解压保存的 pytorch 模型会引发 AttributeError:尽管添加了内联类定义，但无法在 <module '__main__' 上获取属性 'Net' [英] Unpickling saved pytorch model throws AttributeError: Can't get attribute 'Net' on <module '__main__' despite adding class definition inline

问题描述

我正在尝试在 Flask 应用程序中提供 pytorch 模型.当我早些时候在 jupyter 笔记本上运行此代码时，此代码正在运行，但现在我在虚拟环境中运行它，显然即使类定义就在那里，它也无法获得属性Net".所有其他类似的问题都告诉我在同一个脚本中添加保存模型的类定义.但它仍然不起作用.火炬版本是 1.0.1(保存的模型和 virtualenv 都在其中进行了训练)我究竟做错了什么?这是我的代码.

I'm trying to serve a pytorch model in a flask app. This code was working when I ran this on a jupyter notebook earlier but now I'm running this within a virtual env and apparently it can't get attribute 'Net' even though the class definition is right there. All the other similar questions tell me to add the class definition of the saved model in the same script. But it still doesn't work. The torch version is 1.0.1 (where the saved model was trained as well as the virtualenv) What am I doing wrong? Here's my code.

import os
import numpy as np
from flask import Flask, request, jsonify 
import requests

import torch
from torch import nn
from torch.nn import functional as F


MODEL_URL = 'https://storage.googleapis.com/judy-pytorch-model/classifier.pt'


r = requests.get(MODEL_URL)
file = open("model.pth", "wb")
file.write(r.content)
file.close()

class Net(nn.Module):

    def __init__(self):
        super(Net, self).__init__()
        self.fc1 = nn.Linear(input_size, hidden_size)
        self.fc2 = nn.Linear(hidden_size, hidden_size)
        self.fc3 = nn.Linear(hidden_size, output_size)

    def forward(self, x):
        x = torch.sigmoid(self.fc1(x))
        x = torch.sigmoid(self.fc2(x))
        x = self.fc3(x)

        return F.log_softmax(x, dim=-1)

model = torch.load('model.pth')

app = Flask(__name__)

@app.route("/")
def hello():
    return "Binary classification example\n"

@app.route('/predict', methods=['GET'])
def predict():


    x_data = request.args['x_data']

    x_data =  x_data.split()
    x_data = list(map(float, x_data))

    sample = np.array(x_data) 

    sample_tensor = torch.from_numpy(sample).float()

    out = model(sample_tensor)

    _, predicted = torch.max(out.data, -1)

    if predicted.item() == 0: 
         pred_class = "Has no liver damage - ", predicted.item()
    elif predicted.item() == 1:
        pred_class = "Has liver damage - ", predicted.item()

    return jsonify(pred_class)

这是完整的回溯:

Traceback (most recent call last):
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/bin/flask", line 10, in <module>
    sys.exit(main())
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/flask/cli.py", line 894, in main
    cli.main(args=args, prog_name=name)
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/flask/cli.py", line 557, in main
    return super(FlaskGroup, self).main(*args, **kwargs)
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/click/core.py", line 717, in main
    rv = self.invoke(ctx)
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/click/core.py", line 1137, in invoke
    return _process_result(sub_ctx.command.invoke(sub_ctx))
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/click/core.py", line 956, in invoke
    return ctx.invoke(self.callback, **ctx.params)
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/click/core.py", line 555, in invoke
    return callback(*args, **kwargs)
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/click/decorators.py", line 64, in new_func
    return ctx.invoke(f, obj, *args, **kwargs)
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/click/core.py", line 555, in invoke
    return callback(*args, **kwargs)
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/flask/cli.py", line 767, in run_command
    app = DispatchingApp(info.load_app, use_eager_loading=eager_loading)
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/flask/cli.py", line 293, in __init__
    self._load_unlocked()
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/flask/cli.py", line 317, in _load_unlocked
    self._app = rv = self.loader()
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/flask/cli.py", line 372, in load_app
    app = locate_app(self, import_name, name)
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/flask/cli.py", line 235, in locate_app
    __import__(module_name)
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/app.py", line 34, in <module>
    model = torch.load('model.pth')
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/torch/serialization.py", line 368, in load
    return _load(f, map_location, pickle_module)
  File "/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/lib/python3.6/site-packages/torch/serialization.py", line 542, in _load
    result = unpickler.load()
AttributeError: Can't get attribute 'Net' on <module '__main__' from '/Users/judyraj/Judy/pytorch-deployment/flask_app/liver_disease_finder/bin/flask'>

这不能解决我的问题.我不想改变我坚持模型的方式.torch.save() 在虚拟环境之外对我来说很好用.我不介意将类定义添加到脚本中.尽管如此，我还是想看看是什么导致了错误.

This doesn't solve my issue. I do not want to change the way I persist the model. torch.save() worked fine for me outside the virtual env. I don't mind adding the class definition to the script. I'm trying to see what's causing the error despite that.

推荐答案

(这是部分答案)

我不认为 torch.save(model,'model.pt') 在命令提示符下工作，或者当模型从作为 '__main__' 运行的脚本保存时 并从另一个加载.

I don't think torch.save(model,'model.pt') works from the command prompt, or when a model is saved from one script running as '__main__' and loaded from another.

原因是torch必须自动加载用于保存文件的模块，并从__name__中获取模块名称.

The reason is that torch must be automatically loading the module that was used to save the file, and it gets the module name from __name__.

现在是部分部分:目前尚不清楚如何解决此问题，尤其是当您混合使用 virtualenv 时.

Now for the partial part: It's unclear how to fix this issue, especially when you have virtualenvs in the mix.

感谢 Jaentaki开始这个方向的对话.

Thanks to Jatentaki for starting the conversation in this direction.

这篇关于解压保存的 pytorch 模型会引发 AttributeError:尽管添加了内联类定义，但无法在 <module '__main__' 上获取属性 'Net'的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！