TAFSSL

Task-Adaptive Feature Sub-Space Learning for few-shot classification

https://arxiv.org/abs/2003.06670

Code for the experiments

(according to tables and figures in the paper):

Table 1: Transductive setting

python exp_table.py

Table 2: Semi supervised setting

python exp_semi.py

Figure 2: Number of queries in transductive FSL setting

python exp_num_query.py

Figure 3: The affect of the unlabeled data noise on the performance

python exp_noise_semi.py

Figure 4: ICA dimension vs accuracy

python exp_projection_dim.py

Figure 5: Unbalanced

python exp_unbalanced.py

To re-create the feature files:

1. Download miniImageNet / tieredImageNet
2. Generate splits:
   python src/utils/tieredImagenet.py --data path-to-tiered --split split/tiered/
3. Pre-train a model and store the features:
   python ./src/train.py -c $<$path to config file$>$
   python ./src/train.py -c $<$path to config file$>$ --save-features --enlarge

moseslichten/TAFSSL