caow13/brits

Code of NIPS18 Paper: BRITS: Bidirectional Recurrent Imputation for Time Series

README

It is a pytorch implemention of paper "BRITS: Bidirectional Recurrent Imputation for Time Series, Wei Cao, Dong Wang, Jian Li, Hao Zhou, Lei Li Yitan Li. (NerIPS 2018)". The paper can be found here. http://papers.nips.cc/paper/7911-brits-bidirectional-recurrent-imputation-for-time-series

To train the BRIST model, first please unzip the PhysioNet data into raw folder, including the label file Outcomes-a.txt.

To run the model:

• make a empty folder named json, and run inpute_process.py.
• run different models:
  • e.g., RITS_I: python main.py --model rits_i --epochs 1000 --batch_size 64 --impute_weight 0.3 --label_weight 1.0 --hid_size 108
  • for most cases, using impute_weight=0.3 and label_weight=1.0 lead to a good performance. Also adjust hid_size to control the number of parameters

brits's Issues

GRU-D implementation doesn't use input decay to mean

Hi @caow13,

Hope you're doing well.
When going through your code I've noticed that in gru_d.py there is no input decay to mean. In the paper by Che et al here, the decayed input is described as:

Where the second term contains the decay to last seen observation and the third term contains decay to empirical mean.

But in this implementation the code only does a decay to the last seen observation:

It could be that this was intended. Anyway, hope you can find the time to look into this.
Gr,
Noah

How to download the PhysioNet data?

How to download the PhysioNet data?

Why "eval_masks" isn't used in model forward steps?

Need Improvements on Readme file

Hi @caow13 @NIPS-BRITS @Dawn90 ,

The Readme file lacks clear instructions on what is the expected data format your program can accept and step-by-step process for both regression and classification with hyperlinking of the dataset. Along with that the parameters and meaning behind them in the main.py .

Human Activity Data Imputation Experiment

How do you process the human activity data? The description in the paper may not be clear enough to recur the results. Could you please open the data processing code?

Air Quality dataset.

How is the air quality data processed? Could you please provide the code?

BRIST -> BRITS

A typo exists on readme.md

I am getting "ValueError: a must be non-empty" error

forward deltas == backward deltas?

The deltas are identical in both directions

I do not know if this is intended. I do not see it justified in the paper. This means, that backward deltas do not apply to backward mask. I have noticed you have inverted masks twice for backward input processing:

1. first you invert masks when sending an argument to parse_rec function;
2. then you reverse masks again in parse_delta in if dir_ == 'backward'

So whatever bidirectional thing you calculate, the deltas do not align, so the temp decay is wrong, so the gammas... and so on.
Allow me to suggest creating tests using single dimensional data.

Update modules from Python 2 to Python 3

Hi @caow13 @Dawn90 @NIPS-BRITS ,

Codes are in Python 2 module should be bumped to Python 3.

应该把哪部分解压到raw中呢

全部解压进去之后会报错 没有读取数据

Typo in readme file

Readme incorrectly specifies inpute_process.py instead of input_process.py

Maximum recursion level reached on input_process

I am getting this error when I run input_process.py on physionet data, the error is happening on this line:
rec = json.dumps(rec)
do you have any solution for this?
Thanks