使用flask在heroku bert pytorch模型上部署：错误：_pickle.UnpicklingError：无效的加载键，'v'

Question

[尝试在Heroku上部署bert模型。

import torch
import transformers
import numpy as np
from flask import Flask, render_template, request
from model import DISTILBERTBaseUncased

MAX_LEN = 320
TOKENIZER = transformers.DistilBertTokenizer.from_pretrained(
    "distilbert-base-uncased", do_lower_case=True
)
DEVICE = "cpu"
MODEL = DISTILBERTBaseUncased()
MODEL.load_state_dict(torch.load("weight.bin"))
MODEL.to(DEVICE)
MODEL.eval()

app = Flask(__name__)


def sentence_prediction(sentence):
    tokenizer = TOKENIZER
    max_len = MAX_LEN
    comment = str(sentence)
    comment = " ".join(comment.split())

    inputs = tokenizer.encode_plus(
        comment,
        None,
        add_special_tokens=True,
        max_length=max_len,
        pad_to_max_length=True,
    )

    ids = inputs["input_ids"]
    mask = inputs["attention_mask"]

    ids = torch.tensor(ids, dtype=torch.long).unsqueeze(0)
    mask = torch.tensor(mask, dtype=torch.long).unsqueeze(0)

    ids = ids.to(DEVICE, dtype=torch.long)
    mask = mask.to(DEVICE, dtype=torch.long)

    outputs = MODEL(ids=ids, mask=mask)

    outputs = torch.sigmoid(outputs).cpu().detach().numpy()
    return outputs[0][0]


@app.route("/")
def index_page():
    return render_template("index.html")


@app.route("/model")
def models():
    return render_template("model.html")


@app.route("/predict", methods=["POST", "GET"])
def predict():
    if request.method == "POST":
        sentence = request.form.get("text")
        Toxic_prediction = sentence_prediction(sentence)
        return render_template(
            "index.html", prediction_text=np.round((Toxic_prediction * 100), 2)
        )
    return render_template("index.html", prediction_text="")


if __name__ == "__main__":
    app.run(debug=True)

ERROR

MODEL.load_state_dict（torch.load（“ weight.bin”））

2020-05-18T06：32：32.134536 + 00：00 app [web.1]：文件“ /app/.heroku/python/lib/python3.7/site-packages/torch/serialization.py”，行593，正在加载

2020-05-18T06：32：32.134536 + 00：00 app [web.1]：返回_legacy_load（opened_file，map_location，pickle_module，** pickle_load_args）

2020-05-18T06：32：32.134536 + 00：00 app [web.1]：文件“ /app/.heroku/python/lib/python3.7/site-packages/torch/serialization.py”，行763，在_legacy_load中]

2020-05-18T06：32：32.134537 + 00：00 app [web.1]：magic_number = pickle_module.load（f，** pickle_load_args）

2020-05-18T06：32：32.134537 + 00：00 app [web.1]：_pickle.UnpicklingError：无效的加载密钥'v'。

1. 代码在本地运行良好。
2. Heroku部署方法是Github
3. weight.bin大小为255 MB
4. 烧瓶API在本地主机上正常工作

[尝试在Heroku上部署bert模型。导入焊炬导入变压器从烧瓶导入numpy作为np导入Flask，render_template，从模型导入的请求DISTILBERTBaseUncased MAX_LEN = 320 ...

Answer 1

检查错误1.MODEL.load_state_dict（torch.load（“ weight.bin”））->您应该在下面使用或正确检查字母。