Let's make app that recognizes Coquitlam!

Let's get the images of each city

city_types = 'seoul city','coquitlam','paris city', 'new york city'
path = Path('cities')

if not path.exists():
    path.mkdir()
    for o in city_types:
        dest = (path/o)
        dest.mkdir(exist_ok=True)
        results = search_images_bing(key, f'{o}')
        download_images(dest, urls=results.attrgot('content_url'))

fns = get_image_files(path)
fns

failed = verify_images(fns)
failed

failed.map(Path.unlink)

(#16) [None,None,None,None,None,None,None,None,None,None...]

Let's load the data

class DataLoaders(GetAttr):
    def __init__(self, *loaders): self.loaders = loaders
    def __getitem__(self, i): return self.loaders[i]
    train,valid = add_props(lambda i, self: self[i])

cities = 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 = cities.dataloaders(path)
dls.valid.show_batch(max_n=8, nrows = 2)

Let's scale and augment the datas

cities = cities.new(item_tfms=RandomResizedCrop(224, min_scale=0.5),
                        batch_tfms=aug_transforms())
dls = cities.dataloaders(path, num_workers=0) # <- num_workers=0 to prevent window error

Let's build and run a CNN model

learn = cnn_learner(dls, resnet18, metrics=error_rate)
learn.fine_tune(4)

Let's look at the confusion matrix

interp = ClassificationInterpretation.from_learner(learn)
interp.plot_confusion_matrix()

We got an accuracy of 26/90 = 71% (rounded)

interp.plot_top_losses(2,nrows=2)

Let's try to clean up the dataset

cleaner = ImageClassifierCleaner(learn)
cleaner

for idx in cleaner.delete(): cleaner.fns[idx].unlink() # delete

dls = cities.dataloaders(path, num_workers=0)
learn = cnn_learner(dls, resnet18, metrics=error_rate)
learn.fine_tune(5)

Valid_loss doesn't decrease so we stop

interp = ClassificationInterpretation.from_learner(learn)
interp.plot_confusion_matrix()

interp.plot_top_losses(6,nrows=2)

It looks like it has a hard time highlighting seoul city's characteristics as most error comes from seoul images. Suspected factors include seoul having new york like buildings, mountains like coquitlam and brick structures like paris city.

Ignore Below (Deployment IPR)

learn.export()

path = Path()
path.ls(file_exts='.pkl')

learn_inf = load_learner(path/'export.pkl')

learn_inf.dls.vocab

['coquitlam', 'new york city', 'paris city', 'seoul city']

btn_upload = widgets.FileUpload()
btn_upload

img = PILImage.create(btn_upload.data[-1])

out_pl = widgets.Output()
out_pl.clear_output()
with out_pl: display(img.to_thumb(128,128))
out_pl

pred,pred_idx,probs = learn_inf.predict(img)
#hide_output
lbl_pred = widgets.Label()
lbl_pred.value = f'Prediction: {pred}; Probability: {probs[pred_idx]:.04f}'
lbl_pred

Classify Button & Event Handler

btn_run = widgets.Button(description='Classify')
btn_run

def on_click_classify(change):
    img = PILImage.create(btn_upload.data[-1])
    out_pl.clear_output()
    with out_pl: display(img.to_thumb(128,128))
    pred,pred_idx,probs = learn_inf.predict(img)
    lbl_pred.value = f'Prediction: {pred}; Probability: {probs[pred_idx]:.04f}'

btn_run.on_click(on_click_classify)

epoch	train_loss	valid_loss	error_rate	time
0	1.307873	0.871226	0.336207	00:41
1	1.064780	0.831430	0.241379	00:41
2	0.876646	0.767134	0.215517	00:41
3	0.784991	0.738216	0.224138	00:49

epoch	train_loss	valid_loss	error_rate	time
0	1.258363	0.800413	0.347826	00:37
1	1.014135	0.660854	0.243478	00:38
2	0.851025	0.609896	0.243478	00:38
3	0.725140	0.591347	0.217391	00:37
4	0.623130	0.582418	0.226087	00:37