TL;DR: Tensorflow2 offers roughly 3 code styles of model implementation: Sequential, Functional and Subclassing. This project utilizes first style to implement SVHN dataset image classifier, whose level of abstraction is highest among the three. Scroll down to environment section to try out Streamlit demo of implemented image classifiers.

In any frameworks, high-level API sacrifices flexibility for implementation and makes overall code black-box if components are not well understood. However, purpose of it is to reduce amount of code implementation and therefore controlling overall system complexity.

This can also be applied when implementing neural network program in Tensorflow, since this framework provides APIs with wide spectrum of abstraction. For example, when execution logic which contains fixed number of linear steps(i.e. not iterative or autoregressive), this can be simply implemented by Keras Sequential class. Also, APIs for monitoring and controlling training process are prepared within Keras callbacks API.

Although there are APIs with lower level of abstraction to implement exactly same components, this project implements image classifiers using high-level Keras APIs to demonstrate the convenience and simplicity of them. Classifier with MLP, CNN architecture will be trained independently, and each of them will be saved with set of weights which provided best validation accuracy. After training classifiers, they can be examined with test data using Streamlit demo in local environment.

This application is scripted and tested on M2 Apple Silicon machine.

conda env create
conda activate svhn

There are two ways to prepare objects required two run the demo script.

For sake of users who want to try out this demo, two different pretrained models and corresponding CSV files of logs generated during training process are saved in my S3 repository. By executing main.py script using remote command, those objects can be downloaded from that repository, and you are ready to go.

python main.py remote

Optionally, to train and save models in local machine, execute main.py script using local command. Either way, trained models and history files will be saved in directory specified in LOCAL_DIR variable of the main script.

python main.py local

To try out trained models, execute Streamlit demo as following.

streamlit run demo.py