This repository contains the inceptionV3 model pretrained on iNaturalist dataset. The model is converted from tensorflow implementation of work Large Scale Fine-Grained Categorization and Domain-Specific Transfer Learning (https://github.com/richardaecn/cvpr18-inaturalist-transfer)
The pytorch model architecture is defined in myinception.py.
The converted pretrained model can be found here https://drive.google.com/file/d/1VHcS2o0aYtr1MkYQP_uRgEGcM9bIcM_E/view?usp=sharing
##Training on CUB dataset: To train the network on CUB dataset run the following command:
python train_inception_classifier.py --cuda --save_dir='<PATH_WHERE_MODEL_SAVED>'
Note that this will only train the classifier (last fully connected layer) on CUB dataset (200 classes). However the feature extractor (all previous layers) is taken from pretrained iNaturalist model.
##Testing on CUB dataset:
python test_inception_classifier.py --cuda --load='<PATH_WHERE_MODEL_SAVED>/cub_inception_30.pth'
| Pretrained Model | CUB-200 |
|---|---|
| iNat | 89.1 |