Save both model architecture and weights into one file

frank_chu · July 2, 2019, 10:06pm

Is there a way that we can save model architecture and model weights into one file? And at the inference time, we can simply load that file and do the prediction?

yubozhao · July 3, 2019, 12:42am

Hi Frank,

Are you looking for learner.export() and load_learner(FILE_PATH)?

Under the hood. learner.export is grabbing bunch information into dict and use pytorch’s torch.save to save this dict as a pickle. And when we call load_learner. It will call torch.load and pop each attribute one by one to ‘restore’ the state.

You can check out this file to see what the export function stores in the pickle file, https://github.com/fastai/fastai/blob/5b11f1f864c200623a40565f2efb8c3ba6d334ef/fastai/basic_train.py#L232.
args = ['opt_func', 'loss_func', 'metrics', 'true_wd', 'bn_wd', 'wd', 'train_bn', 'model_dir', 'callback_fns']

Let me know is this what you looking for. I am pretty new with fast.ai as well. Feel free to correct me

Cheers

Bo

frank_chu · July 3, 2019, 11:19pm

Thanks! That is really usefull.

But i am looking for sth that can save model and weights into one single file

muellerzr · July 3, 2019, 11:38pm

It is the learn.export() he described above.

https://course.fast.ai/deployment_render.html

The above link shows a good example of how to deploy, and all you need is learn.export. It saves everything into a pickle file that you can load and unpack.

Eg: learn.export('myModel.pkl)

github.com

render-examples/fastai-v3/blob/master/app/server.py

import aiohttp
import asyncio
import uvicorn
from fastai import *
from fastai.vision import *
from io import BytesIO
from starlette.applications import Starlette
from starlette.middleware.cors import CORSMiddleware
from starlette.responses import HTMLResponse, JSONResponse
from starlette.staticfiles import StaticFiles

export_file_url = 'https://www.dropbox.com/s/6bgq8t6yextloqp/export.pkl?raw=1'
export_file_name = 'export.pkl'

classes = ['black', 'grizzly', 'teddys']
path = Path(__file__).parent

app = Starlette()
app.add_middleware(CORSMiddleware, allow_origins=['*'], allow_headers=['X-Requested-With', 'Content-Type'])
app.mount('/static', StaticFiles(directory='app/static'))

This file has been truncated. show original

If you see lines 32 on on the server.py you will see we can call load_learner where we give it a path, and the file we want to get (that pkl file)

And then if you see line 59-64, we get an image and can run learn.predict()

This is synonymous for any deployment platform you want to do, even if you want to do it in a python notebook. Consider load_learner an equivalent to cnn_learner where we can get all of our information from that pkl file. By information I mean we load an architecture with the weights we saved into the export()

Hope that helps!