SOA/RPC>> Neural-Network-From-Scratch>> 返回
项目作者： ankitagupta820

项目描述 ：
Perceptron Neural Network Implementaion from scratch for Multiclass image Classification
高级语言： Python
项目地址: git://github.com/ankitagupta820/Neural-Network-From-Scratch.git
创建时间： 2021-04-07T23:42:10Z
项目社区：https://github.com/ankitagupta820/Neural-Network-From-Scratch
开源协议：MIT License
下载


Neural Network From Scratch

The Multiclass Perceptron Neural Network is developed from Scrach without using any Machine Learning libraries.

Data

We use MNIST dataset of Handwritten digit images normalized into 28x28 (784) bit vectors.

The training set contains 60,000 images. Test set contains 10,000 images.

Model Description

The Model is a Perceptron Neural Network with 1 Hidden Layer.

  • Input Layer - 784 Nodes
  • Hidden Layer - 256 Nodes with Sigmoid as the activation function.
  • Output Layer - 10 Nodes with Softmax as activation function.

Training

The model is fine trained and fine-tuned using Stochastic Gradient decent.
The loss is computed using Cross Entropy loss function.

  • Number of Epochs = 20
  • Learning Rate = 0.005

Results

The model achieves a training accuracy of 94.7% and a test accuracy of 93.4%.

image