Code for Paper: Few-Shot Non-Parametric Learning with Deep Latent Variable Model

This paper is accepted to NeurIPS 2022.
Below is the instruction for doing classification with trained generative models.

Requirements

python>=3.7
numpy
torch
torchvision
pillow
tqdm
tensorboardX
scikit-learn

or

pip install -r requirements

Download Trained Generative Models

wget https://www.dropbox.com/s/2426ilsu2jjuykf/params.zip

unzip and put params folder under neural_compressor/model

NPC

python main.py --dataset mnist --compressor gzip --data_dir data --distance NCD --shot 10

NPC-LV

NPC-LV consists of 3 steps:

1. Train a latent variable generative model
2. Combined ANS with trained generative model to form a compressor
3. Use compressor and compressor-based distance metric for classification

Step 2&3: Classification

This repo contains trained model parameters so we can do classification directly using the command below. By default, only 100 test samples and 100 training samples will be used, which will take about half an hour to run.

python main.py --compressor bbans --online --dataset mnist --shot 10

To use the same test indicies as the paper, which will include 1000 test samples, use --replicate. This process will require ~4 hours on 10-shot.

Below is the table of the result reported in the paper:

	MNIST	FashionMNIST	CIFAR-10
5-shot	77.6+-0.4	74.1+-3.2	35.3+-2.9
10-shot	84.6+-2.1	77.2+-2.2	36.0+-1.8
50-shot	91.4+-0.6	83.2+-0.7	37.4+-1.2
100-shot	93.6	84.5	40.2

Step 1 (Optional): Train a Generative Model

cd neural_compressor
python -m model.mnist_train --nz=2 --width=63

neural_compressor contains files for training a hierarchical VAE. By default, trained generative model's parameters are saved under neural_compressor/model/params/${dataset}/nz2.

NPC-LV (large scale)

To replicate the experiment with 1000 test samples and 1000 training samples, we provide the compressed files to run knn directly, as compression takes a long time (1-2days). Compressed files can be downloaded from here (~2.1G).

After downloading and unzipping, pass compressed directories to command, e.g.,:

python main.py --compressor bbans --dataset mnist --replicate --c_train_dir ../state_array/mnist/nz2/train1000 --c_test_dir ../state_array/mnist/nz2/test1000 --c_combined_dir ../state_array/mnist/nz2/avg/test1000_train1000

Similarly, to replicate paper's result for fewer shot, add --shot 10.