  --exp-dir $ROOT/exp_p3 \
  --lang-dir $ROOT/data/lang_char \
  --manifest-dir $ROOT/data/fbank_p3 \
  --use-ctc 1 --use-transducer 0 --ctc-loss-scale 1.0 \
  --causal 1 \
  --chunk-size "16,32,64,-1" \
  --left-context-frames "64,128,256,-1" \
  --max-duration 300 \
  --use-fp16 0 \
  --num-workers 8 \
  --enable-musan False \
  --on-the-fly-feats True \
  --base-lr 0.02   # warm-start lr; nan at batch~350 was a bad data sample (recurs at any lr), handled by skip-nan guard in train.py