twitter-research / meta-learning-lstm Goto Github PK

View Code? Open in Web Editor NEW

363.0 17.0 162.0 297 KB

This repo contains the source code accompanying a scientific paper with the same name.

License: MIT License

Lua 90.76% Python 9.24%

meta-learning-lstm's Introduction

meta-learning-lstm

This repo contains the code for the following paper: https://openreview.net/pdf?id=rJY0-Kcll

Dependencies

The following libaries are necessary:

torch-autograd
torch-ipc (use version from commit 'c1b2984c4c2dae085005d385996f4c0660173b27')
torch-Dataset
moses

Training

Splits corresponding to meta-training, meta-validation, and meta-testing are placed in data/miniImagenet/. Download corresponding imagenet images and place in folder called images and place folder in data/miniImagenet/.

To train a model:

th train/run-train.lua --task [1-shot or 5-shot task] --data config.imagenet --model [model name]

For example, to run matching-nets:

th train/run-train.lua --task config.5-shot-5-class --data config.imagenet --model config.baselines.train-matching-net

And, to run LSTM meta-learner for 5-shot task:

th train/run-train.lua --task config.5-shot-5-class --data config.imagenet --model config.lstm.train-imagenet-5shot

Contact

For questions about miniImagenet format, please contact Sachin Ravi at email given in the paper.

meta-learning-lstm's People

Contributors

Stargazers

Watchers

Forkers

alexxnica kryndex benjamesbabala jdc08161063 shiyongde pizpizpie xiongweiwu rintukutum yiliangnie ml-lab shuangjiexu katyprogrammer tvkpz december-boy aojunzhou fenoms vgsatorras lim0606 haofusheng andrewhess jayvischeng changebio shuiyiwensi ankitrock004 ecada spencerx pandinosaurus b2220333 haiminzhang skmkedar christinaliang linranran wangxueliangustc bigdatasciencegroup guangxxie zhangxlaurence iamsile zizhao pinglmlcv megayeye gccnosharp suvarna-kadam chenghao-ch94 bai-yeqi zihaozhang9 caihengyu520 irelia7u sudarshan1994 wangyuchn cindyguyuehu123 wpfhtl thoamsdong godencrystal wujinlonglovezhangmiao1314 cheer00 yuanwanglll xiangriconglin kwanegx yikeqicn ymleifdu aprameyabharadwaj szitha qinglu0330 ryany1994 jufangshen jakezdk vishwajitvishnu oscarcarli mtminus sunhongmin225 ahmedfrikha 952942497 brando90 anujshah1003 hyoje42 thanhkaist jxw2018 eric-moreno overgter liuchaoxd shengzhang90 nan2020 timchen0618 shaotengliu shourav zerozhupan jiunhaojhan hyperpotatoneo guardian-in-the-wf keyufeng rees-c aslyc zwb-1224171056 colinmatthewgeorge87 stories-z seamus-lee ychen041 jaeikyoo glare-ni lightconewja

meta-learning-lstm's Issues

erroc

Using data-loader: dataset.data-loader
data dir: /home/eric/meta-learning-lstm-master/data/miniImagenet/miniImagenet/images
/home/eric/torch/install/bin/luajit: ./dataset/miniImagenet.lua:14: attempt to index local 'fid' (a nil value)
stack traceback:
./dataset/miniImagenet.lua:14: in function 'loadSplit'
./dataset/miniImagenet.lua:142: in function 'f'
/home/eric/torch/install/share/lua/5.1/moses.lua:104: in function 'each'
./dataset/miniImagenet.lua:140: in function <./dataset/miniImagenet.lua:111>
./dataset/data-loader.lua:34: in function <./dataset/data-loader.lua:13>
./train/train.lua:26: in function <./train/train.lua:6>
train/run-train.lua:16: in main chunk
[C]: in function 'dofile'
...eric/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk
[C]: at 0x00405d50
what does this mean?
what should i do?

Access the miniImageNet dataset

Hi, I wonder how I can download the miniImageNet dataset. For some filenames in the csv file, I cannot find its corresponding url using the ImageNet api (this should return the (filename, url) pairs): http://www.image-net.org/api/text/imagenet.synset.geturls.getmapping?wnid=[wnid].

For example, "n0185567200000003.jpg" is the first one in val.csv, but when I replace the [wnid] with n01855672 to query the api, I cannot find any filename called "n01855672_3" or "n01855672_00000003". So I wonder if your filename is exactly the same as the original one in ImageNet. If not, can you tell me the mapping? Or can you upload the dataset somewhere and share it? Thanks!

why does it print nothing

when i run "th train/run-train.lua --task config.5-shot-5-class --data config.imagenet --model config.lstm.train-imagenet-5shot"
it just print
" Learner nParams: 32901
Meta-learner params
{
1 :
{
W : CudaTensor - size: 24x80
b : CudaTensor - size: 1x80
}
2 :
{
WI : CudaTensor - size: 22x1
cI : CudaTensor - size: 32901x1
WF : CudaTensor - size: 22x1
bF : CudaTensor - size: 1x1
bI : CudaTensor - size: 1x1
}
}"
what's the result of the train?

Attempt to serialize closure './dataset/miniImagenet.lua:31' with 1 upvalues; first is: '_'.

