Joalon a software engineer

Villager or not - Single label classification using the fastai framework

Last post we created a dataset for a ‘villager or not a villager’ classification task. This time we’ll train a model and see how well we can get a neural network to recognize said villager. We’ll use a resnet34 model created with the fastai framework, the same as from lesson 1 in the fastai deep learning course 2019.

This code is on Github. I did most of the work on a Google Colab notebook.

Here’s the code for creating a neural net and training it on the dataset:

from fastai.datasets import untar_data, download_data
from import ImageDataBunch, cnn_learner, get_image_files, imagenet_stats, models, error_rate
from google.colab import drive

drive.mount('/content/gdrive', force_remount=False)

path = untar_data('');
training_path = path/'training'
fnames = get_image_files(training_path)
pat = r'/([^/]+)_\d+.png$'

data = ImageDataBunch.from_name_re(training_path, pat=pat, fnames=fnames)

learn = cnn_learner(data, models.resnet34, metrics=error_rate)

learn.fit_one_cycle(10)'/content/gdrive/My Drive/Villager or not/first-villager-model')

Here you can see it running: Training

Note! I did have to add a GPU to the colab notebook, otherwise the training time for each epoch was 10 minutes instead of 10 seconds.


Choose GPU

Let’s see how it did, we can plot a confusion matrix:

interp = ClassificationInterpretation.from_learner(learn)

Confusion matrix

That’s cool, it’s gotten 5 images wrong in the whole set. In the last post we concluded that the image generation was a bit flawed, now is a good time to check which images the model was most confused about.

interp.plot_top_losses(9, figsize=(15,11))

Top losses

Not very strange that these images have a high loss, they’re clearly wrongly labelled. It’s the perfect time to use a cool widget from the fastai library called the ImageCleaner to do some relabeling:

ds, idxs = DatasetFormatter().from_toplosses(learn)
ImageCleaner(ds, idxs, path)

Image cleaner

I had to spend quite some time relabeling images. The generator wasn’t as good as I expected, this model has probably overfit to be able to do so good on the wrongly labeled images.

A water obstacle

Note: I couldn’t run the ImageCleaner on Google Colab since it doesn’t allow running widgets so, to run it, I had to use my laptop. Unfortunately I don’t have a current GPU so I had to install pytorch as CPU only:

pip3 install
pip3 install
pip3 install fastai

Pytorchs website has a really good tool for getting installation instructions for your own environment: CPU only pytorch

The ImageCleaner widget doesn’t actually change your dataset on disk, it creates a cleaned.csv file in the training path which can be loaded and retrained with:

df = pd.read_csv(path/'cleaned.csv', header='infer')

db = (ImageList.from_df(df, path)

learn = cnn_learner(db, models.resnet34, metrics=error_rate)
learn = learn.load('villager-or-not')

We’ll need the export.pkl file from executing learn.export() to create a production model and run predictions on.

To finally test our model let’s grab a couple of screenshots the model hasn’t seen before to run inference with. I took some from the AoE 2 Wiki and cropped them. The champion is a unit the model hasn’t seen before, it’ll be interesting to see how it handles that.

Test image Test image2 Test image3

The directory you’re running the following code in should contain the aforementioned export.pkl.

from import load_learner, open_image
from PIL import Image

learn = load_learner('./')
image_names = ['villager-test.png', 'gold-test.png', 'champion-test.png']

for i in range(3):
    image = open_image(image_names[i])
    prediction = learn.predict(image)
    print(image_names[i] + ': ' + prediction[0])

Prediction result

It works! Although it classified the champion as a villager it has never seen another unit so that’s an honest mistake. Next lesson is multi label classification, stay tuned.

comments powered by Disqus