Awesome Open Source
Awesome Open Source

Character Embeddings Recurrent Neural Network Text Generation Models

Inspired by Andrej Karpathy's The Unreasonable Effectiveness of Recurrent Neural Networks.

This repository attempts to replicate the models, with slight modifications, in different python deep learning frameworks.

Frameworks

Default Model Specification

Layer Type Output Shape Param # Remarks
Embedding (64, 64, 32) 3136 vocab size: 98, embedding size: 32
Dropout (64, 64, 32) 0 dropout rate: 0.0
LSTM (64, 64, 128) 82432 output size: 128
Dropout (64, 64, 128) 0 dropout rate: 0.0
LSTM (64, 64, 128) 131584 output size: 128
Dropout (64, 64, 128) 0 dropout rate: 0.0
Dense (64, 64, 98) 12642 output size: 98

Training Specification

  • Batch size: 64
  • Sequence length: 64
  • Number of epochs: 32
  • Learning rate: 0.001
  • Max gradient norm: 5.0

Setup

# clone repo
git clone [email protected]:yxtay/char-rnn-text-generation.git && cd char-rnn-text-generation

# create conda environment
conda env create -f=environment.yml

# activate environment
source activate dl

Usage

Training

usage: <framework>_model.py train [-h] --checkpoint-path CHECKPOINT_PATH 
                                  --text-path TEXT_PATH
                                  [--restore [RESTORE]]
                                  [--seq-len SEQ_LEN]
                                  [--embedding-size EMBEDDING_SIZE]
                                  [--rnn-size RNN_SIZE] 
                                  [--num-layers NUM_LAYERS]
                                  [--drop-rate DROP_RATE]
                                  [--learning-rate LEARNING_RATE]
                                  [--clip-norm CLIP_NORM] 
                                  [--batch-size BATCH_SIZE]
                                  [--num-epochs NUM_EPOCHS]
                                  [--log-path LOG_PATH]

optional arguments:
  -h, --help            show this help message and exit
  --checkpoint-path CHECKPOINT_PATH
                        path to save or load model checkpoints
  --text-path TEXT_PATH
                        path of text file for training
  --restore [RESTORE]   whether to restore from checkpoint_path or from
                        another path if specified
  --seq-len SEQ_LEN     sequence length of inputs and outputs (default: 64)
  --embedding-size EMBEDDING_SIZE
                        character embedding size (default: 32)
  --rnn-size RNN_SIZE   size of rnn cell (default: 128)
  --num-layers NUM_LAYERS
                        number of rnn layers (default: 2)
  --drop-rate DROP_RATE
                        dropout rate for rnn layers (default: 0.0)
  --learning-rate LEARNING_RATE
                        learning rate (default: 0.001)
  --clip-norm CLIP_NORM
                        max norm to clip gradient (default: 5.0)
  --batch-size BATCH_SIZE
                        training batch size (default: 64)
  --num-epochs NUM_EPOCHS
                        number of epochs for training (default: 32)
  --log-path LOG_PATH   path of log file (default: main.log)

Example:

python tf_model.py train \
    --checkpoint=checkpoints/tf_tinyshakespeare/model.ckpt \
    --text=data/tinyshakespeare.txt

Sample logs:

Text Generation

usage: <framework>_model.py generate [-h] --checkpoint-path CHECKPOINT_PATH
                                     (--text-path TEXT_PATH | --seed SEED)
                                     [--length LENGTH] [--top-n TOP_N]
                                     [--log-path LOG_PATH]

optional arguments:
  -h, --help            show this help message and exit
  --checkpoint-path CHECKPOINT_PATH
                        path to load model checkpoints
  --text-path TEXT_PATH
                        path of text file to generate seed
  --seed SEED           seed character sequence
  --length LENGTH       length of character sequence to generate (default:
                        1024)
  --top-n TOP_N         number of top choices to sample (default: 3)
  --log-path LOG_PATH   path of log file (default: main.log)

Example:

python tf_model.py generate \
    --checkpoint=checkpoints/tf_tinyshakespeare/model.ckpt \
    --seed="KING RICHARD"

Sample output:

KING RICHARDIIIIl II I tell thee,
As I have no mark of his confection,
The people so see my son.

SEBASTIAN:
I have men's man in the common to his sounds,
And so she said of my soul, and to him,
And too marry his sun their commanded
As thou shalt be alone too means
As he should to thy sensess so far to mark of
these foul trust them fringer whom, there would he had
As the word of merrous and subject.

GLOUCESTER:
A spack, a service the counsel son and here.
What is a misin the wind and to the will
And shall not streaks of this show into all heard.

KING EDIN YORK:
I will be suppet on himself tears as the sends.

KING EDWARD IV:
No looks and them, and while, a will, when this way.

BAPTHIO:
A mortain and me to the callant our souls
And the changed and such of the son.

CORIOLANUS:
I will, so show me with the child to the could sheep
To beseence, and shall so so should but hear
Than him with her fair to be that soul,
Whishe it is no meach of my lard and
And this, and with my love and the senter'd with marked
And her should

Benchmarks

Below are training duration and loss on tinyshakespeare.txt.

Framework Duration (s) Loss
Keras 5270 1.42505
TensorFlow 3003 1.45795
PyTorch 5868 1.32285
Chainer 4954 1.22930
MXNet 7348 1.34199
Alternatives To Char Rnn Text Generation
Select To Compare


Alternative Project Comparisons
Related Awesome Lists
Top Programming Languages
Top Projects

Get A Weekly Email With Trending Projects For These Topics
No Spam. Unsubscribe easily at any time.
Python (797,356
Deep Learning (36,026
Tensorflow (22,288
Pytorch (20,778
Neural (16,761
Neural Network (15,419
Rate (13,974
Keras (10,425
Generation (8,392
Recurrent Neural Networks (4,397
Embeddings (3,552
Mxnet (1,244
Chainer (763
Text Generation (602
Long Short Term Memory (48
Character Embeddings (10