Prototypical Networks for Few shot Learning in PyTorch

Overview

Prototypical Networks for Few shot Learning in PyTorch

Simple alternative Implementation of Prototypical Networks for Few Shot Learning (paper, code) in PyTorch.

Prototypical Networks

As shown in the reference paper Prototypical Networks are trained to embed samples features in a vectorial space, in particular, at each episode (iteration), a number of samples for a subset of classes are selected and sent through the model, for each subset of class c a number of samples' features (n_support) are used to guess the prototype (their barycentre coordinates in the vectorial space) for that class, so then the distances between the remaining n_query samples and their class barycentre can be minimized.

Prototypical Networks

T-SNE

After training, you can compute the t-SNE for the features generated by the model (not done in this repo, more infos about t-SNE here), this is a sample as shown in the paper.

Reference Paper t-SNE

Omniglot Dataset

Kudos to @ludc for his contribute: https://github.com/pytorch/vision/pull/46. We will use the official dataset when it will be added to torchvision if it doesn't imply big changes to the code.

Dataset splits

We implemented the Vynials splitting method as in [Matching Networks for One Shot Learning]. That sould be the same method used in the paper (in fact I download the split files from the "offical" repo). We then apply the same rotations there described. In this way we should be able to compare results obtained by running this code with results described in the reference paper.

Prototypical Batch Sampler

As described in its PyDoc, this class is used to generate the indexes of each batch for a prototypical training algorithm.

In particular, the object is instantiated by passing the list of the labels for the dataset, the sampler infers then the total number of classes and creates a set of indexes for each class ni the dataset. At each episode the sampler selects n_classes random classes and returns a number (n_support + n_query) of samples indexes for each one of the selected classes.

Prototypical Loss

Compute the loss as in the cited paper, mostly inspired by this code by one of its authors.

In prototypical_loss.py both loss function and loss class à la PyTorch are implemented.

The function takes in input the batch input from the model, samples' ground truths and the number n_suppport of samples to be used as support samples. Episode classes get infered from the target list, n_support samples get randomly extracted for each class, their class barycentres get computed, as well as the distances of each remaining samples' embedding from each class barycentre and the probability of each sample of belonging to each episode class get finmally computed; then the loss is then computed from the wrong predictions probabilities (for the query samples) as usual in classification problems.

Training

Please note that the training code is here just for demonstration purposes.

To train the Protonet on this task, cd into this repo's src root folder and execute:

$ python train.py

The script takes the following command line options:

  • dataset_root: the root directory where tha dataset is stored, default to '../dataset'

  • nepochs: number of epochs to train for, default to 100

  • learning_rate: learning rate for the model, default to 0.001

  • lr_scheduler_step: StepLR learning rate scheduler step, default to 20

  • lr_scheduler_gamma: StepLR learning rate scheduler gamma, default to 0.5

  • iterations: number of episodes per epoch. default to 100

  • classes_per_it_tr: number of random classes per episode for training. default to 60

  • num_support_tr: number of samples per class to use as support for training. default to 5

  • num_query_tr: nnumber of samples per class to use as query for training. default to 5

  • classes_per_it_val: number of random classes per episode for validation. default to 5

  • num_support_val: number of samples per class to use as support for validation. default to 5

  • num_query_val: number of samples per class to use as query for validation. default to 15

  • manual_seed: input for the manual seeds initializations, default to 7

  • cuda: enables cuda (store True)

Running the command without arguments will train the models with the default hyperparamters values (producing results shown above).

Performances

We are trying to reproduce the reference paper performaces, we'll update here our best results.

Model 1-shot (5-way Acc.) 5-shot (5-way Acc.) 1 -shot (20-way Acc.) 5-shot (20-way Acc.)
Reference Paper 98.8% 99.7% 96.0% 98.9%
This repo 98.5%** 99.6%* 95.1%° 98.6%°°

* achieved using default parameters (using --cuda option)

** achieved running python train.py --cuda -nsTr 1 -nsVa 1

° achieved running python train.py --cuda -nsTr 1 -nsVa 1 -cVa 20

°° achieved running python train.py --cuda -nsTr 5 -nsVa 5 -cVa 20

Helpful links

.bib citation

cite the paper as follows (copied-pasted it from arxiv for you):

@article{DBLP:journals/corr/SnellSZ17,
  author    = {Jake Snell and
               Kevin Swersky and
               Richard S. Zemel},
  title     = {Prototypical Networks for Few-shot Learning},
  journal   = {CoRR},
  volume    = {abs/1703.05175},
  year      = {2017},
  url       = {http://arxiv.org/abs/1703.05175},
  archivePrefix = {arXiv},
  eprint    = {1703.05175},
  timestamp = {Wed, 07 Jun 2017 14:41:38 +0200},
  biburl    = {http://dblp.org/rec/bib/journals/corr/SnellSZ17},
  bibsource = {dblp computer science bibliography, http://dblp.org}
}

License

This project is licensed under the MIT License

Copyright (c) 2018 Daniele E. Ciriello, Orobix Srl (www.orobix.com).

Owner
Orobix
Orobix
code for paper"A High-precision Semantic Segmentation Method Combining Adversarial Learning and Attention Mechanism"

PyTorch implementation of UAGAN(U-net Attention Generative Adversarial Networks) This repository contains the source code for the paper "A High-precis

Tong 8 Apr 25, 2022
Photo2cartoon - 人像卡通化探索项目 (photo-to-cartoon translation project)

人像卡通化 (Photo to Cartoon) 中文版 | English Version 该项目为小视科技卡通肖像探索项目。您可使用微信扫描下方二维码或搜索“AI卡通秀”小程序体验卡通化效果。

Minivision_AI 3.5k Dec 30, 2022
Soft actor-critic is a deep reinforcement learning framework for training maximum entropy policies in continuous domains.

