Handwriting | Predict & Synthesis

Proposed implementation of Alex Graves paper

Results

Handwritten synthesis of a dactilographed text

text : Welcome to my repository

Attention windows

Random handwriting generation

How to use this repository

This repository is divided into 4 folders:

• data
• notebooks
• models
• utils

Notebooks folder

• Notebook to explore the data
• Notebook for training models
• Notebook to view the results

Models folder

• the dummy file gathers the functions able to generate the strokes (unconditional and conditional) from the models learned and stored in ../data/files
• the dataloader file contains the two classes allowing to manipulate easily the dataset with pytorch
• the networks file contains the two classes modeling the two networks for the first two first questions
• the trainers file contains the classes that allow to train the networks preceding
• the files folder contains the saved weights of networks and others

Utils folder

• contains the function plot_stroke in the _init_
• contains the Onehot Encoder class which allows to work with Onehot encoded sentences.

Data folder

• contains the raw data :
  • strokes-py3.npy contains 6000 sequences of points that correspond to handwritten sentences
  • sentences.txt contains the corresponding text sentences

Requirements

• numpy==1.17.4
• matplotlib==3.1.1
• torch==1.3.1