Set up Windows OS for Fastai v2

Preparation - Set up GPU drivers

Follow the instructions in Configure GPU on windows page first.

Install Fastai v2.

What did not work: - The Fastbook install instructions don’t work as the automatic Pip install steps fail. Manually running Pip install also results in conflicts. - Running conda install -c fastchan fastai will not resolve as conda satsolver takes forever without results.

What worked: - Install Anaconda individual edition - Create a new env: conda create -n fastaiv2 without any packages. This provides conda a fresh start and makes it easy for the solver - Then run conda install -c fastai -c pytorch fastai to install all fastaiv2 and all of its dependencies. - If you want to run the notebooks, then run conda install jupyter to install Jupyter - Optional: Clone the v2 book repo: https://github.com/fastai/fastbook - Run pip install -U fastbook to install the book’s deps and files on disk.

Verify GPU is picked up

From terminal or Notebook, run the following

import torch
torch.cuda.is_available()

If you get a True, you are good to go. Else revisit the config doc page listed above.

Verify FastAI v2 can be imported

Run the quickstart example from the book (or copy below):

from fastai.vision.all import *
path = untar_data(URLs.PETS)/'images'

def is_cat(x): return x[0].isupper()
dls = ImageDataLoaders.from_name_func(
    path, get_image_files(path), valid_pct=0.2, seed=42,
    label_func=is_cat, item_tfms=Resize(224))

learn = cnn_learner(dls, resnet34, metrics=error_rate)
learn.fine_tune(1)

As you run the above, open the Task Manager and monitor the GPU usage. You should see a spike while the CPU rate seems constant. The training should go much faster. Instead if you notice CPU peaking to 95-100%, then verify if torch.cuda.is_available() returns True. If it does not, then recheck the driver and OS version steps from above.

Congrats! You are all set to use fastaiv2 on Windows OS without needing a dual boot for Linux or installing the Windows subsystem for linux options.