A simple and modular Tensorflow model development environment to handle sequence-to-sequence models.

Developing models to solve a problem for a data set at hand, requires lot of trial and error methods. Which includes and not limited to:

• Preparing the ground truth or data set for training and testing
  • Collecting the data from online or open data sources
  • Getting the data from in-house or client database
• Pre-processing the data set
  • Text cleaning
  • NLP processing
  • Meta feature extraction etc.,
• Data iterators, loading and looping the data examples for model while training and testing
  • In memory - All data is held in RAM and looped in batches on demand
  • Reading from the disk on demand in batches
  • Maitaining different feature sets (i.e number of features and its types) for the model
• Models
  • Maintaining different models for same set of features
  • Good visulaizing and debugging environment/tools
  • Start and pause the training at will
• Model Serving
  • Load a particular model from the pool of available models for a particular data set

Below two Git repos got our attention:

• https://github.com/guillaumegenthial/sequence_tagging
• https://github.com/Franck-Dernoncourt/NeuroNER

Both of the projects are excellent in their own way, however they lack few things like support for different dataset and models in a modular way, which plays a key role in a customer facing project(s). Where nature of the data changes as the project evolves.

• To come up with an software architecture to try different models on different dataset
• Which should take care of:
  • Pre-processing the data
  • Preparing the data iterators for training, validation and testing for set of features and their types
  • Use a model that aligns with the data iterator feature type
  • Train the model in an iterative manner, with fail safe
  • Use the trained model to predict on new data
• Keep the model core logic independent of the current architecture

A few object-oriented principles are used in the python scripts for ease of extensibility and maintenance.

What we solved using this code?

• Top level accuracies on open Conll dataset 2003
• Extract information from patent documents for form filling, from historical data entries from the Database records.

• Handling Dataset and Preprocessing
• Data iterators
  • Dataset may have one or more features like words, characters, positional information of words etc.,
  • Extract those and convert word/characters to numeric ids, pad them etc.,
  • Enforces number of features and their types, so that set of models can work on down the line
• Models should agree with data iterator features types and make use of the aviable features to train the data

• Tensorflow Estimators is used for training/evaluating/saving/restoring/predicting

Each experiment starts based on a dataset.

Let use CoNLL data set, since it is provided as part this repo

• conll_csv_experiments
  • config
    • config.ini # all one time config goes here
  • data
    • train.txt
    • test.txt
    • val.txt
  • preprocessed_data
    • train/
    • val/
    • test/
  • data_iterator_1
    • model_v0
      • config_1
      • config_2

• Models
  • Variable Length BiLSTM with CRF

The whole pacakage is tested on CoNLL data set for software integrity, and results are not tuned yet!

Check here for more details on how to rest it on CoNLL data set.

• Python 3.5
• tensorflow-gpu r1.4
• spaCy
• tqdm
• tmux
• overrides

#run following command for one time password verification
ssh-copy-id "[email protected]"

ssh [email protected]

# One time setup
tmux new -s your_name
export PATH=/home/rpx/anaconda3/bin:$PATH

### Note following environment is already setup, 
### no need to replicate unles you wanted different versions
conda create -n tensorflow-gpu python=3.5 anaconda
export LD_LIBRARY_PATH=/home/rpx/softwares/cudnn6/cuda/lib64:$LD_LIBRARY_PATH
source activate tensorflow-gpu
python --version

conda create -n tensorflow-gpu python=3.5 anaconda
source activate tensorflow-gpu

pip install tensorflow_gpu
pip install spacy
python -m spacy download en_core_web_md
pip install tqdm
pip install overrides

cd ~/experiments/
mkdir your_name
cd your_name

git clone https://gitlab.pramati.com/imaginea-labs/i-tagger

Day to day use

tmux a -t your_name

### run only if you previous tmux session was closed completly
source activate tensorflow-gpu
export PATH=/home/rpx/anaconda3/bin:$PATH
export LD_LIBRARY_PATH=/home/rpx/softwares/cudnn6/cuda/lib64:$LD_LIBRARY_PATH

• Walk Through of Tensorflow APIs

python src/commands/patent_dataset.py --mode=preprocess
python src/commands/patent_dataset.py --mode=train
python src/commands/patent_dataset.py --mode=retrain --model-dir=PATH TO Model
python src/commands/patent_dataset.py --mode=predict --model-dir=PATH TO Model --predict-dir=PATH to Prediction files

TODOs:

• Remove all default params
• Tune the model for CoNLL dataset
• Test code and Documentation
• Cleaning of the code
• More on LSTM basics/tutorials