np.save 带来的误差影响

[dbcSep03]

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
from FlagEmbedding import BGEM3FlagModel
import pandas as pd
import numpy as np
if __name__ == "__main__":
    model = BGEM3FlagModel('BAAI/bge-m3',  use_fp16=True) 
    df = pd.read_parquet("00000.parquet")
    npy_data = np.load("00000.npy")
    texts_all = [temp["text"] for _, temp in df.iterrows()]
    for index, text in enumerate(texts_all):
        print(text[:100])
        embedding1 = model.encode([text], max_length=2048)['dense_vecs'][0]
        embedding2 = model.encode([text], max_length=2048)['dense_vecs'][0]
        test_embed = npy_data[index]
        print(test_embed.dtype) # float16
        print(embedding1.dtype) # float16
        print(embedding2.dtype) # float16
        print(embedding1.shape) # (1024,)
        print(embedding2.shape) # (1024,)
        print(test_embed.shape) # (1024,)
        print(embedding1 == test_embed) # [False False False ... False False False]
        print(embedding2 == test_embed) # [False False False ... False False False]
        print(embedding2 == embedding1) # [ True  True  True ...  True  True  True]
        print(np.allclose(test_embed, embedding2, atol=1e-3))  # 允许 1e-3 误差  # False
        similarity = embedding1 @ test_embed.T
        print(similarity) # 0.999
        break

测试了一下 结果如下所示，想问一下 np.save带来的误差影响有多少？ 为什么会有那么大量的偏差？

[hanhainebula]

你好，@dbcSep03。这个误差属于正常现象，在下游检索场景中，这种误差基本不会对检索结果产生较大影响。