Feature Extractor Stacking for Cross-domain Few-shot Meta-learning,
Hongyu Wang, Eibe Frank, Bernhard Pfahringer, Michael Mayo, Geoffrey Holmes
The quickest way to verify the main results is to use cached logits, which can be downloaded from Zenodo:
In the root directory "FeatureExtractorStacking/", please create a directory named "dataset/", and create three child directories for the fine-tuning algorithms "tsa/", "url/", and "flute/". Uncompress the downloaded data into these directories. Every cached episode should have a path such as "FeatureExtractorStacking/datasets/tsa/traffic_sign/0.pt".
Please install the environment with anaconda
conda env create -f environment.yml
To test ConFES and ReFES on the cached episodes, please run
python test.py
Fine-tuning algorithms can be selected by modifying the DATASETS_DIR string in "test.py", which uses TSA logits by default.
In order to make our cached data available, we had to store the logits in half-precision floats. Therefore, the results obtained here may differ slightly from the ones in the paper. We can provide our original full-precision cache if requested.
To only run a subset of the 600 episodes for a quicker demonstration, please edit the dataset length in the "test.py" file, from
def __len__(self):
return len(os.listdir(self.dataset_dir))
to (for example, if we only want 5 episodes)
def __len__(self):
return 5
To plot heatmaps for an episode, please run
python plot.py
The plots will be saved in "heatmaps/". The episode can be selected by editing "plot.py":
CACHEFILENAME = 'datasets/tsa/traffic_sign/0.pt'
We are actively working on code that performs fine-tuning and FES on few-shot episodes with raw images, and saves snapshots and meta-classifiers in files for later practical use.