Writing Type Training

Code for training a deep learning model to detect the writing type of document images.

Intro

The component can be used for classifying input files to 'machinewritten' and 'handwritten' document and documents containing both writing types.

The following input file formats are accepted:

• .jpg
• .png
• .tiff

Setup

Creating virtual environment using Anaconda

The component can be safely installed and used in its own virtual environment. For example an Anaconda environment 'writing_type' can be created with the command

Installing dependencies in LINUX (Tested with Ubuntu 20.04)

conda create -n writing_type python=3.7

Now the virtual environment can be activated by typing

conda activate writing_type

Install required dependencies/libraries by typing

pip install -r requirements.txt

The latter command must be executed in the folder where the requirements.txt file is located.

NB! If you are having problems with the above command, installing this library can help. conda install -c conda-forge poppler

Training

Before training the data should be organized into a train folder, where there are two folders: an empty folder, containing empty images, and a ok folder for non-empty images. An example of the organization is depicted below.

├──writing_type_training
      ├──models
      ├──data
      |   ├──train
      |       ├──combination
      |       ├──hand_written
      |       └──type_written
      ├──runs
      ├──train.py
      ├──utils.py
      ├──dataset.py
      ├──constants.py
      └──requirements.txt

An example of training command. More arguments can be found in the train.py file or by running python train.py -h.

python train.py --epochs 10 --run_name 'run1'

By default, the code uses a pretrained Imagenet model. If you want to train from scratch you can specify it in the training command as below.

python train.py --epochs 10 --run_name 'run1' --from_scratch True

If you want to use data augmentations (i.e. rotate, colorjitter, sharpness, blur, affine and erasing), you can specify it in the training command as below.

python train.py --epochs 10 --run_name 'run1' --own_transform True

If you want to use different learning rates for base model and classification head, you can specify it in the training command as below.

python train.py --epochs 10 --run_name 'run1' --double_lr True

If you want to use freeze the base model during training, you can specify it in the training command as below.

python train.py --epochs 10 --run_name 'run1' --freeze True

If you want to change learning rate, number of workers or batch size, you can specify it in the training command as below.

python train.py --epochs 10 --run_name 'run1' --lr 0.01 --num_workers 4 --batch 8

All of the examples above can be used in one command.

Visualizing training metrics with Tensorboard

You can visualize training metrics with Tensorboard by running a following command in the training folder.

tensorboard --logdir runs

After this you can in your preferred browser go to this link http://localhost:6006/. There you can see how your training is progressing.

Information about the saved the models

The models are saved into ./runs/models folder.

The code saves two models based on different metrics. The first model is saved based on a "fitness" score that is calculated 0.75 * balanced f1 score + 0.25 * balanced_accuracy. The second model is saved based on validation loss.