Multi-label Classification of Amazonian Land Space

Python

Deep Learning

Computer Vision

Project

Author

Jake Gehri

Published

October 19, 2022

import pandoc
from fastbook import *
from fastai import *

creds = '{"username":"jakegehri","key":"3d0213a52bd1816d21037e941bc77569"}'

# Creates a cred path for kaggle datasets to be downloaded in comand line

cred_path = Path('~/.kaggle/kaggle.json').expanduser() 

if not cred_path.exists():
    cred_path.parent.mkdir(exist_ok=True)
    cred_path.write_text(creds)
    cred_path.chmod(0o600)

path = Path('./planet/planet')

path.ls()

(#4) [Path('planet/planet/train_classes.csv'),Path('planet/planet/test-jpg'),Path('planet/planet/sample_submission.csv'),Path('planet/planet/train-jpg')]

train = (path/'train-jpg').ls()

train

(#40479) [Path('planet/planet/train-jpg/train_19921.jpg'),Path('planet/planet/train-jpg/train_24619.jpg'),Path('planet/planet/train-jpg/train_21510.jpg'),Path('planet/planet/train-jpg/train_31089.jpg'),Path('planet/planet/train-jpg/train_33277.jpg'),Path('planet/planet/train-jpg/train_11172.jpg'),Path('planet/planet/train-jpg/train_14671.jpg'),Path('planet/planet/train-jpg/train_29521.jpg'),Path('planet/planet/train-jpg/train_27535.jpg'),Path('planet/planet/train-jpg/train_13323.jpg')...]

data = (path/'train_classes.csv')

df = pd.read_csv(data)

df.head()

	image_name	tags
0	train_0	haze primary
1	train_1	agriculture clear primary water
2	train_2	clear primary
3	train_3	clear primary
4	train_4	agriculture clear habitation primary road

dblock = DataBlock()

dsets = dblock.datasets(df)

dsets[0]

(image_name         train_0
 tags          haze primary
 Name: 0, dtype: object,
 image_name         train_0
 tags          haze primary
 Name: 0, dtype: object)

def get_x(r): return path/'train-jpg'/(r['image_name'] + '.jpg')
def get_y(r): return r['tags'].split(' ')

dblock = DataBlock(get_x=get_x, get_y=get_y)
dsets = dblock.datasets(df)
dsets.train[0]

(Path('planet/planet/train-jpg/train_1398.jpg'),
 ['agriculture', 'partly_cloudy', 'primary', 'road'])

dblock = DataBlock(blocks = (ImageBlock, MultiCategoryBlock), 
                   get_x=get_x, get_y=get_y)

dsets = dblock.datasets(df)
dsets.train[0]

(PILImage mode=RGB size=256x256,
 TensorMultiCategory([1., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 1., 1., 0., 0., 0., 0.]))

dblock = DataBlock(blocks = (ImageBlock, MultiCategoryBlock), 
                   splitter=RandomSplitter(seed=42),
                   get_x=get_x, get_y=get_y)

dsets = dblock.datasets(df)

len(dsets.train), len(dsets.valid)

(32384, 8095)

dblock = DataBlock(blocks = (ImageBlock, MultiCategoryBlock), 
                   splitter=RandomSplitter(seed=42),
                   get_x=get_x, get_y=get_y,
                   item_tfms=RandomResizedCrop(128, min_scale=0.35)
                  
)

dls = dblock.dataloaders(df)

dls.show_batch()

Experiment 1

learn = vision_learner(dls, resnet18, metrics=accuracy_multi)
learn.fine_tune(6, freeze_epochs=2)
learn.recorder.plot_loss()

/usr/local/lib/python3.9/dist-packages/torchvision/models/_utils.py:208: UserWarning: The parameter 'pretrained' is deprecated since 0.13 and will be removed in 0.15, please use 'weights' instead.
  warnings.warn(
/usr/local/lib/python3.9/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	accuracy_multi	time
0	0.242502	0.150174	0.945827	00:25
1	0.147816	0.126972	0.951968	00:20

epoch	train_loss	valid_loss	accuracy_multi	time
0	0.127971	0.109930	0.957991	00:23
1	0.119986	0.104264	0.960476	00:23
2	0.109712	0.098336	0.962301	00:38
3	0.104570	0.097734	0.962555	00:49

