1- What is that?

Trying out the concepts taught in fastai lesson 1 with a new dataset.

Inspired by this Superman comic, I’m building a deep learning classifier to distinhuish between images of birds, planes and supermen.

!pwd
/content

Workflow

Here’s the process:

  • Choose categories- bird, plane, superman
  • Download 1 image for each and check they’re ok
  • Download 30 images for each into directories named by category
  • Check images are valid
  • Download pre-trained model
  • Feed dir of images into model, fine tune, check performance
!pip install -Uqq fastai duckduckgo_search
     |█████▏                          | 10 kB 24.3 MB/s eta 0:00:01     |██████████▍                     | 20 kB 15.9 MB/s eta 0:00:01     |███████████████▋                | 30 kB 20.8 MB/s eta 0:00:01     |████████████████████▉           | 40 kB 13.2 MB/s eta 0:00:01     |██████████████████████████      | 51 kB 7.1 MB/s eta 0:00:01     |███████████████████████████████▎| 61 kB 8.4 MB/s eta 0:00:01     |████████████████████████████████| 62 kB 1.5 MB/s 
     |███▍                            | 10 kB 25.6 MB/s eta 0:00:01     |██████▉                         | 20 kB 33.5 MB/s eta 0:00:01     |██████████▏                     | 30 kB 38.0 MB/s eta 0:00:01     |█████████████▋                  | 40 kB 43.0 MB/s eta 0:00:01     |█████████████████               | 51 kB 44.0 MB/s eta 0:00:01     |████████████████████▍           | 61 kB 47.6 MB/s eta 0:00:01     |███████████████████████▊        | 71 kB 50.3 MB/s eta 0:00:01     |███████████████████████████▏    | 81 kB 51.8 MB/s eta 0:00:01     |██████████████████████████████▌ | 92 kB 53.5 MB/s eta 0:00:01     |████████████████████████████████| 96 kB 6.5 MB/s 
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
flask 1.1.4 requires click<8.0,>=5.1, but you have click 8.1.3 which is incompatible.
from duckduckgo_search import ddg_images
from fastcore.all import *

Data

Searching for images

bird = ddg_images('bird', max_results=2)
  • Bird is a list of dicts
  • wrap inside L
  • use .itemgot('image') on L to extract the urls only
bird
[{'title': 'Pair these bird events with Great Backyard Bird Count - pennlive.com',
  'image': 'https://www.pennlive.com/resizer/UeXGnWaOE07lC4n_J5lJ0a1M-_M=/1280x0/smart/advancelocal-adapter-image-uploads.s3.amazonaws.com/image.pennlive.com/home/penn-media/width2048/img/entertainment_impact/photo/shutterstock-58364638jpg-6932548563f01558.jpg',
  'thumbnail': 'https://tse1.mm.bing.net/th?id=OIP.tAh4IvhbxtzpJUCMwC-djQHaGe&pid=Api',
  'url': 'https://www.pennlive.com/entertainment/2015/02/pair_these_bird_events_with_gr.html',
  'height': 1118,
  'width': 1280,
  'source': 'Bing'},
 {'title': "Help Count Birds for Science during Audubon's Annual Christmas Bird Count - Restore the ...",
  'image': 'https://www.mississippiriverdelta.org/files/2015/12/BUFH-wing-feathers.jpg',
  'thumbnail': 'https://tse2.mm.bing.net/th?id=OIP.1xYUOvqR7iH9FMoKBawtIAHaF1&pid=Api',
  'url': 'https://mississippiriverdelta.org/help-count-birds-for-science-during-audubons-annual-christmas-bird-count/',
  'height': 1340,
  'width': 1700,
  'source': 'Bing'}]
L(bird).itemgot('image')
(#2) ['https://www.pennlive.com/resizer/UeXGnWaOE07lC4n_J5lJ0a1M-_M=/1280x0/smart/advancelocal-adapter-image-uploads.s3.amazonaws.com/image.pennlive.com/home/penn-media/width2048/img/entertainment_impact/photo/shutterstock-58364638jpg-6932548563f01558.jpg','https://www.mississippiriverdelta.org/files/2015/12/BUFH-wing-feathers.jpg']

