Category-wise Fine-Tuning

• Conference Paper
• Journal Preprint

Introduction

This code demonstrates using CFT_BP_ASL and CFT_GA to fine-tune models trained with Assume Negative (AN) (i.e., Negative mode) on the MS-COCO dataset.

1. Download the MS-COCO 2014 dataset.

2. Put the training image folder (train2014), validation image folder (val2014), and annotation folder (annotations) under the same folder (eg., COCO) like this:

    -----COCO  
       |----train2014  
       |----val2014  
       |----annotations
   

3. Run pip install -r requirements.py to install the necessary packages.

4. Config the dataset location in config.py.

5. Config the known label proportion in config.py.

6. Run python train.py. This will train a classification model with AN and save the trained model in output/train/best.pth.

7. Run CFT_prepare.py. This will generate and store (feature vector z, label y) pairs of the dataset to output/CFT/cache for preparation of CFT. This generation can dramatically speed up CFT.

8. Run CFT_optimize.py. This will respectively use CFT_BP_ASL and CFT_GA to fine-tune the train model. The parameters of the fine-tuned classification layer will be saved in output/CFT.

9. Run validate.py to see the classification performances of the trained model, the trained model after CFT_BP_ASL, and the trained model after CFT_GA. The result is saved in output/valid/logs.

Acknowledgment

Part of the codes in this repository are from:

• Alibaba-MIIL/ASL
• Alibaba-MIIL/PartialLabelingCSL