Generative Handwriting using LSTM Mixture Density Network with TensorFlow

Last update: Nov 24, 2022

Related tags

Deep Learning write-rnn-tensorflow

Overview

Generative Handwriting Demo using TensorFlow

An attempt to implement the random handwriting generation portion of Alex Graves' paper.

See my blog post at blog.otoro.net for more information.

How to use

I tested the implementation on TensorFlow r0.11 and Pyton 3. I also used the following libraries to help:

svgwrite
IPython.display.SVG
IPython.display.display
xml.etree.ElementTree
argparse
pickle

Training

You will need permission from these wonderful people people to get the IAM On-Line Handwriting data. Unzip lineStrokes-all.tar.gz into the data subdirectory, so that you end up with data/lineStrokes/a01, data/lineStrokes/a02, etc. Afterwards, running python train.py will start the training process.

A number of flags can be set for training if you wish to experiment with the parameters. The default values are in train.py

--rnn_size RNN_SIZE             size of RNN hidden state
--num_layers NUM_LAYERS         number of layers in the RNN
--model MODEL                   rnn, gru, or lstm
--batch_size BATCH_SIZE         minibatch size
--seq_length SEQ_LENGTH         RNN sequence length
--num_epochs NUM_EPOCHS         number of epochs
--save_every SAVE_EVERY         save frequency
--grad_clip GRAD_CLIP           clip gradients at this value
--learning_rate LEARNING_RATE   learning rate
--decay_rate DECAY_RATE         decay rate for rmsprop
--num_mixture NUM_MIXTURE       number of gaussian mixtures
--data_scale DATA_SCALE         factor to scale raw data down by
--keep_prob KEEP_PROB           dropout keep probability

Generating a Handwriting Sample

I've included a pretrained model in /save so it should work out of the box. Running python sample.py --filename example_name --sample_length 1000 will generate 4 .svg files for each example, with 1000 points.

IPython interactive session.

If you wish to experiment with this code interactively, just run %run -i sample.py in an IPython console, and then the following code is an example on how to generate samples and show them inside IPython.

[strokes, params] = model.sample(sess, 800)
draw_strokes(strokes, factor=8, svg_filename = 'sample.normal.svg')
draw_strokes_random_color(strokes, factor=8, svg_filename = 'sample.color.svg')
draw_strokes_random_color(strokes, factor=8, per_stroke_mode = False, svg_filename = 'sample.multi_color.svg')
draw_strokes_eos_weighted(strokes, params, factor=8, svg_filename = 'sample.eos.svg')
draw_strokes_pdf(strokes, params, factor=8, svg_filename = 'sample.pdf.svg')

Have fun-

License

MIT

Generative Handwriting using LSTM Mixture Density Network with TensorFlow

Related tags

Overview

Generative Handwriting Demo using TensorFlow

How to use

Training

Generating a Handwriting Sample

IPython interactive session.

License

Owner

hardmaru

Monocular 3D pose estimation. OpenVINO. CPU inference or iGPU (OpenCL) inference.

High-Fidelity Pluralistic Image Completion with Transformers (ICCV 2021)

It is modified Tensorflow 2.x version of Mask R-CNN

Super Pix Adv - Offical implemention of Robust Superpixel-Guided Attentional Adversarial Attack (CVPR2020)

Repo for 2021 SDD assessment task 2, by Felix, Anna, and James.

PyTorch implementations of Top-N recommendation, collaborative filtering recommenders.

Learning recognition/segmentation models without end-to-end training. 40%-60% less GPU memory footprint. Same training time. Better performance.

Bottom-up attention model for image captioning and VQA, based on Faster R-CNN and Visual Genome

Code for WECHSEL: Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models.

Source code for Fixed-Point GAN for Cloud Detection

Pose Transformers: Human Motion Prediction with Non-Autoregressive Transformers

Anonymize BLM Protest Images

Revisiting Discriminator in GAN Compression: A Generator-discriminator Cooperative Compression Scheme (NeurIPS2021)

A python comtrade load library accelerated by go

An educational AI robot based on NVIDIA Jetson Nano.

SE3 Pose Interp - Interpolate camera pose or trajectory in SE3, pose interpolation, trajectory interpolation

This program generates a random 12 digit/character password (upper and lowercase) and stores it in a file along with your username and app/website.

Exploring Simple 3D Multi-Object Tracking for Autonomous Driving (ICCV 2021)

Save-restricted-v-3 - Save restricted content Bot For telegram

Simple renderer for use with MuJoCo (>=2.1.2) Python Bindings.