NOTE
There is a newer version of this toolkit, which addresses most of the shortcomings of this version. Please refer to the new tutorial & examples here

Introducing Pytorch Toolkit (PyTk)

A Keras-like API to train, evaluate and test Pytorch Models

Jump straight to the step-by-step tutorial

Pytorch, with it's Pythonic interface and dynamic graph evaluation makes it much easier to debug your deep learning models compared with Tensorflow 1.0. However, on the flip side, it's still a low level library, which requires us to (often) write the same (or almost same) boilerplate code to train, evaluate & test our model. Some ML authors love to code this way as they feel in control of what they create.

Me? I hate writing the same boilerplate code repeatedly. I'd rathen have my library handle the mundane tasks while I focus on the tasks that really matter viz. designing my model architecture, preparing data for my model & tuning the hyper-parameters to get the best performance. Keras does a remarkable job of providing a clean & simple API to train, evaluate & test models. With the Pytorch Toolkit I aim to bring the ease that Keras provides to Pytorch. Most of the API is similar to the Keras API, so Keras users should find it very easy to understand.

Features of the Pytorch Toolkit

• Keras-like API to train model and evaluate model performance (e.g. fit() and evaluate()) and make predictions ( e.g. predict())
• Support for torchvision datasets - with fit_dataset(), evaluate_dataset() and predict_module() calls
• Convenience class PytkModule from which to extend your model - this class provides the Keras-like API
• Full support for using nn.Sequential API, via the PytkModuleWrapper class, which also provides the same functions listed above
• Support for saving and loading model states with save() and load() methods
• Support for a variety of metrics (like accuracy, f1-score, precision, recall), calculated for each training epoch
• Keras-like progress display while model trains
• Support for Early Stopping of training

I must confess, I'm not good at coming up with snazzy names for the libraries I create. I tried several acronyms and finally settled on a rather unfancy name Pytorch Toolkit (or PyTk). If you can come up with a really cool name, please let me know!

This Github repository includes the tooklit, along with several examples on how to use it. Also included is a step-by-step tutorial which gradually introduces you to the complete API included in the toolkit. All functions & classes are included in just 1 Python file (ingeniously named pytorch_toolkit.py - I did warn you that I am not good at coming up with names, didn't I?). I have not yet created a module - maybe someday...

Installing Pytorch Toolkit

Since all functions & classes are included in just 1 file - pytorch_toolkit.py, there are strictly no special installation steps required to use this toolkit.

• Clone this repository - so you get entire Pytorch Tooklit and several example files

• Copy the pytorch_toolkit.py file into your project's directory and you are done! (Alternately, if you don't like several copies scattered across your disk drive, copy this file to any one folder in your Python SYSPATH)

• At the top of your code file (or Jupyter Notebook) and after all your other imports, enter the following code to import the toolkit into your project - I use the pytk alias - you can use whatever you prefer.

  import pytorch_tooklit as pytk

• This library depends on several other Python libraries, viz:

  • Numpy
  • pandas
  • pathlib
  • matplotlib
  • seaborn
  • scikit-learn
  • Pytorch (of course!)
  • torchsummary - if you want to see a Keras-like summary of your model (optional!)

I am assuming that you have these installed already - if you are an aspiring Data Scientist or ML enthusiast, you would have these (except perhaps Pytorch & torchsummary). Please refer to the respective module documentation on how to install these libraries.

Testing the installation

• I assume you have followed the instructions above & have installed the pre-requisites, including Pytorch itself.

• Start with a new Python code file (or Jupyter notebook) - add the following line at the top (after all your other imports):

  # ... your imports including Pytorch imports
  
  import pytorch_tooklit as pytk

• Just run this file/ code-cell - if you don't get any import errors, you are done - celebrate your success!! :)

  • Should you get any import errors, please correct them by installing the respective module/library (Pytorch Tooklit dependencies are mentioned above)
  • Repeat the above step (running file) until you get no errors

Features of the Pytorch Toolkit

This tookkit was inspired by Keras' clean API to train, evaluate and test models. Much of the functions provided mimick those from the Keras API. If you are already using Keras, you should notice the similarities immmediately.

The toolkit provides:

• A custom class - PytkModule from which your custom models should derive
• Keras-like API to train your models - fit(...), fit_dataset(...) functions and a show_plots(...) function to plot loss and accuracy metric across epochs, so you can see if your model is overfitting or underfitting.
• Several pre-defined metrics, like Accuracy, F1-Score, MSE, MAE etc., which can be tracked during training.
• Keras-like API to evaluate model's performance post training - evaluate(...) and avaluate_dataset(...) functions
• Functions to save & load model's state - pytk.load() and save(...) functions.
• Keras-like API to run predictions - predict(...) and predict_module(...) calls.
• Early stopping of training based on several criteria (e.g. validation loss not improving)

I am also including several fully working examples (as Python files or as Jupyter notebooks), where I have applied this API to solve several ML problems. The step-by-step tutorial will refer to one or more of these examples. I'll be adding more example to this Github repository, so please check back for changes.

If you are excited about starting with the Pytorch Toolkit, jump to the step-by-step tutorial right away!

Classification Examples:

• pyt_breast_cancer.py - binary classification on the Wisconsin Breast Cancer dataset using Pytorch ANN
• pyt_iris.py - multiclass classification of scikit-learn Iris dataset using Pytorch ANN
• pyt_wine.py - multiclass classification of scikit-learn Wine dataset
• pyt_mnist_dnn.py - MNIST digits multiclass classification with Pytorch ANN
• pyt_cifar10_cnn.py - multiclass classification with CNN on CIFAR-10 dataset
• Pytorch-Fruits360(Kaggle)_CNN.ipynb - iPython Notebook for the Kaggle Fruits360 multiclass classification problem
• Pytorch-Malaria Cell Detection(Kaggle)_CNN.ipynb - iPython Notebook for the Kaggle Malaric Cell Detection Dataset binary classification problem

Regression Examples:

• pyt_regression.py - univariate regression on synthesized data
• pyt_salary_regression.py - multivariate regression on salary data (@see csv_filed/salary_data.csv)

I will be adding mode examples over the course of time. Keep watching this repository :).

Hope you enjoy using the Pytorch Toolkit - my small contribution to the Pytorch community. Feedback is welcome.