Experiment 2

learn2 = vision_learner(dls, resnet34, metrics=accuracy_multi)
learn2.fine_tune(6, freeze_epochs=2)
learn2.recorder.plot_loss()

epoch	train_loss	valid_loss	accuracy_multi	time
0	0.244604	0.147677	0.947447	00:51
1	0.148016	0.123402	0.953312	00:25

epoch	train_loss	valid_loss	accuracy_multi	time
0	0.123248	0.106454	0.959452	00:59
1	0.118266	0.104211	0.960942	00:34
2	0.111732	0.098766	0.962381	00:42
3	0.105815	0.096162	0.963245	00:34
4	0.099851	0.094583	0.964037	00:34
5	0.093574	0.095035	0.963848	00:34

Experiment 3

learn3 = vision_learner(dls, resnet34, metrics=partial(accuracy_multi, thresh=0.2))
learn3.fine_tune(6, freeze_epochs=2)
learn3.recorder.plot_loss()

epoch	train_loss	valid_loss	accuracy_multi	time
0	0.240184	0.151341	0.934862	00:38
1	0.151022	0.122403	0.939382	00:27

epoch	train_loss	valid_loss	accuracy_multi	time
0	0.124375	0.106883	0.950107	00:35
1	0.118021	0.101975	0.951757	00:34
2	0.109789	0.099863	0.952055	00:34
3	0.104578	0.096811	0.951851	00:34
4	0.095859	0.095459	0.953784	00:34
5	0.092178	0.095330	0.954445	01:02

Experiment 4

learn4 = vision_learner(dls, resnet34, metrics=partial(accuracy_multi, thresh=0.7))
learn4.fine_tune(6, freeze_epochs=2)
learn4.recorder.plot_loss()

epoch	train_loss	valid_loss	accuracy_multi	time
0	0.248779	0.148406	0.942892	00:52
1	0.147272	0.123920	0.950325	00:37

epoch	train_loss	valid_loss	accuracy_multi	time
0	0.125626	0.109230	0.955863	00:48
1	0.118446	0.103957	0.956015	00:34
2	0.112594	0.099723	0.958754	00:37
3	0.106374	0.096328	0.958689	00:51
4	0.099089	0.094283	0.960499	00:34
5	0.095276	0.094695	0.961879	00:39

test_dl = learn2.dls.test_dl(get_image_files(path/'test-jpg'))
test_dl.show_batch()

preds, _ = learn2.get_preds(dl=test_dl)

thresh = 0.5
labelled_preds = [' '.join([learn2.dls.vocab[i] for i,p in enumerate(pred) if p > thresh]) for pred in preds]

labelled_preds[:5]

['agriculture clear primary road',
 'bare_ground clear',
 'clear primary',
 'clear primary',
 'clear primary']

fnames = []

for name in os.listdir((path/'test-jpg')):
    fnames.append(name)

fnames[0:5]

['test_36099.jpg',
 'test_27503.jpg',
 'test_15453.jpg',
 'test_20695.jpg',
 'test_5439.jpg']

learn2.predict((path/'test-jpg'/(fnames[0])))

((#4) ['agriculture','clear','primary','road'],
 TensorBase([ True, False, False, False, False,  True, False, False, False, False, False, False,  True,  True, False, False, False]),
 TensorBase([9.0538e-01, 3.7246e-04, 4.8710e-02, 5.2741e-04, 2.3139e-04, 9.9834e-01, 1.1549e-05, 2.1592e-04, 2.8083e-01, 2.6772e-01, 1.3102e-03, 5.9414e-04, 9.9820e-01, 7.1026e-01, 3.3888e-03,
             1.9953e-02, 1.1821e-01]))

sample = pd.read_csv(path/'sample_submission.csv')
sample

	image_name	tags
0	test_0	primary clear agriculture road water
1	test_1	primary clear agriculture road water
2	test_2	primary clear agriculture road water
3	test_3	primary clear agriculture road water
4	test_4	primary clear agriculture road water
...	...	...
61186	file_9995	primary clear agriculture road water
61187	file_9996	primary clear agriculture road water
61188	file_9997	primary clear agriculture road water
61189	file_9998	primary clear agriculture road water
61190	file_9999	primary clear agriculture road water

61191 rows × 2 columns

df = pd.DataFrame({'image_name':fnames, 'tags':labelled_preds}, columns=['image_name', 'tags'])

df

	image_name	tags
0	test_36099.jpg	agriculture clear primary road
1	test_27503.jpg	bare_ground clear
2	test_15453.jpg	clear primary
3	test_20695.jpg	clear primary
4	test_5439.jpg	clear primary
...	...	...
40664	test_12322.jpg	clear primary water
40665	test_10596.jpg	agriculture partly_cloudy primary road
40666	test_567.jpg	partly_cloudy primary
40667	test_23428.jpg	agriculture clear primary
40668	test_10099.jpg	clear primary

40669 rows × 2 columns