Deploy fastai-v2 model on windows machine

Part 1 (2020)
1

I have a model trained for discriminating different infestations in fruits. I want to deploy this model on a windows-10 machine. Can someone elaborate how to go about it?

1 Like
2

fastai isn’t really compatible with windows. if you cannot get a Linux machine to deploy on, you could try the WSL on windows 10 and deploy there!

3

You can try via dockers here are details => Deployment ready template for creating responsive web app for Fastai2 Vision models

4

Issue is that my earlier model used to run on Windows with Fastai1. Is there no workaround?
Alternatively how can we continue to use Fastai1 in Colab ? further can we install Fastai1 on a windows machine using earlier method.

5

I have exactly the same problem my earlier model run fine on windows with Fastai1 but now I have issue using the latest Fastai-v2. In Fastai1 installation I follow the instruction of https://medium.com/@pierre_guillou/how-to-install-fastai-v1-on-windows-10-ca1bc370dce4.
Can somebody please help?
Thanks

6

Hi,

I can get FastAI working on a conda environment in Win 10 using this environment.yml file

name: fastai
channels:
- fastai
- pytorch
- defaults
dependencies:
- fastcore
- fastai
- fastprogress>=0.1.22
- jupyter
- matplotlib
- nbdev
- pandas
- pillow
- pip
- python>=3.6
- pytorch>=1.3.0
- pyyaml
- requests
- torchvision>=0.5

and running the following command in a Conda shell
conda env create -f environment.yml

7

Just curious, shouldn’t an earlier model be portable in some way?

8

Oh! that should solve the problem.
Can you suggest how to create a yml file in windows? I tried with Notebook++ but always got a .txt file on saving.
Alternately if you can provide a link to the environment.yml file.

Regards

9

Try name in quotes when saving
or
In file Explorer, check the View->File Name Extensions
Then you can rename the file
or
Start Dos Command and
rename a.txt a.yml

10

Hi,
I was successful in installing the fastai using environment.yml. However there seems to be something wrong. When I tried running intro notebook I am getting the error:
No module named ‘fastai2’
Any hint what might be wrong?

11

Hi,
Thanks.
I trained my model using Colab. I exported the model to run it on Windows 10.
The following command run fine: from fastai.vision.all import
But the problem is when I use the following command
load_learner(‘C:/Python/Test/Model/’) (export.pkl is inside the directory).
I got error:
File “C:\Users…\Anaconda3\envs\fastai\lib\site-packages\torch\serialization.py”, line 210, in init
** super(_open_file, self).init(open(name, mode))**
PermissionError: [Errno 13] Permission denied: 'C:/Python/Test/Model/'

Anybody have access to this directory. I also try moving the model to different directory got the same error.

Thanks again

12

instead of fastai2, use fastai

fastai2 was a temporary name during development

13

Yeah, I just got this also when setting up a brand new Windows install - fixed it by downloading the latest fastai source from github again, deleting my existing folder (which I had called “fastai2”, btw) and extracting the github .zip to “fastai-master” and then renaming that new folder to be just “fastai”. Started my kernel and loaded the notebook again, and all fixed now.

14

I trained on colab and exported the model.
For the local machine I installed miniconda 64bits with Python 3.8 and created fastai virtual environment . I followed the installation instructions on the fastai Github page. I am getting
AttributeError: Can’t get attribute ‘FlattenedLoss’ on <module ‘fastai.layers’ from ‘/home/lenovo/miniconda3/envs/fastai2/lib/python3.8/site-packages/fastai/layers.py’>

I even tried updating the fastai installation but the error persists.
Whats wrong?

15

Any progress? I am also trying to deploy my model into a Windows local machine.

16

Could not deploy to Windows machine. Shifted to Linux.

17

Hi - were you able to find a way to get around the “permission denied” message in this use case?

18

I was able to get a tabular model deployed on Windows 10 using Flask. The toughest problem was getting the model loaded. Here are the key things I learned getting this to work:

  1. The path definition was critical. I needed to include the following to get one that worked:

     import pathlib
 temp = pathlib.PosixPath
 pathlib.PosixPath = pathlib.WindowsPath

  2. The only variation of the load_learner statement that worked was with a single argument defined as follows. I had to use os.path.join to get a fully qualified path while avoiding Windows weirdness. The model file adult_sample_model_may16.pkl is in the same directory as the Python module:

         path = Path(os.getcwd())
     full_path = os.path.join(path,'adult_sample_model_may16.pkl')
     learner = load_learner(full_path)

Once I got through those two hurdles, I was able to successfully run inference with the model on my Windows system.