classification

Image classification

The goal here is to propose a training script and a dataset to train a wildfire classification model.

Setup

Python 3.6 (or higher), pip and Git are required to train your models with PyroVision. Install the training-specific dependencies:

git clone https://github.com/pyronear/pyro-vision.git
pip install -e "pyro-vision/.[training]"

Quick Tour

The script comes with multiples arguments that you can explore:

python references/classification/train.py --help

OpenFire

You can also use freely our open-source dataset:

python references/classification/train.py path/to/dataset/folder --openfire --arch rexnet1_0x --lr 1e-3 -b 32 --grad-acc 2 --epochs 100 --device 0 --prefetch-size 512

If you prefer to run this in Google Colab, we have a starter notebook for you!

Custom datasets

When we use datasets where we are not owners of the data, we unfortunately cannot share them publicly. However, we still want you to be able to train on your own dataset. If you intend to do so, your dataset should follow the folder's hierarchy below:

CustomDataset
├── train
│   └── images
│       ├── 0
│       │   ├── no_fire_train_image_first.jpg
│       │   ├── ...
│       │   └── no_fire_train_image_last.jpg
│       └── 1
│           ├── fire_train_image_first.jpg
│           ├── ...
│           └── fire_train_image_last.jpg
└── val
    └── images
        ├── 0
        │   ├── no_fire_val_image_first.jpg
        │   ├── ...
        │   └── no_fire_val_image_last.jpg
        └── 1
            ├── fire_val_image_first.jpg
            ├── ...
            └── fire_val_image_last.jpg

Once this is the case, you can train your model in a similar fashion as for OpenFire:

python train.py path/to/dataset/folder --model rexnet1_0x --lr 1e-3 -b 16 --epochs 20 --device 0

Available architectures

The list of supported architectures is available here.