Doppelganger

%load_ext autoreload
%autoreload 2

Imports

from fastai.vision.all import *
from fastai.vision.widgets import *
from aiking.data.external import *

path = get_ds("DoppelGanger"); path.ls()

(#8) [Path('/Landmark2/pdo/aiking/data/DoppelGanger/Javier_Bardem'),Path('/Landmark2/pdo/aiking/data/DoppelGanger/Mercedes_Masohn'),Path('/Landmark2/pdo/aiking/data/DoppelGanger/Sarah_Shahi'),Path('/Landmark2/pdo/aiking/data/DoppelGanger/Jeffrey_Dean_Morgan'),Path('/Landmark2/pdo/aiking/data/DoppelGanger/Brad_Pitt'),Path('/Landmark2/pdo/aiking/data/DoppelGanger/Benicio_Del_Toro'),Path('/Landmark2/pdo/aiking/data/DoppelGanger/Cuba_Gooding_Jr'),Path('/Landmark2/pdo/aiking/data/DoppelGanger/Donald_Faison')]

!exa --grid {path}

Benicio_Del_Toro  Cuba_Gooding_Jr  Javier_Bardem        Mercedes_Masohn
Brad_Pitt         Donald_Faison    Jeffrey_Dean_Morgan  Sarah_Shahi

Define Datablock & Dataloaders

doppelganger = DataBlock(
    blocks = (ImageBlock, CategoryBlock),
    get_items=get_image_files,
    get_y=parent_label,
    splitter=RandomSplitter(valid_pct=0.2, seed = 42),
    item_tfms=RandomResizedCrop(224,min_scale=0.3),
    batch_tfms = aug_transforms()
)

doppelganger

dls = doppelganger.dataloaders(path); dls

<fastai.data.core.DataLoaders at 0x7f5559f8d0d0>

dls.train.show_batch(max_n=8,nrows=2,unique=True)

dls.train.show_batch(max_n=8,nrows=2)

Training and Fine Tuning

learn = cnn_learner(dls, resnet18, metrics= [ accuracy, error_rate]); learn

learn.fine_tune(8)

epoch	train_loss	valid_loss	accuracy	error_rate	time
0	2.733192	1.106454	0.567308	0.432692	00:13

epoch	train_loss	valid_loss	accuracy	error_rate	time
0	1.466980	0.813958	0.682692	0.317308	00:14
1	1.280565	0.585941	0.783654	0.216346	00:13
2	1.090317	0.491283	0.826923	0.173077	00:12
3	0.893104	0.491068	0.850962	0.149038	00:12
4	0.746640	0.476913	0.836538	0.163462	00:12
5	0.645121	0.464664	0.846154	0.153846	00:13
6	0.558647	0.457159	0.850962	0.149038	00:12
7	0.504691	0.463198	0.850962	0.149038	00:12

Classification Interpretation

interp = ClassificationInterpretation.from_learner(learn); interp

interp.plot_confusion_matrix()

interp.plot_top_losses(k=30, nrows=30)

ML Driven Data Cleaning

cleaner = ImageClassifierCleaner(learn)
cleaner

Note

Couldnot figure out any good cases to be cleaned. Although being in suit seems to play an important role for classifer particularily for Cuba Gooding Jr.

?ImageClassifierCleaner

Init signature: ImageClassifierCleaner(learn, opts=(), height=128, width=256, max_n=30)
Docstring:      A widget that provides an `ImagesCleaner` with a CNN `Learner`
File:           ~/anaconda3/envs/aiking/lib/python3.8/site-packages/fastai/vision/widgets.py
Type:           type
Subclasses:     

Model Export

c = Config()
# c.model_path.mkdir()
# (c.model_path/'doppelganger.pkl').unlink()
learn.export(c.learner_path/'doppelganger.pkl')

Inference

learn_inf = load_learner(c.learner_path/'doppelganger.pkl')

uploader = widgets.FileUpload(); uploader

img = PILImage.create(uploader.data[0])
img

learn_inf.predict(img)

('Mercedes_Masohn',
 tensor(6),
 tensor([2.4761e-05, 2.0251e-06, 4.3338e-04, 3.2580e-03, 6.0018e-05, 7.6089e-07,
         9.8427e-01, 1.1951e-02]))

uploader = widgets.FileUpload(); uploader

img = PILImage.create(uploader.data[0])
img

learn_inf.predict(img)

('Javier_Bardem',
 tensor(4),
 tensor([4.2255e-06, 3.2207e-06, 5.6138e-08, 2.0617e-05, 9.9911e-01, 8.4978e-04,
         3.7117e-07, 1.3715e-05]))

Note

Model Got Sarah Shahi wrong and Javier Bardem Correct. What to do to improve things further?