Wildlife Embeddings

A pytorch lightning implementation of the WBIA Piev2 Plugin: https://github.com/WildMeOrg/wbia-plugin-pie-v2.

Data Format

A csv file should be created for your dataset with the following columns:

• annot: unique integer identifier for each datapoint
• image: image name (not full path)
• name: individual name (incomparable sides of the same individual should have different names)
• encounter: annotations with the same encounter will not be compared during validation or testing. If unsure set equal to the annot column.
• x: left bounding box coordinate
• y: top bounding box coordinate
• w: bounding box width
• h: bounding box height
• theta: rotation of the bounding box

Training

During training we evaluate the model on the validation data every two epochs and retain the checkpoint with the highest 1-vs-all top1 accuracy.

Arguments (defaults)

• --name: name for logging

• --version: version number for logging

• --data-file: path to the data csv file

• --data-dir: path to the image directory

• --eval-cutoff: training is done with individuals with > eval-cutoff encounters, the rest are used for validation/testing

• --image-size: input image size (256)

• --train-transforms: data augmentation for training, see src/data/transforms.py (resize, affine, color_jitter, grayscale, blur, center_crop, normalize)

• --eval-transforms: data augmentation for validation/testing (resize, center_crop, normalize)

• --num-copies: see src/data/sampler.py (4)

• --num-instances: see src/data/sampler.py (4)

• --batch-size: batch size per gpu (64)

• --num-workers: number of dataloader workers (8)

• --embedding-dim: output embedding size (512)

• --lr: learning rate (1e-5)

• --wd: weight decay (5e-4)

• --fixbase-epoch: freeze the weights of the model excluding fully-connected layers for this many epochs during training (1)

• --margin: triplet loss margin (0.3)

• --weight-t: triplet loss weight (1.0)

• --weight-x: cross-entropy loss weight (1.0)

• --gpus: number of gpus

• --max_epochs: maximum number of epochs

Also included are all the pytorch-lightning trainer flags: https://pytorch-lightning.readthedocs.io/en/latest/api/pytorch_lightning.trainer.trainer.Trainer.html#pytorch_lightning.trainer.trainer.Trainer