iMaterialist Challenge (FGVC6, 2019), 13th place
iMaterialist Challenge on Product Recognition (FGVC6, CVPR 2019): kaggle page
I use the following 5 CNN backbones pretrained on ImageNet. All pretrained models are from pretrained-models.pytorch.
- NASNet-A-Large
- SENet154
- InceptionResNet-v2
- ResNet152
- ResNet101
- Learning rate: 4e-4
- Weight decay: 5e-5
- Epoch: all models converge in around 10 epochs
Nothing special, average of all probabilities.