Using data-loader: dataset.data-loader
data dir: /home/eric/meta-learning-lstm-master/data/miniImagenet/miniImagenet/images
/home/eric/torch/install/bin/luajit: /home/eric/torch/install/share/lua/5.1/dataset/Reader.lua:52: Attempt to serialize closure './dataset/miniImagenet.lua:31' with 1 upvalues; first is: '_'.
Consider using ipc.workqueue.writeup() instead.
stack traceback:
[C]: in function 'map'
/home/eric/torch/install/share/lua/5.1/dataset/Reader.lua:52: in function 'Reader'
/home/eric/torch/install/share/lua/5.1/dataset/Dataset.lua:71: in function 'sampledBatcher'
./dataset/miniImagenet.lua:147: in function 'f'
/home/eric/torch/install/share/lua/5.1/moses.lua:104: in function 'each'
./dataset/miniImagenet.lua:140: in function <./dataset/miniImagenet.lua:111>
./dataset/data-loader.lua:34: in function <./dataset/data-loader.lua:13>
./train/train.lua:26: in function <./train/train.lua:6>
train/run-train.lua:16: in main chunk
[C]: in function 'dofile'
...eric/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk
[C]: at 0x00405d50
I do not know how to deal with it.anyone can help ?

miniimagenet access?

Hi, I was wondering where I can download the mini_imagenet dataset used in your paper? thanks.

dataset

Hi,
Thanks for you sharing.
Where can I download corresponding imagenet images?

What is the difference between learning to learn by gradient descent paper ?

Hello

Nice to read your work. However, i am confused between the difference between your work and Learning to Learn by Gradient Descent paper. I think this two idea is the same. Can you tell me the difference between this two paper?

mini-imagenet

How to load mini-imagenet dataset? Thank you

./dataset/miniImagenet.lua:14: attempt to index local 'fid' (a nil value) stack traceback:

mldl@mldlUB1604:/ub16_prj/meta-learning-lstm$
mldl@mldlUB1604:/ub16_prj/meta-learning-lstm$ th train/run-train.lua --task config.5-shot-5-class --data config.imagenet --model config.baselines.train-matching-net
Training with options: {
dataLoader : "dataset.data-loader"
episodeSamplerKind : "permutation"
dataName : "dataset.miniImagenet"
nTest :
{
1 : 100
2 : 250
3 : 600
}
batchSize : 75
nDepth : 3
nTrainShot : 5
nIn : 84
model : "config.baselines.train-matching-net"
nTestShot :
{
1 : 1
2 : 5
}
useCUDA : true
printPer : 1000
rawDataDir : "data/miniImagenet/"
useDropout : true
test : "-"
task : "config.5-shot-5-class"
metaLearner : "model.baselines.matching-net"
nEpisode : 75000
embedModel : "model.baselines.simple-embedding"
learningRate : 0.0001
optimMethod : "adam"
nValidationEpisode : 100
steps : 3
learner : "model.matching-net-classifier"
nClasses :
{
test : 5
train : 5
val : 5
}
nExamples : 20
nEval : 15
data : "config.imagenet"
}
Using data-loader: dataset.data-loader
unzipping: /home/mldl/ub16_prj/meta-learning-lstm/data/miniImagenet/miniImagenet/images.zip
unzip: cannot find or open /home/mldl/ub16_prj/meta-learning-lstm/data/miniImagenet/miniImagenet/images.zip, /home/mldl/ub16_prj/meta-learning-lstm/data/miniImagenet/miniImagenet/images.zip.zip or /home/mldl/ub16_prj/meta-learning-lstm/data/miniImagenet/miniImagenet/images.zip.ZIP.
data dir: /home/mldl/ub16_prj/meta-learning-lstm/data/miniImagenet/miniImagenet/images
/home/mldl/torch/install/bin/luajit: ./dataset/miniImagenet.lua:14: attempt to index local 'fid' (a nil value)
stack traceback:
./dataset/miniImagenet.lua:14: in function 'loadSplit'
./dataset/miniImagenet.lua:142: in function 'f'
/home/mldl/torch/install/share/lua/5.1/moses.lua:104: in function 'each'
./dataset/miniImagenet.lua:140: in function <./dataset/miniImagenet.lua:111>
./dataset/data-loader.lua:34: in function <./dataset/data-loader.lua:13>
./train/train.lua:26: in function <./train/train.lua:6>
train/run-train.lua:16: in main chunk
[C]: in function 'dofile'
...mldl/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk
[C]: at 0x00405d50
mldl@mldlUB1604:~/ub16_prj/meta-learning-lstm$

Attempt to serialize closure './dataset/miniImagenet.lua:31' with 1 upvalues

/home/i/torch/install/bin/luajit: /home/i/torch/install/share/lua/5.1/dataset/Reader.lua:52: Attempt to serialize closure './dataset/miniImagenet.lua:31' with 1 upvalues; first is: '_'. Consider using ipc.workqueue.writeup() instead.

What wrong?

How to recover exact images from imagenet?

Hi Sachin,

Great work!
I'm trying to recover the images for miniImagenet dataset you provided and can't seem to figure out.
I did notice that wordnet id is prefix of the fileids in the dataset. However, rest of the field doesn't seem to match the file names in imagenet.

I would really appreciate your help on this!

Thanks,
Tsendee