xingxingzhang / dress Goto Github PK

Hi,
If I'm not mistaken, the computation of FKGL is off due to the behaviour of int division in python2 (result of division casted to int instead of a float).

This would probably change the results in the DRESS paper (and subsequent papers who used the script).

Shouldn't there be a module called 'utils/shortcut.lua' included as well? Thanks!

Could you please release .840B.300d.wiki.full.aner.t7 file? I download the pertained wiki word embedding online but it doesn't work.

I try to reproduce the result (mostly for BLEU) for encoder-decoder with attention (in the paper it said 88% in BLEU), but I tried the model you provide "model_0.001.256.2L.we.full.2l.ft0.t7" with both test file in wikilarge and wikismall. I see the best result I got is from Wikilarge (which is 59.80% lower than what you reported in the paper). I wonder if I use correct test file or correct checkpoint?

`th generate_pipeline.lua --modelPath ../model_0.001.256.2L.we.full.2l.ft0.t7 --dataPath ../../text_simplification_data/train/dress/wikismall/PWKP_108016.tag.80.aner.test --outPathRaw outraw2.txt --oriDataPath ../../text_simplification_data/train/dress/wikismall/PWKP_108016.tag.80.aner.ori.test --oriMapPath ../../text_simplification_data/train/dress/wikismall/PWKP_108016.tag.80.aner.map.t7

th generate_pipeline.lua --modelPath ../model_0.001.256.2L.we.full.2l.ft0.t7 --dataPath ../../text_simplification_data/train/dress/wikilarge/wiki.full.aner.test --outPathRaw outraw2.txt --oriDataPath ../../text_simplification_data/train/dress/wikilarge/wiki.full.aner.ori.test --oriMapPath ../../text_simplification_data/train/dress/wikilarge/wiki.full.aner.map.t7`

To be honest, for now, I tried your dataset with another encoder-decoder model (Transformer, no RNN inside), which supposed to be work better at least in memory part. And I have also concerns with your model configuration:
why turn off the ignore case, I think the uppercase or lowercase makes no difference for the words in this data set?

Also based on your decode result, I also see many of short sentences (which drop performance but maybe good for FK score), the reason is wikilarge contains noisy kwoodsen data (http://homepages.inf.ed.ac.uk/kwoodsen/wiki.html#wiki-data), the data teach model to generate short sentences.

Could you please add some open source license to the data set? e.g. Creative Commons.

Thanks!

Hi

Whichever .sh file I try to run doesn't run because there is no cudnn.sh file, how to get that ?

Based on Table 1 of your paper, I saw WikiSmall performance is much lower than WikiLarge. The table indicates three results are based on the different test set.

I know for WikiLarge, you use 8 references test set. But for WikiSmall, I wonder which test set you are using? Is it PWKP_108016.tag.80.aner.test.src/PWKP_108016.tag.80.aner.test.dst (contain 100 test sample) in WikiSmall folder?

in file train_atuo_encoder.sh

where are the files?

train=/afs/inf.ed.ac.uk/group/project/img2txt/encdec/dataset/wiki-full/norm_all/auto-encoder-data/train.txt
valid=/afs/inf.ed.ac.uk/group/project/img2txt/encdec/dataset/wiki-full/norm_all/auto-encoder-data/valid.txt
test=/afs/inf.ed.ac.uk/group/project/img2txt/encdec/dataset/wiki-full/norm_all/auto-encoder-data/test.txt

Hi,

I would like to clarify regarding the Newsela data setup:

1. Am I right that the originally released data in 2015 (1,130 articles) was used in the paper? (That is, the text file in “newsela_data_share-20150302” in the Newsela release)
2. Following the description in Section 5 by having the first 1,070 articles for training, the next 30 for development, and the next 30 for testing, followed by filtering out sentence pairs corresponding to alignment levels 0-1, 1-2, and 2-3 gave me numbers of sentence pairs that are different from the paper (94,944 training, 2,531 development, and 2,462 test sentences). How can I come up with 94,208 training sentence pairs, 1,129 development sentence pairs, and 1,076 test sentence pairs as stipulated in the paper?

