bertinetto / r2d2 Goto Github PK

View Code? Open in Web Editor NEW

114.0 10.0 18.0 37 KB

[ICLR'19] Meta-learning with differentiable closed-form solvers

Home Page: http://www.robots.ox.ac.uk/~luca/r2d2.html

License: MIT License

Python 85.83% Shell 14.17%

iclr2019 few-shot-learning meta-learning classification deep-learning machine-learning computer-vision pytorch

r2d2's Introduction

Meta-learning with differentiable closed-form solvers.

Paper (published at ICLR 2019)

arXiv: https://arxiv.org/abs/1805.08136
OpenReview: https://openreview.net/forum?id=HyxnZh0ct7

Please refer to it as:

@inproceedings{
bertinetto2018metalearning,
title={Meta-learning with differentiable closed-form solvers},
author={Luca Bertinetto and Joao F. Henriques and Philip Torr and Andrea Vedaldi},
booktitle={International Conference on Learning Representations},
year={2019},
url={https://openreview.net/forum?id=HyxnZh0ct7},
}

Data setup

In scripts/train/conf/fewshots.yaml, specify the location of your custom $DATASET_PATH (data.root_dir).
Download Omniglot, CIFAR-FS and miniImageNet the above format. Original datasets from here and here.
Download and extract one or more datasets in your custom $DATASET_PATH folder, the code assumes the following structure (example):

$DATASET_PATH
├── miniimagenet
│   ├── data
│   │   ├── n01532829
|   |   |── ...
│   │   └── n13133613
│   ├── splits
│   │   └── ravi-larochelle
|   |   |   ├── train.txt
|   |   |   ├── val.txt
|   |   |   └── test.txt
├── omniglot
|   ...
├── cifarfs 
|   ...

Repo setup (with Conda)

Set up conda environment: conda env create -f environment.yml.
source activate fsrr
Install torchnet: pip install git+https://github.com/pytorch/tnt.git@master.
Install the repo package: pip install -e .
source deactivate fsrr

Run

scripts/train/experiments.sh contains all the experiments of the paper (train+eval) in blocks of three lines, e.g.

expm_folder=mini_r2d2 
python run_train.py --log.exp_dir $expm_folder --data.dataset miniimagenet --data.way 16 --model.drop 0.1 --base_learner.init_adj_scale 1e-4 
python ../eval/run_eval.py --data.test_episodes 10000 --data.test_way 5 --data.test_shot 1 --model.model_path ../train/results/$expm_folder/best_model.1shot.t7 
python ../eval/run_eval.py --data.test_episodes 10000 --data.test_way 5 --data.test_shot 5 --model.model_path ../train/results/$expm_folder/best_model.5shot.t7

Note

Some of the files of this repository (e.g. data loading and training boilerplate routines) are the result of a modification of prototypical networks code and contain a statement in their header.

r2d2's People

Contributors

Stargazers

Watchers

Forkers

abdelpakey b2220333 peterzcc mata62n xiaoanshi hyzcn arnoutdevos chunyanzhao ryutian xhchrn chenzl23 zialiu chrisyxue weiyucheng jw447 nahad3 lyapunovstability zixind

r2d2's Issues

run problem

When I run the ‘scripts/train/experiments.sh’ command, I get a ‘no run_train.py’ or 'No module named 'fewshots'' problem. Is it wrong with me running the directory? Or is it a directory structure problem with the code?

If more than one value is outlined, it means their intervals overlap

Dear author,“I meant they overlap when considering the ranges of performance between mean-confidence to mean+confidence.“According to your opinion，so what do you make of these sets of data？

Error when running training

Hi,

I am trying to run:

expm_folder=mini_lrd2_iter1
python run_train.py --data.cuda True --log.exp_dir $expm_folder --data.dataset miniimagenet --data.way 10 --data.test_way 5 --model.drop 0.1 --data.batch_size 200 --model.groupnorm False --train.learning$

And get the following error:

RuntimeError: expand(torch.cuda.FloatTensor{[140, 1]}, size=[140]): the number of sizes provided (1) must be greater or equal to the number of dimensions in the tensor (2)

Code about SGD learning of CNN, alpha, beta and lambda

Thank you for your published code, it seems that the code does not contain the SGD learning part of CNN, alpha, beta and lambda. Could you talk about the implementation or release those operations? Thank you!

If more than one value is outlined, it means their intervals overlap.

Dear author, I have seen the overlapping part of multiple value intervals in the paper, but I don't quite understand what this means. Could you please explain it?

Invalid data download link

Hi, thank you so much for sharing the code!
The links for downloading miniImagenet and Omniglot are invalid (not the official ones).
Espicially for the miniImage one, how am I supposed to get the data as well as the txt files?
Thanks!

probelm when running

Hi luca,
I set up the repo like that in readme.But when I goes to "Install the repo package: pip install -e .", my terminal says"-e option requires 1 argument". What does this mean?

Accuracy gaps on CIFAR-FS dataset

Hi, thank you for sharing these well-written codes!

When I ran experiments with default setting on CIFAR-FS dataset, I notice quite large performance gaps compared to the results reported in the paper (near 4% for both 5-shot and 1-shot configurations). For the default setting, I mean the setting described in scripts/train/conf/fewshots.yaml

I checked the setting described in the paper I think it's consistent with the default setting in the repo. If the default setting is not the optimal, could you provide the optimal settings for the three datasets?

Best regards