infer_batch.py

import glob
import os
import sys
import time

from diarizen.pipelines.inference import DiariZenPipeline

"""
Usage: use with two processes.
cd /workspace/DiariZen
python3 infer_batch.py /workspace/MERLIon_audio /workspace/MERLIon_rttm_diarizen 0 2 > /workspace/MERLIon_shard0.log 2>&1 &
python3 infer_batch.py /workspace/MERLIon_audio /workspace/MERLIon_rttm_diarizen 1 2 > /workspace/MERLIon_shard1.log 2>&1 &
"""

audio_dir, rttm_out_dir = sys.argv[1], sys.argv[2]
shard_id, num_shards = (int(sys.argv[3]), int(sys.argv[4])) if len(sys.argv) == 5 else (0, 1)

os.makedirs(rttm_out_dir, exist_ok=True)

diar_pipeline = DiariZenPipeline.from_pretrained(
    "BUT-FIT/diarizen-wavlm-large-s80-md-v2",
    rttm_out_dir=rttm_out_dir,
)

wavs = sorted(glob.glob(os.path.join(audio_dir, "*.wav")))[shard_id::num_shards]
print(f"shard {shard_id}/{num_shards}: {len(wavs)} wav files", flush=True)

for i, wav in enumerate(wavs, 1):
    sess = os.path.splitext(os.path.basename(wav))[0]
    rttm_path = os.path.join(rttm_out_dir, f"{sess}.rttm")
    if os.path.exists(rttm_path) and os.path.getsize(rttm_path) > 0:
        print(f"[{i}/{len(wavs)}] {sess} already done, skip", flush=True)
        continue
    t0 = time.time()
    diar_pipeline(wav, sess_name=sess)
    print(f"[{i}/{len(wavs)}] {sess} done in {time.time() - t0:.0f}s", flush=True)

print("all done", flush=True)