Skip to main content

How-To Guides

Installation​

pip install -r requirements.txt

Species classification on a single image input​

Note: Some sample camera trap species images are available in the dir. data/sample_images/. During training, we utlize multiple heterogeneous contexts in the multimodal KG. However, at inference time, only the image is used to perform species classification.

  1. Download the required data bash download_data.sh.
  2. Evaluate the pretrained model on a given image.
python eval_image.py --ckpt-path <PATH TO TRAINED CKPT> --img-path <PATH TO IMG FILE>

Data Preprocessing​

This will download the iWildCam2020-WILDS dataset and Open Tree of Life taxonomy and pre-process them.

bash preprocess.sh

Note: The dir. data/iwildcam_v2.0/train/ contains images for all splits.

Training a model from scratch​

We consider different training settings that comprise of combination of different context types such as taxonomy, location, and time.

Train using only images linked to species labels​

python -u main.py --data-dir data/iwildcam_v2.0/ --img-dir data/iwildcam_v2.0/train/ --save-dir CKPT_DIR > CKPT_DIR/log.txt

Train using species labels and taxonomy contexts​

python -u main.py --data-dir data/iwildcam_v2.0/ --img-dir data/iwildcam_v2.0/train/ --save-dir CKPT_DIR --add-id-id > CKPT_DIR/log.txt

Train using species labels and location contexts​

python -u main.py --data-dir data/iwildcam_v2.0/ --img-dir data/iwildcam_v2.0/train/ --save-dir CKPT_DIR --add-image-location > CKPT_DIR/log.txt

Train using species labels and time contexts​

python -u main.py --data-dir data/iwildcam_v2.0/ --img-dir data/iwildcam_v2.0/train/ --save-dir CKPT_DIR --add-image-time > CKPT_DIR/log.txt

Train using species labels, taxonomy, and time contexts​

python -u main.py --data-dir data/iwildcam_v2.0/ --img-dir data/iwildcam_v2.0/train/ --save-dir CKPT_DIR --add-id-id --add-image-time > CKPT_DIR/log.txt

Train using species labels, taxonomy, and location contexts​

python -u main.py --data-dir data/iwildcam_v2.0/ --img-dir data/iwildcam_v2.0/train/ --save-dir CKPT_DIR --add-id-id --add-image-location > CKPT_DIR/log.txt

evaluate a model (specify split=val/test/id_val/id_test)​

python eval.py --ckpt-path <PATH TO TRAINED CKPT> --split test --data-dir data/iwildcam_v2.0/ --img-dir data/iwildcam_v2.0/train/