This repository is no longer maintained. Please use our new Softlearning package instead. Soft Actor-Critic Soft actor-critic is a deep reinforcement

Tuomas Haarnoja 752 Jan 07, 2023
Official PyTorch implemention of our paper "Learning to Rectify for Robust Learning with Noisy Labels".

WarPI The official PyTorch implemention of our paper "Learning to Rectify for Robust Learning with Noisy Labels". Run python main.py --corruption_type

Haoliang Sun 3 Sep 03, 2022
Learning View Priors for Single-view 3D Reconstruction (CVPR 2019)

Learning View Priors for Single-view 3D Reconstruction (CVPR 2019) This is code for a paper Learning View Priors for Single-view 3D Reconstruction by

Hiroharu Kato 38 Aug 17, 2022
Earthquake detection via fiber optic cables using deep learning

Earthquake detection via fiber optic cables using deep learning Author: Fantine Huot Getting started Update the submodules After cloning the repositor

Fantine 4 Nov 30, 2022
RNG-KBQA: Generation Augmented Iterative Ranking for Knowledge Base Question Answering

RNG-KBQA: Generation Augmented Iterative Ranking for Knowledge Base Question Answering Authors: Xi Ye, Semih Yavuz, Kazuma Hashimoto, Yingbo Zhou and

Salesforce 72 Dec 05, 2022
"MST++: Multi-stage Spectral-wise Transformer for Efficient Spectral Reconstruction" (CVPRW 2022) & (Winner of NTIRE 2022 Challenge on Spectral Reconstruction from RGB)

MST++: Multi-stage Spectral-wise Transformer for Efficient Spectral Reconstruction (CVPRW 2022) Yuanhao Cai, Jing Lin, Zudi Lin, Haoqian Wang, Yulun Z

Yuanhao Cai 274 Jan 05, 2023
Material for my PyConDE & PyData Berlin 2022 Talk "5 Steps to Speed Up Your Data-Analysis on a Single Core"

5 Steps to Speed Up Your Data-Analysis on a Single Core Material for my talk at the PyConDE & PyData Berlin 2022 Description Your data analysis pipeli

Jonathan Striebel 9 Dec 12, 2022
RMTD: Robust Moving Target Defence Against False Data Injection Attacks in Power Grids

RMTD: Robust Moving Target Defence Against False Data Injection Attacks in Power Grids Real-time detection performance. This repo contains the code an

0 Nov 10, 2021
A rule-based log analyzer & filter

Flog 一个根据规则集来处理文本日志的工具。 前言 在日常开发过程中,由于缺乏必要的日志规范,导致很多人乱打一通,一个日志文件夹解压缩后往往有几十万行。 日志泛滥会导致信息密度骤减,给排查问题带来了不小的麻烦。 以前都是用grep之类的工具先挑选出有用的,再逐条进行排查,费时费力。在忍无可忍之后决

上山打老虎 9 Jun 23, 2022
Gesture-Volume-Control - This Python program can adjust the system's volume by using hand gestures

Gesture-Volume-Control This Python program can adjust the system's volume by usi

VatsalAryanBhatanagar 1 Dec 30, 2021
The "breathing k-means" algorithm with datasets and example notebooks

The Breathing K-Means Algorithm (with examples) The Breathing K-Means is an approximation algorithm for the k-means problem that (on average) is bette

Bernd Fritzke 75 Nov 17, 2022
An index of recommendation algorithms that are based on Graph Neural Networks.

An index of recommendation algorithms that are based on Graph Neural Networks.

FIB LAB, Tsinghua University 564 Jan 07, 2023
DISTIL: Deep dIverSified inTeractIve Learning.

DISTIL: Deep dIverSified inTeractIve Learning. An active/inter-active learning library built on py-torch for reducing labeling costs.

decile-team 110 Dec 06, 2022
Keras Implementation of The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation by (Simon Jégou, Michal Drozdzal, David Vazquez, Adriana Romero, Yoshua Bengio)

The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation: Work In Progress, Results can't be replicated yet with the m

Yad Konrad 196 Aug 30, 2022
CC-GENERATOR - A python script for generating CC

CC-GENERATOR A python script for generating CC NOTE: This tool is for Educationa

Lêkzï 6 Oct 14, 2022
(NeurIPS 2021) Pytorch implementation of paper "Re-ranking for image retrieval and transductive few-shot classification"

SSR (NeurIPS 2021) Pytorch implementation of paper "Re-ranking for image retrieval and transductivefew-shot classification" [Paper] [Project webpage]

xshen 29 Dec 06, 2022
Benchmark spaces - Benchmarks of how well different two dimensional spaces work for clustering algorithms

benchmark_spaces Benchmarks of how well different two dimensional spaces work fo

Bram Cohen 6 May 07, 2022
🔥 TensorFlow Code for technical report: "YOLOv3: An Incremental Improvement"

🆕 Are you looking for a new YOLOv3 implemented by TF2.0 ? If you hate the fucking tensorflow1.x very much, no worries! I have implemented a new YOLOv

3.6k Dec 26, 2022