/Diffusion-Handwriting-Generation

Primary LanguagePythonMIT LicenseMIT

Diffusion Models for Hand-writing generation

Tensorflow implementation for Diffusion Models for Hand-Writing Generation

1. Download dataset

First, download and extract the contents of the following files:

  • lineStrokes-all.tar.gz (Pen strokes in XML)
  • lineImages-all.tar.gz (Images for style encoding)
  • ascii-all.tar.gz (Sequences of text)

from this dataset URL and extract the contents to ./data

2. Training

To start training, log in to wandb.ai account on the CLI Then, run the train.py script with the following arguments :

python3 train.py --steps <number_of_epochs>
					--batchsize <batch_size>
					--seqlen <max_text_sequence_length>
					--textlen <max_text_length>
					--width <style_img_width>
					--warmup <lr_scheduler_warmup_step>
					--dropout <dropout_rate>
					--num_attlayers <number_attention_layers>
					--print_every <show_loss_every_m_steps>
					--save_every <ckpt_weights_every_n_steps>

3. Inference

To run sample inference, run the inference.py :

python3 inference.py --textstring <text_you_want>
						--writersource <style_image>
						--name <output_file_name>
						--weights <path_to_weights.h5>