Datasets:
File size: 2,605 Bytes
9332074 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 | import sys,subprocess,json,time,hashlib
from pathlib import Path
p=subprocess.run([sys.executable,"-m","pip","install","--quiet","onnxruntime","tokenizers"],capture_output=True,text=True,timeout=180)
assert p.returncode==0,"CPU embedding dependencies unavailable"
import numpy as np,onnxruntime as ort
from tokenizers import Tokenizer
from huggingface_hub import HfApi,hf_hub_download
D=Path("dataset-repair");pairs=json.loads((D/"translation_candidates.json").read_text())
repo="sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
revision="e8f8c211226b894fcb81acc59f3b34ba3efd5f42"
model_path=hf_hub_download(repo,"onnx/model_quint8_avx2.onnx",revision=revision,token=False)
tokenizer_path=hf_hub_download(repo,"tokenizer.json",revision=revision,token=False)
tok=Tokenizer.from_file(tokenizer_path);tok.enable_truncation(max_length=128)
tok.enable_padding(pad_id=tok.token_to_id("<pad>"),pad_token="<pad>")
opt=ort.SessionOptions();opt.intra_op_num_threads=2;opt.inter_op_num_threads=1
session=ort.InferenceSession(model_path,sess_options=opt,providers=["CPUExecutionProvider"])
input_names={v.name for v in session.get_inputs()}
scores=[];started=time.monotonic()
for offset in range(0,len(pairs),32):
batch=pairs[offset:offset+32]
texts=[s for row in batch for s in [row["source_text"],row["target_text"]]]
enc=tok.encode_batch(texts)
inputs={"input_ids":np.asarray([x.ids for x in enc],dtype=np.int64),"attention_mask":np.asarray([x.attention_mask for x in enc],dtype=np.int64),"token_type_ids":np.asarray([x.type_ids for x in enc],dtype=np.int64)}
output=session.run(None,{k:v for k,v in inputs.items() if k in input_names})[0]
if output.ndim==3:
mask=inputs["attention_mask"][...,None]
vectors=(output*mask).sum(1)/mask.sum(1).clip(1)
else:vectors=output
vectors=vectors/np.linalg.norm(vectors,axis=1,keepdims=True).clip(1e-9)
scores.extend(float(np.dot(vectors[i],vectors[i+1])) for i in range(0,len(vectors),2))
if offset%256==0:
progress={"done":len(scores),"total":len(pairs),"seconds":round(time.monotonic()-started,1)}
(D/"alignment_progress.json").write_text(json.dumps(progress));print("CPU_ALIGNMENT",json.dumps(progress),flush=True)
assert len(scores)==len(pairs) and all(np.isfinite(scores))
(D/"alignment_scores.json").write_text(json.dumps(scores))
meta={"model":repo,"revision":revision,"runtime":"ONNX Runtime CPU, portable AVX2 UINT8","file":"onnx/model_quint8_avx2.onnx","max_tokens":128,"records":len(pairs),"elapsed_seconds":time.monotonic()-started}
(D/"alignment_metadata.json").write_text(json.dumps(meta,indent=2))
print(json.dumps(meta,indent=2))
|