Thank you.

Regards,
Christian

Hello,

Is there any Python/ Tensorflow implementation of this work?

Thanks.

Would it be possible to add a list of dependencies which need to be installed before running the code?

When I train an Encoder-Decoder Attention Model using "sh run_std.sh", I get the following error:

/home/qiang/torch/extra/cutorch/lib/THC/THCTensorIndex.cu:321: void indexSelectLargeIndex(TensorInfo<T, IndexType>, TensorInfo<T, IndexType>, TensorInfo<long, IndexType>, int, int, IndexType, IndexType, long) [with T = float, IndexType = unsigned int, DstDim = 2, SrcDim = 2, IdxDim = -2]: block: [56,0,0], thread: [63,0,0] Assertion srcIndex < srcSelectDimSize failed.
THCudaCheck FAIL file=/home/qiang/torch/extra/cutorch/lib/THC/generic/THCStorage.c line=32 error=59 : device-side assert triggered
/home/qiang/torch/install/bin/luajit: cuda runtime error (59) : device-side assert triggered at /home/qiang/torch/extra/cutorch/lib/THC/generic/THCStorage.c:32
stack traceback:
[C]: at 0x7fbc8f5b6050
[C]: in function '__index'
layers/EMaskedClassNLLCriterion.lua:18: in function 'forward'
nnets/EncDecAWE.lua:391: in function 'opfunc'
/home/qiang/torch/install/share/lua/5.1/optim/sgd.lua:44: in function 'optimMethod'
nnets/EncDecAWE.lua:468: in function 'trainBatch'
train.lua:40: in function 'train'
train.lua:162: in function 'main'
train.lua:269: in function 'main'
train.lua:272: in main chunk
[C]: in function 'dofile'
...iang/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk
[C]: at 0x00405e90
Lock freed

Usage instructions:

To obtain and lock an id: ./gpu_lock.py --id
The lock is automatically freed when the parent terminates

To get an id that won't be freed: ./gpu_lock.py --id-to-hog
You must manually free these ids: ./gpu_lock.py --free

More info: http://homepages.inf.ed.ac.uk/imurray2/code/gpu_monitoring/

embedFile should be a torch object , not only the glove.840B.300d.txt file

but, where is the file?

`require 'shortcut'

local WordEmbed = torch.class('WordEmbeddingFT')

-- input should be torch7 file
function WordEmbed:__init(embedFile)
self.embed, self.word2idx, self.idx2word = unpack( torch.load(embedFile) )
xprintln('load embedding done!')
self.lowerCase = true
for _, word in ipairs(self.idx2word) do
if word ~= word:lower() then
self.lowerCase = false
end
end
print('lower case: ')
print(self.lowerCase)
end`

for cuda 7.5 please point to https://developer.nvidia.com/cuda-75-downloads-archive

When trying to train an encoder decoder model with attention by calling the script experiments/wikilarge/encdeca/train.sh, I get the error message: "module 'shortcut' not found:No LuaRocks module found for shortcut", although this file does exists in utils/shortcut.lua. Do I need to compile anything in advance such that require('shortcut') and other modules will work?

For your evaluation, I thought you use evaluation code from Joshua.
They have both SARI and STAR, could I ask is there any special reason you use STAR rather than STAR?

Hi, I have some problems when using the STAR scripts.

I can get the correct STAR score with the DRESS output (27.37). However, when I replace the DRESS output with random sentence, it get higher STAR score (33.7).

Am I using the script correctly? Do you have similar results during the experiment?

Hello,

Could you provide details on the pre-processing applied to the dataset? For example, which tokenizer was used (with which options)? Thank you.

Hi,
I was able to find the predictions of the models in the test set, but I was wondering if you could make available the predictions of the models (or even just the best one) on the validation set.

Thanks!