01: Intro

Setup

Import fastbook and patch IPython for VS Code

from fastbook import *
from IPython.display import clear_output, DisplayHandle
def update_patch(self, obj):
    clear_output(wait=True)
    self.display(obj)
DisplayHandle.update = update_patch

Set up a path to store the images

data_path = Path('data')/'01'
data_path.mkdir(parents=True,exist_ok=True)

Preview Images

Set up a function to try out different keywords and see how they look

def preview(keyword: string):
    path = data_path / f"{keyword}.jpg"
    download_url(search_images_ddg(keyword, max_images=1)[0], path)
    return Image.open(path).to_thumb(256, 256)

preview("rugby")

100.75% [98304/97571 00:02<00:00]

preview("nfl")

100.88% [245760/243610 00:00<00:00]

preview("soccer")

107.98% [81920/75868 00:00<00:00]

preview("afl")

100.18% [139264/139007 00:00<00:00]

Download Dataset

searches = "rugby","afl","nfl","soccer"
dataset_path = data_path/"datasets"

if not (dataset_path).exists():
    for o in searches:
        dest = (dataset_path/o)
        dest.mkdir(parents=True,exist_ok=True)
        results = search_images_ddg(f"{o} game")
        download_images(dest, urls=results[:200])
        resize_images(dest, max_size=400, dest=dest)

Build Data Block

failed = verify_images(get_image_files(dataset_path))
failed.map(Path.unlink)

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

dls = DataBlock(
    blocks=(ImageBlock, CategoryBlock),
    get_items=get_image_files,
    splitter=RandomSplitter(valid_pct=0.2, seed=42),
    get_y=parent_label,
    item_tfms=[Resize(192, method='squish')]
).dataloaders(dataset_path)

dls.show_batch(max_n=6)

Train Model

learn = vision_learner(dls, resnet18, metrics=error_rate)
learn.fine_tune(30)

epoch	train_loss	valid_loss	error_rate	time
0	1.331276	1.282537	0.398230	00:00
1	1.073042	1.108479	0.362832	00:00
2	0.873794	0.998766	0.371681	00:00
3	0.718180	0.987133	0.389381	00:00
4	0.583215	1.016907	0.362832	00:00
5	0.476426	1.067516	0.345133	00:00
6	0.383498	1.137339	0.362832	00:00
7	0.314319	1.200328	0.353982	00:00
8	0.261772	1.208668	0.336283	00:00
9	0.218976	1.175529	0.327434	00:00
10	0.184930	1.139289	0.300885	00:00
11	0.156622	1.109112	0.292035	00:00
12	0.133540	1.112555	0.292035	00:00
13	0.114173	1.126750	0.300885	00:00
14	0.098229	1.136930	0.300885	00:00
15	0.084903	1.140284	0.318584	00:00
16	0.074624	1.148302	0.327434	00:00
17	0.064603	1.167305	0.318584	00:00
18	0.055911	1.179664	0.318584	00:00
19	0.048461	1.191116	0.327434	00:00
20	0.042165	1.178158	0.336283	00:00
21	0.036902	1.185483	0.327434	00:00
22	0.032348	1.170821	0.336283	00:00
23	0.028613	1.168687	0.336283	00:00
24	0.025399	1.167978	0.345133	00:00
25	0.022398	1.166666	0.345133	00:00
26	0.019721	1.173368	0.345133	00:00
27	0.017580	1.174313	0.345133	00:00
28	0.015551	1.172876	0.345133	00:00
29	0.013663	1.171674	0.345133	00:00

Test Model

learn.show_results()

sport,_,probs = learn.predict(PILImage.create(data_path/'nfl.jpg'))
print(f"this is {sport}")
print("\n--probabilities--")
for i, o in enumerate(dls.vocab):
    print(f"{o}: {probs[i]:.4f}")

this is nfl

--probabilities--
afl: 0.0008
nfl: 0.9891
rugby: 0.0013
soccer: 0.0089