April 8, 2021

594 words 3 mins read

HendrikStrobelt/LSTMVis

HendrikStrobelt/LSTMVis

Visualization Toolbox for Long Short Term Memory networks (LSTMs)

repo name HendrikStrobelt/LSTMVis
repo link https://github.com/HendrikStrobelt/LSTMVis
homepage
language Jupyter Notebook
size (curr.) 5779 kB
stars (curr.) 1001
created 2016-06-03
license BSD 3-Clause “New” or “Revised” License

Visual Analysis for State Changes in RNNs

More information about LSTMVis, an introduction video, and the link to the live demo can be found at lstm.seas.harvard.edu

Also check out our new work on Sequence-to-Sequence models on github or the live demo at http://seq2seq-vis.io/

Changes in V2.1

  • update to Python 3.7++ (thanks to @nneophyt)

Changes in V2

  • new design and server-backend
  • discrete zooming for hidden-state track
  • added annotation tracks for meta-data and prediction
  • added training and extraction workflow for tensorflow
  • client is now ES6 and D3v4
  • some performance enhancements on client side
  • Added Keras tutorial here (thanks to Mohammadreza Ebrahimi)

Install

Please use python 3.7 or later to install LSTMVis.

Clone the repository:

git clone https://github.com/HendrikStrobelt/LSTMVis.git; cd LSTMVis

Install python (server-side) requirements using pip:

python -m venv  venv3
source venv3/bin/activate
pip install -r requirements.txt

Download & Unzip example dataset(s) into <LSTMVis>/data/05childbook:

Children Book - Gutenberg - 2.2 GB

Parens Dataset - 10k small - 0.03 GB

start server:

source venv3/bin/activate
python lstm_server.py -dir <datadir>

For the example dataset, use python lstm_server.py -dir data

open browser at http://localhost:8888 - eh voila !

Adding Your Own Data

If you want to train your own data first, please read the Training document. If you have your own data at hand, adding it to LSTMVis is very easy. You only need three files:

  • HDF5 file containing the state vectors for each time step (e.g. states.hdf5)
  • HDF5 file containing a word ID for each time step (e.g. train.hdf5)*
  • Dict file containing the mapping from word ID to word (e.g. train.dict)*

A schematic representation of the data:

Data Format

*If you don’t have these files yet, but a space-separated .txt file of your training data instead, check out our text conversion tool

Data Directory

LSTMVis parses all subdirectories of <datadir> for config files lstm.yml. A typical <datadir> might look like this:

<datadir>
├── paren  		        <--- project directory
│   ├── lstm.yml 		<--- config file
│   ├── states.hdf5 	        <--- states for each time step
│   ├── train.hdf5 		<--- word ID for each time step
│   └── train.dict 		<--- mapping word ID -> word
├── fun .. 

Config File

a simple example of an lstm.yml is:

name: children books  # project name
description: children book texts from the Gutenberg project # little description

files: # assign files to reference name
  states: states.hdf5 # HDF5 files have to end with .h5 or .hdf5 !!!
  train: train.hdf5 # word ids of training set
  words: train.dict # dict files have to end with .dict !!

word_sequence: # defines the word sequence
  file: train # HDF5 file
  path: word_ids # path to table in HDF5
  dict_file: words # dictionary to map IDs from HDF5 to words

states: # section to define which states of your model you want to look at
  file: states # HDF5 files containing the state for each position
  types: [
        {type: state, layer: 1, path: states1}, # type={state, output}, layer=[1..x], path = HDF5 path
        {type: state, layer: 2, path: states2},
        {type: output, layer: 2, path: output2}
  ]

Intrigued ? Here is more..

Check out our documents about:

Credits

LSTMVis is a collaborative project of Hendrik Strobelt, Sebastian Gehrmann, Bernd Huber, Hanspeter Pfister, and Alexander M. Rush at Harvard SEAS.

comments powered by Disqus