Fine-Grained Visual Categorization using Meta-Learning Optimization with Sample Selection of Auxiliary Data
The source code of the ECCV 2018 paper: Fine-Grained Visual Categorization using Meta-Learning Optimization with Sample Selection of Auxiliary Data. Paper available at: https://arxiv.org/abs/1807.10916
The code is used to train a MetaFGNet with L-Bird Subset and CUB-200-2011 dataset as the source and target dataset respectively. The extention to other source and target datasets is direct.
It concludes five parts:
- L_Bird_pretrain: Train a model for the classification task of L-Bird Subset based on the model that pre-trained on the ImageNet.
- MetaFGNet_without_Sample_Selection: Train the MetaFGNet without sample selection of L_Bird Subset
- Sample_Selection: Select the target-related samples from L_Bird Subset
- MetaFGNet_with_Sample_Selection: Train the MetaFGNet with sample selection of L_Bird Subset
- Fine_tune_for_final_results: Fine-tune the MetaFGNet model on the target dataset for better and final result.
We split the whole program into five parts for better understanding and reuse.
- The 'regularized meta-learning objective' is implemented in the second part and the fourth part.
- The proposed sample selection method is implemented in the third part.
We also provide some intermediate results for quickly implementation and verification. They can be downloaded from:
This code is completed with the cooperation of Hui Tang
We provide a supplementary material to demonstrate how to evaluate the gradient of meta-learning loss at Gradient for Meta-Learning Loss
If you have any questions, feel free to contact me at: firstname.lastname@example.org.