from IPython.display import YouTubeVideo
YouTubeVideo('-Sw87Dyh3Kss')

Setup

!nvidia-smi

Fri Oct 15 16:26:50 2021       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.74       Driver Version: 460.32.03    CUDA Version: 11.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla V100-SXM2...  Off  | 00000000:00:04.0 Off |                    0 |
| N/A   35C    P0    23W / 300W |      0MiB / 16160MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

!pip install -Uqqq fastai

from fastai.vision.all import *

Baixando o dataset

L(dir(URLs))

(#95) ['ADULT_SAMPLE','AG_NEWS','AMAZON_REVIEWS','AMAZON_REVIEWS_POLARITY','BIWI_HEAD_POSE','BIWI_SAMPLE','CALTECH_101','CAMVID','CAMVID_TINY','CARS'...]

URLs.PETS

'https://s3.amazonaws.com/fast-ai-imageclas/oxford-iiit-pet.tgz'

path = untar_data(URLs.PETS)/'images'
path

Path('/root/.fastai/data/oxford-iiit-pet/images')

EDA (Explorando o Dataset)

files = get_image_files(path)

files

(#7390) [Path('/root/.fastai/data/oxford-iiit-pet/images/British_Shorthair_204.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/boxer_165.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/basset_hound_125.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/Bombay_185.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/wheaten_terrier_92.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/yorkshire_terrier_47.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/Siamese_31.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/British_Shorthair_74.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/beagle_157.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/shiba_inu_176.jpg')...]

Image.open(path/'pug_116.jpg')

Image.open(path/'British_Shorthair_201.jpg')

Dataloader

def labeler(x): return 'Gato' if x.name[0].isupper() else 'Cachorro'

labeler(path/'British_Shorthair_201.jpg')

'Gato'

labeler(path/'pug_116.jpg')

'Cachorro'

files = get_image_files(path)
files

(#7390) [Path('/root/.fastai/data/oxford-iiit-pet/images/British_Shorthair_204.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/boxer_165.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/basset_hound_125.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/Bombay_185.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/wheaten_terrier_92.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/yorkshire_terrier_47.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/Siamese_31.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/British_Shorthair_74.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/beagle_157.jpg'),Path('/root/.fastai/data/oxford-iiit-pet/images/shiba_inu_176.jpg')...]

labeler(files[0])

'Gato'

dblock = DataBlock(
    blocks = [ImageBlock, CategoryBlock], # blocos que formata os dados
    splitter = RandomSplitter(valid_pct = 0.2, seed = 42),
    get_items = get_image_files, # função que carrega os dados
    get_y = labeler,
    # get_x = ... 
    item_tfms = Resize(224),
)
dls = dblock.dataloaders(path, num_workers = 2, bs = 64)
dls.show_batch()

xb, yb = dls.one_batch()

Learner

learn = cnn_learner(dls, resnet50, metrics=accuracy)

/usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:718: UserWarning: Named tensors and all their associated APIs are an experimental feature and subject to change. Please do not use them for anything important until they are released as stable. (Triggered internally at  /pytorch/c10/core/TensorImpl.h:1156.)
  return torch.max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)

Transfer Learning

learn.fine_tune(3)

0.999323 ** 12

0.9919061815543482

Verificação

learn.show_results()

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

(996+1+1+480) / 7390

0.2

Testando com uma nova imagem

import ipywidgets as widgets
uploader = widgets.FileUpload()
uploader

img = PILImage.create(uploader.data[0])
label, _, probs = learn.predict(img)
if label == 'Gato':
    print(f'Isso é um {label} com {(probs[1].item()*100):.1f}% de certeza')
else:
    print(f'Isso é um {label} com {(probs[0].item()*100):.1f}% de certeza')
img.to_thumb(300)

Isso é um Gato com 100.0% de certeza

Problemas a serem resolvidos

E se a foto for de um leão?
E se o cachorro for o Pluto?
E se tivermos gatos e cachorros na foto?
E se a foto for de uma cadeira?

Data Augmentation

dblock = DataBlock(
    blocks = [ImageBlock, CategoryBlock],
    get_items = get_image_files,
    get_y = labeler,
    item_tfms = Resize(256),
    batch_tfms = aug_transforms(mult = 2, size = 224)
)
dls = dblock.dataloaders(path, num_workers = 2)

dls.show_batch(unique = True)

epoch	train_loss	valid_loss	accuracy	time
0	0.075581	0.028851	0.991204	00:52
1	0.034629	0.010006	0.995940	00:51
2	0.016202	0.002749	0.999323	00:51