Cricket vs Baseball
Deep Neural network attempts to distinguish between whether someone is playing baseball or cricket
This is a deep learning trained model that attempts to distinguish between cricket and baseball players
The result will be either of: Batsman (cricket) Bowler (cricket) Batter (baseball) Pitcher (baseball).
The model was trained using image search data from Bing. However, the data was quite bad (with a lot of cartoons and unrelated images) and hence the classifier is not very good. For example, the stumps in a cricketer's image is primarily attributed to a batsman and if stumps appears in a bowler's image, it is likely to be classified as batsman.
This is the reason why I moved to handling another problem, Feling Classifier. That blog entry details the step-by-step discussion on how the neural network classifier was developed using Fastai, PyTorch and Bing Image search
In this blog entry, a very brief description of the steps that went into building this model is given, espcially to both show how we collected the images as well as the reason why this wasn't much successful by showing examples of the training data. The steps are cursory and some outputs (such as training) are not shown here (I was trying to reduce the size of the jupyter notebook and instead ended up removing all outputs).
Even then, the app was built and is hosted on binder here:
key = 'xxpas324sc4r3r32' # This is not the actual key, but something random to hide my key
To ensure that the images are as relevant as possible, I used a search term with the name of the sport prefixed to it. This means, I had to use a different image folder name.
image_types = 'cricket batsman','cricket bowler','baseball batter', 'baseball pitcher'
image_folder_names = 'batsman', 'bowler', 'batter', 'pitcher'
path = Path('cricball')
if not path.exists():
path.mkdir()
for type,foldername in zip(image_types,image_folder_names):
dest = (path/foldername)
dest.mkdir(exist_ok=True)
results = search_images_bing(key, f'{type}')
download_images(dest, urls=results.attrgot('content_url'))
And remove any corrupted images
fns = get_image_files(path)
failed = verify_images(fns)
failed.map(Path.unlink);
sportsmen = DataBlock(
blocks=(ImageBlock, CategoryBlock),
get_items=get_image_files,
splitter=RandomSplitter(valid_pct=0.2, seed=42),
get_y=parent_label,
item_tfms=Resize(128))
dls = sportsmen.dataloaders(path)
sportsmen = sportsmen.new(
item_tfms=RandomResizedCrop(224, min_scale=0.5),
batch_tfms=aug_transforms())
dls = sportsmen.dataloaders(path)
learn = cnn_learner(dls, resnet18, metrics=error_rate)
learn.fine_tune(4)
interp = ClassificationInterpretation.from_learner(learn)
interp.plot_confusion_matrix()
interp.plot_top_losses(18, nrows=3)
Honestly, this is disappointingly bad set of data. So many cartoons and sketches, and even some completely irrelevant images. So we clean these up a bit.
cleaner = ImageClassifierCleaner(learn)
cleaner
Unfortunately, the widget is not showing up above, but well, you can refer to how it looks in the Feline Classifier blog entry.
for idx in cleaner.delete(): cleaner.fns[idx].unlink()
for idx,cat in cleaner.change(): shutil.move(str(cleaner.fns[idx]), path/cat)
sportsmen = sportsmen.new(
item_tfms=RandomResizedCrop(224, min_scale=0.5),
batch_tfms=aug_transforms())
dls = sportsmen.dataloaders(path)
learn = cnn_learner(dls, resnet18, metrics=error_rate)
learn.fine_tune(4)
interp = ClassificationInterpretation.from_learner(learn)
interp.plot_confusion_matrix()
interp.plot_top_losses(18, nrows=3)
Slightly better, but the mismatch in the prediction says that NN is struggling in its task. But the confusion matrix seems okay here. And we say we will stop this whole thing here.
So we just export the model and live with it
learn.export()
The app is developed in the same way as Feline classifier and is available in this github repository: Cricket vs Baseball
I still think it works decently and the app is hosted on Binder and you can access it here:
I will conquer this problem better another time