datamatica - Telecommunication towers component clasification

Introduction

The first attempt for developing this idea started as the homework assignment for lesson 1: Deep Learning 2019 (v3) fast.ai course.

For that homework I toke my domain expertise on telecommunication towers to build an image clasifier which could hopefuly recognize different tower components.

In may 2022 I was coursing live the 2022 version of the course, which is done in collaboration with The University of Queensland and its now called Practical Deep Learning for Coders.

I’m running this notebook on an old GTX-1070 NVIDIA GPU.

Import Libraries

from fastai.vision.all import *
import timm
import plotly.express as px

Dataset

I curated an image dataset of multiple towers I worked with in the past several years. And choosed 514 images in 8 relatively “easy” to distinguish categories (components).

The dataset was stored in google drive and is shared here.

Local Dataset

For the 2022 course, thanks to the help of many great people in fastai forums, I was able to install fastai locally and to use my local GPU on WSL2.

path = Path("photos")

train_path = path / 'train'
valid_path = path / 'valid'

labels = [label.parts[-1] for label in train_path.iterdir()]
train_quantity = [len(list(each.iterdir())) for each in train_path.iterdir()]
valid_quantity = [len(list(each.iterdir())) for each in valid_path.iterdir()]
df = pd.DataFrame()

df['label'] = labels * 2
df['set'] = ['train'] * 8 + ['valid'] * 8
df['quantity'] = train_quantity + valid_quantity

df

	label	set	quantity
0	base_plate	train	87
1	grounding_bar	train	52
2	identification	train	30
3	ladder	train	37
4	light	train	69
5	lightning_rod	train	33
6	platform	train	57
7	transmission_lines	train	29
8	base_plate	valid	29
9	grounding_bar	valid	15
10	identification	valid	8
11	ladder	valid	11
12	light	valid	22
13	lightning_rod	valid	10
14	platform	valid	16
15	transmission_lines	valid	9

is_train = df.set == 'train'
df[is_train].quantity.sum(), df[~is_train].quantity.sum()

(394, 120)

fig = px.bar(
    df, x="set", y="quantity",
    color='label', barmode='group',
    height=400
)
fig.show()

The Data

The data was hand picked from a huge tower photoset. To start, I choose these eight easy distinguishable components to be clasified:

Base plate
Grounding bar
Identification
Ladder
Light
Lightning rod
Platform
Transmission lines

There are two folders, one for the training (train) and the other for the validation set (valid).

print(path.ls())   
print('*'*100)
(path/'train').ls()

[Path('photos/train'), Path('photos/valid')]
****************************************************************************************************

(#8) [Path('photos/train/base_plate'),Path('photos/train/grounding_bar'),Path('photos/train/identification'),Path('photos/train/ladder'),Path('photos/train/light'),Path('photos/train/lightning_rod'),Path('photos/train/platform'),Path('photos/train/transmission_lines')]

tower_parts_fns = get_image_files(path)
tower_parts_fns

(#514) [Path('photos/train/base_plate/Ac102-Corozopando-(64).jpg'),Path('photos/train/base_plate/Ac102-Corozopando-(75).jpg'),Path('photos/train/base_plate/camaguan-087.jpg'),Path('photos/train/base_plate/camaguan-098.jpg'),Path('photos/train/base_plate/cantv el yoco 015.JPG'),Path('photos/train/base_plate/cantv-capanaparo-011.jpg'),Path('photos/train/base_plate/cantv-cinaruco-018.jpg'),Path('photos/train/base_plate/cantv-cinaruco-025.jpg'),Path('photos/train/base_plate/cartanal-(7).jpg'),Path('photos/train/base_plate/CHUSPITA-II-AC-72-MTS-002.jpg')...]

failed = verify_images(tower_parts_fns)
print(failed)

[]

DataLoaders

tower_parts = DataBlock(
    blocks=(ImageBlock, CategoryBlock), 
    get_items=get_image_files, 
    splitter=GrandparentSplitter(train_name='train', valid_name='valid'),
    get_y=parent_label,
    item_tfms=Resize(224)
)

dls = tower_parts.dataloaders(path)

%%time
for _ in dls.train: pass

CPU times: user 449 ms, sys: 506 ms, total: 955 ms
Wall time: 20.4 s

dls.train.show_batch(max_n=16, nrows=4, figsize=(10,10))

Learner

learn = vision_learner(dls, resnet34, metrics=error_rate)

fastai’s `lr_find`

%%time
learn.lr_find()

CPU times: user 24.4 s, sys: 8.72 s, total: 33.2 s
Wall time: 5min 38s

SuggestedLRs(valley=0.0014454397605732083)

dls.num_workers

Note

I experimented by setting dls.num_workers = 4 and it didn’t make any difference in the time it takes to run lr_find(), even though that, by watching at the progress bar, it seemed that the bottleneck was in pre-processing the batch. Not in GPU.

len(dls.train), len(dls.train.get_idxs())

(6, 394)

len(dls.valid), len(dls.valid.get_idxs())

(2, 120)

dls.drop_last

True

Training

%%time
learn.fine_tune(3)

epoch	train_loss	valid_loss	error_rate	time
0	2.751579	0.932106	0.316667	00:44

epoch	train_loss	valid_loss	error_rate	time
0	1.000472	0.463910	0.166667	00:43
1	0.632878	0.272529	0.091667	00:43
2	0.429861	0.208375	0.050000	00:43

CPU times: user 8.31 s, sys: 4.01 s, total: 12.3 s
Wall time: 2min 55s

interp = ClassificationInterpretation.from_learner(learn)
interp.plot_confusion_matrix(figsize=(8,8))

interp.plot_top_losses(9)

learn.export('models/tower_parts_model')
learn.save("exported_model_from_fastai", with_opt=False)

Path('models/exported_model_from_fastai.pth')

Conclusions

I took about 500 pictures and 3 epochs to finetune a small pre-trained model to make it recognize 8 components with an error rate of about 6%.
learn.lr_find() took about 5 minutes to run, more that the fine tuning.
There are some categories that normally appear in one picture at the same time. I solved a classification problem to simplify, but maybe the actual problem should be a multi-class classificacion.