NJODE

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #golang #android #deep-learning #windows #cli #kubernetes #vue #mcp

(★ 30)

code for "Neural Jump Ordinary Differential Equations"

NJODE 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Neural Jump Ordinary Differential Equations

[![DOI](https://zenodo.org/badge/269355048.svg)](https://zenodo.org/badge/latestdoi/269355048)

This repository is the official implementation of 
[Neural Jump Ordinary Differential Equations: Consistent Continuous-Time Prediction and Filtering](https://openreview.net/forum?id=JFKR3WqwyXR).

For a short summary of the paper see our video-presentation at the 
[ICLR conference](https://iclr.cc/virtual/2021/poster/3339).
Below we also provide the poster summarizing our work that we presented there.

![alt text](https://raw.githubusercontent.com/HerreraKrachTeichmann/NJODE/master/ICLR-conference-material/NJODE_poster.png)

## Requirements

This code was executed using Python 3.7.

To install requirements, download this Repo and cd into it. Then:

```setup
pip install -r requirements.txt
```

## Training & Testing

To train and test the model(s) of the paper, run these commands (other 
hyperparameters can be changed in the main section of demo.py):

go to the source directory:
```train
cd NJODE
```

for Black-Scholes model:
```train
python demo.py --dataset='BlackScholes'
```

for Heston model:
```train
python demo.py --dataset='Heston'
```

for Ornstein-Uhlenbeck model:
```train
python demo.py --dataset='OrnsteinUhlenbeck'
```

If no dataset for the model was generated yet, it will be generated 
automatically before the training starts.
ATTENTION: if you run the demo for a pretrained model first, a dataset with 
only 100 samples (instead of 20'000) will be generated for plotting. This should
be deleted before training a new model, such that a bigger dataset is generated
and used.

The model is trained and concurrently saved and tested after every <save_every>
epoch.
The plots that are generated are saved in "../data/saved_models/id-<model_id>/plots/"
and the training progress is printed.

## Pre-trained Models

Pre-trained models for each of the 3 stochastic models are distributed with the
code and saved in "../data/saved_models/id-x" for x=1,2,3.
These pre-trained models can be loaded and used to generate sample paths with 
the following commands:

go to the source directory:
```demo
cd controlled_ODE_RNN
```

- for Black-Scholes model:
```demo
python demo.py --model_id=1
```

- for Heston model:
```demo
python demo.py --model_id=2
```

- for Ornstein-Uhlenbeck model:
```demo
python demo.py --model_id=3
```

If no dataset for the model was generated yet, a small version of the dataset 
with 100 samples will be generated automatically, such that plots can be 
produced.
ATTENTION: this dataset should be replaced with a bigger one for training (the 
datasets are saved in "../data/training_data/" and can be deleted there).

The pretrained models are loaded and used for plotting. No training. The plots
are saved in "../data/saved_models/id-x/plots/" for x=1,2,3.

## Empirical convergence study
Models for the empirical convergence study were trained using parallel_train.py
with the model parameters specified below
```
# ==========================================================================
# parallel training for convergence analysis
# ==========================================================================
```
in the main part of the file parallel_train.py.