This code finetunes an Inception V3 model on the iNaturalist 2017 competition dataset. You can read about the results in this blog post.
The network was trained on Ubuntu 16.04 using PyTorch 0.1.12. Each training epoch took about two hours using a GTX 1080. The links for the raw data are available here. We also provide a pretrained model that can be downloaded from here. Every epoch the code will save a checkpoint and also save the current best model according to validation accuracy.
Training for about 70 epochs (note the model pretty much converges after 30 epochs) results in a top one accuracy of 64.44% and and top five of 85.34% on the validation set. Below are the results broken down by super category.
| Super Category | Num Classes | Val Top 5 Acc (%) |
|---|---|---|
| Plantae | 2,101 | 85.96 |
| Insecta | 1,021 | 89.66 |
| Aves | 964 | 85.40 |
| Reptilia | 289 | 74.56 |
| Mammalia | 186 | 79.48 |
| Fungi | 121 | 89.44 |
| Amphibia | 115 | 75.09 |
| Mollusca | 93 | 83.00 |
| Animalia | 77 | 85.39 |
| Arachnida | 56 | 88.95 |
| Actinopterygii | 53 | 79.90 |
| Chromista | 9 | 81.94 |
| Protozoa | 4 | 91.78 |
By setting the following flags it's possible to generate a submission file for the competition.
evaluate = True
save_preds = True
resume = 'model_path/iNat_2017_InceptionV3.pth.tar' # path to trained model
val_file = 'ann_path/test2017_3_public_use.json' # path to test file
rootdir = 'test_ims_path/ignat_test_images_2017/' # path to test images
op_file_name = 'inat2017_test_preds.csv' # submission filename