entrepreneur-interet-general / tf-han

Tensorflow Multilabel Hierarchical Attention Networks

About

This package implements the Hierarchical Attention Networks from Yang et al. 2016.

Main modifications are multilabel output with sigmoid_cross_entropy_with_logits and the addition of dense layers before the classification layers. Reason for this is to give the classifier more abstraction power throught these layers instead of relying on one single neuron to make sense of the document representation output from the 2 layers of bi-directionnal RNNs.

Code style is Black.

Install

Main dependencies are python (3.6), tensorflow (1.7), pandas, sklearn and PyYAML.

$ git clone [e-mail]/vict0rsch/tf-han.git
$ cd tf-han
$ pip install -r requirements.txt

Usage

Outline

The idea here is to split the work between several calsses:

A model is the set of tensorflow operations that transform and input_tensor into the output_tensor (no shit)

A trainer trains (and validate) a model: it takes care of getting the data, computing the loss, optimizing the model, running the validation steps, saving the graph etc.

An experiment runs various trainers with various hyper parameters.

All these classes have an hp attribute which represents their hyperparameters.

Available classes

Configuring an experiment

All (hyper) parameters can be defined in a yaml configuration file. List of parameters and their values can be found there: parameters.md.

A default configuration file is provided -> default_conf.yaml.

The idea is to have two separate sets of parameters:

• Fixed parameters like the source of data, the type of trainer or model or anything you want to keep constant across trainers during an experiment
• Randomizable parameters like the batch_size or layers or optimizer or learning_rate or anything you want to change between trainers

Randomizable parameters

They can be randomized in a few ways:

• From a range or from a list of values
• Sampled according a uniform, normal or deterministic (in the order of the range or list) distribution.

Data format

Running an Experiment

$ pwd
path/to/tf-han
$ ls
readme.md        requirements.txt test.py          text_classifier
$ ipython

>>> import text_classifier as tcl
>>> exp = tcl.Experiment(conf_path='path/to/conf.yaml')
>>> exp.run(10)

File structure

---

Credits