Gerating Sentences from a Continuous Space

Tensorflow implementation of Generating Sentences from a Continuous Space.

Prerequisites

1. Python packages:
   • Python 3.4 or higher
   • Tensorflow r0.12
   • Numpy

Setting up the environment:

1. Clone this repository:

git clone https://github.com/Chung-I/Variational-Recurrent-Autoencoder-Tensorflow.git

2. Set up conda environment:

conda create -n vrae python=3.6
conda activate vrae

3. Install python package requirements:

pip install -r requirements.txt

Usage

Training:

python vrae.py  --model_dir models --do train --new True

Reconstruct:

python vrae.py --model_dir models --do reconstruct --new False --input input.txt --output output.txt

Sample (this script read only the first line of input.txt, generate num_pts samples, and write them into output.txt):

python vrae.py --model_dir models --do sample --new False --input input.txt --output output.txt

Interpolate (this script requires that input.txt consists of only two sentences; it generate num_pts interpolations between them, and write those interpolated sentences into output.txt)::

python vrae.py --model_dir models --do interpolate --new False --input input.txt --output output.txt

model_dir: The location of the config file config.json and the checkpoint file.

do: Accept 4 values: train, encode_decode, sample, or interpolate.

new: create models with fresh parameters if set to True; else read model parameters from checkpoints in model_dir.

config.json

Hyperparameters are not passed from command prompt like that in tensorflow/models/rnn/translate/translate.py. Instead, vrae.py reads hyperparameters from config.json in model_dir.

Below are hyperparameters in config.json:

• model:

  • size: embedding size, and encoder/decoder state size.
  • latent_dim: latent space size.
  • in_vocab_size: source vocabulary size.
  • out_vocab_size: target vocabulary size.
  • data_dir: path to the corpus.
  • num_layers: number of layers for encoder and decoder.
  • use_lstm: use lstm for encoder and decoder or not. Use BasicLSTMCell if set to True; else GRUCell is used.
  • buckets: A list of pairs of [input size, output size] for each bucket.
  • bidirectional: bidirectional_rnn is used if set to True.
  • probablistic: variance is set to zero if set to False.
  • orthogonal_initializer: orthogonal_initializer is used if set to True; else uniform_unit_scaling_initializer is used.
  • iaf: inverse autoregressive flow is used if set to True.
  • activation: activation for encoder-to-latent layer and latent-to-decoder layer.
    • elu: exponential linear unit.
    • prelu: parametric linear unit. (default)
    • None: linear.

• train:

  • batch_size
  • beam_size: beam size for decoding. Warning: beam search is still under implementation. NotImplementedError would be raised if beam_size is set to be greater than 1.
  • learning_rate: learning rate parameter passed into AdamOptimizer.
  • steps_per_checkpoint: save checkpoint every steps_per_checkpoint steps.
  • anneal: do KL cost annealing if set to True.
  • kl_rate_rise_factor: KL term weight is increasd by this much every steps_per_checkpoint steps.
  • max_train_data_size: Limit on the size of training data (0: no limit).
  • feed_previous: If True, only the first of decoder_inputs will be used (the "GO" symbol), and all other decoder inputs will be generated by: next = embedding_lookup(embedding, argmax(previous_output)). In effect, this implements a greedy decoder. It can also be used during training to emulate http://arxiv.org/abs/1506.03099. If False, decoder_inputs are used as given (the standard decoder case).
  • kl_min: the minimum information constraint. Should be a non-negative float (where 0 is no constraint).
  • max_gradient_norm: gradients will be clipped to maximally this norm.
  • word_dropout_keep_prob: probability of randomly replacing some fraction of the conditioned-on word tokens with the generic unknown word token UNK. when equal to 0, the decoder sees no input.

• reconstruct:

  • feed_previous
  • word_dropout_keep_prob

• sample:

  • feed_previous
  • word_dropout_keep_prob
  • num_pts: sample num_pts points.

• interpolate:

  • feed_previous
  • word_dropout_keep_prob
  • num_pts: sample num_pts points.

Data

Penn TreeBank corpus is included in the repo. We also provide a Chinese poem corpus, its preprocessed version (set {"model":{"data_dir": "<corpus_dir>"}} in <model_dir>/config.json to it), and its pretrained model (set model_dir to it), all of which can be found here.