Planet UAS dataset: Example of the aitlas toolbox for multi label image classification

This notebook shows a sample implementation of a multi label image classification using the aitlas toolbox using the Planet UAS multi label dataset.

from aitlas.datasets import PlanetUASMultiLabelDataset
from aitlas.models import ResNet50MultiLabel
from aitlas.transforms import ResizeCenterCropFlipHVToTensor, ConvertToRGBResizeCenterCropToTensor
from aitlas.utils import image_loader

Load the dataset

dataset_config = {
    "data_dir": "./data/PlanetUAS/images",
    "csv_file": "./data/PlanetUAS/multilabels.txt"
}
dataset = PlanetUASMultiLabelDataset(dataset_config)

Show images from the dataset

fig1 = dataset.show_image(1000)
fig2 = dataset.show_image(30)
fig3 = dataset.show_batch(15)

Inspect the data

dataset.show_samples()

dataset.data_distribution_table()

fig = dataset.data_distribution_barchart()

Load train and test splits

train_dataset_config = {
    "batch_size": 16,
    "shuffle": True,
    "num_workers": 4,
    "data_dir": "./data/PlanetUAS/images",
    "csv_file": "./data/PlanetUAS/train.csv"
}

train_dataset = PlanetUASMultiLabelDataset(train_dataset_config)
train_dataset.transform = ResizeCenterCropFlipHVToTensor()

Setup and create the model for training

epochs = 10
model_directory = "./experiments/PlanetUAS"
model_config = {
    "num_classes": 17,
    "learning_rate": 0.0001,
    "pretrained": True,
    "threshold": 0.5,
    "metrics": ["accuracy", "precision", "recall", "f1_score"]
}
model = ResNet50MultiLabel(model_config)
model.prepare()

Training and evaluation

model.train(
    train_dataset=train_dataset,
    epochs=epochs,
    model_directory=model_directory,
    run_id='1',
)

Predictions

model_path = "./experiments/PlanetUAS/checkpoint.pth.tar"
#labels = PlanetUASMultiLabelDataset.labels
labels = ["haze", "primary", "agriculture", "clear", "water", "habitation", "road", "cultivation", "slash_burn",
          "cloudy", "partly_cloudy", "conventional_mine", "bare_ground", "artisinal_mine", "blooming",
          "selective_logging", "blow_down"]
transform = ConvertToRGBResizeCenterCropToTensor()
model.load_model(model_path)

image = image_loader('./data/predict/image1.tif')
fig = model.predict_image(image, labels, transform)

image = image_loader('./data/predict/image2.tif')
fig = model.predict_image(image, labels, transform)

image = image_loader('./data/predict/image3.tif')
fig = model.predict_image(image, labels, transform)

image = image_loader('./data/predict/image4.tif')
fig = model.predict_image(image, labels, transform)