Build search_images using these

def search_images(keyword, max_images=30):
    return L(ddg_images(keyword, max_results=max_images)).itemgot('image')
url = search_images('flying bird', 1)
url
(#1) ['http://1.bp.blogspot.com/-6anwlEoD-Ec/UmvgxLKbxsI/AAAAAAAAZhQ/z64F1RGv7iU/s1600/Flying+Birds+Wallpapers+%25283%2529.jpg']

Downloading and viewing single images

from fastdownload import download_url
download_url(url[0], dest='bird.jpg', show_progress=False)
Path('bird.jpg')
from fastai.vision.all import *
im = Image.open('bird.jpg')
im.to_thumb(256,256)

download_url(search_images('flying plane', 1)[0], dest='plane.jpg', show_progress=False)
Image.open('plane.jpg').to_thumb(256, 256)

download_url(search_images('flying superman', 1)[0], dest='superman.jpg', show_progress=False)
Image.open('superman.jpg').to_thumb(256, 256)

Download many images

import os
searches = ['bird', 'plane', 'superman']
path = Path('is_that_a')

for s in searches:
    dest = (path/s)
    dest.mkdir(exist_ok=True, parents=True)
    download_images(dest, urls=search_images(f'flying {s}'))
    resize_images(path/s, max_size=400, dest=path/s)
failed = verify_images(get_image_files(path))
failed.map(Path.unlink)
len(failed)
0

Fine-tune model

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(path, bs=32)
dls.show_batch(max_n=12)

learn = vision_learner(dls, resnet18, metrics=error_rate)
learn.fine_tune(3)
/usr/local/lib/python3.7/dist-packages/torchvision/models/_utils.py:209: UserWarning: The parameter 'pretrained' is deprecated since 0.13 and will be removed in 0.15, please use 'weights' instead.
  f"The parameter '{pretrained_param}' is deprecated since 0.13 and will be removed in 0.15, "
/usr/local/lib/python3.7/dist-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and will be removed in 0.15. The current behavior is equivalent to passing `weights=ResNet18_Weights.IMAGENET1K_V1`. You can also use `weights=ResNet18_Weights.DEFAULT` to get the most up-to-date weights.
  warnings.warn(msg)
Downloading: "https://download.pytorch.org/models/resnet18-f37072fd.pth" to /root/.cache/torch/hub/checkpoints/resnet18-f37072fd.pth
epoch train_loss valid_loss error_rate time
0 2.054615 1.164085 0.352941 00:08
epoch train_loss valid_loss error_rate time
0 1.213407 0.438084 0.176471 00:12
1 0.745004 0.064237 0.000000 00:13
2 0.510461 0.036753 0.000000 00:13

Evaluate

is_that_a,_,probs = learn.predict(PILImage.create('bird.jpg'))
print(f"That is a: {is_that_a}.")
print(f"Probability: {probs[0]:.4f}")

im = Image.open('bird.jpg')
im.to_thumb(256,256)
That is a: bird.
Probability: 0.9944

is_that_a,_,probs = learn.predict(PILImage.create('plane.jpg'))
print(f"That is a: {is_that_a}.")
print(f"Probability: {probs[1]:.4f}")

im = Image.open('plane.jpg')
im.to_thumb(256,256)
That is a: plane.
Probability: 0.9989

is_that_a,_,probs = learn.predict(PILImage.create('superman.jpg'))
print(f"That is a: {is_that_a}.")
print(f"Probability: {probs[2]:.4f}")

im = Image.open('superman.jpg')
im.to_thumb(256,256)
That is a: superman.
Probability: 0.9998