| import json, soundfile as sf, torch |
| from transformers import AutoModelForCTC, AutoProcessor |
| M="/root/models/joint_cont_best" |
| proc=AutoProcessor.from_pretrained(M) |
| model=AutoModelForCTC.from_pretrained(M,torch_dtype=torch.bfloat16).cuda().eval() |
| rows=[json.loads(l) for l in open("/root/devhard/devhard_linsna.jsonl")] |
| def norm(s): return " ".join(str(s).replace("|"," ").split()) |
| out=[] |
| with torch.inference_mode(): |
| for i in range(0,len(rows),8): |
| b=rows[i:i+8] |
| au=[sf.read(r["audio"],dtype="float32")[0] for r in b] |
| x=proc(au,sampling_rate=16000,return_tensors="pt",padding=True) |
| x={k:v.to("cuda",dtype=torch.bfloat16 if v.dtype==torch.float32 else v.dtype) for k,v in x.items()} |
| pid=model(**x).logits.float().argmax(-1).cpu().numpy() |
| for r,h in zip(b,proc.batch_decode(pid)): |
| out.append({"id":r["id"],"lang":r.get("lang"),"ref":r["text"],"hyp":norm(h)}) |
| json.dump(out,open("/root/devhard_joint_hyps.json","w",encoding="utf-8"),ensure_ascii=False) |
| print("DECODE_DEVHARD_DONE",len(out)) |
|
|
