/character_recognizer_pytorch

Primary LanguagePython

Character Recognizer with PyTorch

This project aims to show how to use PyTorch to recognize handwritten characters.

Before we begin, we must make sure to setup!

  1. We first clone this repo
$ git clone https://github.com/character_recognizer_pytorch
$ cd character_recognizer_pytorch
  1. Create a new environment with conda and install requirements
$ conda create --name character_recognizer_pytorch python=3.8
$ conda activate character_recognizer_pytorch
$ pip install -r requirements.txt

Using PyTorch to Recognize Characters

We can run an experiment, for now, using two public datasets:

  • MNIST: MNIST stands for Mini-NIST, where NIST is the National Institute of Standards and Technology, which compiled a dataset of handwritten digits and letters in the 1980s. MNIST is Mini because it only included digits.
  • EMNIST: EMNIST is a repackaging of the original dataset, which also includes letters, but presented in the popularized MNIST format. You can see a publication about it here.

After installing the libraries, use python to run run_experiment.py

$ export PYTHONPATH=. # Defines root as python workdir
$ python training/run_experiment.py --max_epochs=5 --gpus=0 --num_workers=20 --model_class=MLP --data_class=MNIST

Where:

  • --max_epochs: Controls the number of complete passes through the training dataset.
  • --gpus: ID of GPU used to train the model. 0 stands for CPU and 1 for GPU or -1 to use all GPUs available.
  • --num_workers: How many subprocesses to use for data loading. 0 means that the data will be loaded in the main process.
  • --model_class: Class for modeling available in model_core.models. Contains the Neural Network Architecture.
  • --data_class: Class for downloading and preparing data available in model_core.data. Contains all steps to download and preprocess data. It can be MNIST or EMNIST.

P.S.: Project based on labs from Spring 2021 Full Stack Deep Learning Course