Repository currently contains:
- Baseline models -
- Binary classification - model to classify whether image contains or doesnt contain an elephant;
- Classification models - using pretrained resnet and desnet models that are pretrained on imagenet
- Siamese models - using contrastive loss and triplet loss
- MetaFGNet model - code adopted from - https://github.com/YBZh/MetaFGNet/tree/master/MetaFGNet_with_Sample_Selection
Experiments:
- Pretrained Resent and Densenet using weighted cross entropy
- Pretrained Resent and Densenet using imbalanced data sampler
- Pretrained Resent and Densenet using top 5 classes of the elephant images
- Binary classification model.
- Siamese Network - Contrastive and Triplet Loss
- MetaFGNet model - trained on resnet34 and densenet models.
python3 train.py --batch_size=256 --epochs=2000 --data_path='data/dataset/top5' --output_path='output/output_resnet_sampler_top5_lr_0.0001' --model_path='models/model_resnet_lr_0.0001_sampler_top5' --model_name='resnet' --use_sampler=True --use_top5=True
- Pytorch Tutorials - https://github.com/pytorch/tutorials
- Triplet Loss - https://medium.com/@crimy/one-shot-learning-siamese-networks-and-triplet-loss-with-keras-2885ed022352