VitorBigelli / fashion-mnist

A convolutional neural network build with TensorFlow JS to classify images in relation to the famous Fashion Mnist

Fashion Mnist TensorFlow JS Classifier

This repository contains a Convolutional Neural Network build with TensorFlow JS that trains on the Fashion Mnist sprited dataset and allow user to draw images to predict, all this running on the Chrome Web Browser.

Setting up

In order to the model runs properly you need to download Google Chrome and install the Web Server for Chrome extension.

Running

1. Fork and clone this repository.

2. Open the Web Server for Chrome extension. You can do so by accessing the chrome://apps url in the Chrome search bar.

3. Click the "Open Folder" button, select and open the directory to where you cloned this repository.

4. Click the generated Web Server URL. It will display for you the files of the directory.

5. Click the index.html file and the application will automatically load and train.

6. When training is done you will be able to draw in the black canvas to make predictions.

Model

Architecture

_________________________________________________________________ 
Layer (type)                 Output shape              Param #   
================================================================= 
conv2d_Conv2D1 (Conv2D)      [null,26,26,16]           160       
_________________________________________________________________ 
max_pooling2d_MaxPooling2D1  [null,13,13,16]           0         
_________________________________________________________________ 
conv2d_Conv2D2 (Conv2D)      [null,11,11,64]           9280      
_________________________________________________________________ 
max_pooling2d_MaxPooling2D2  [null,5,5,64]             0         
_________________________________________________________________ 
flatten_Flatten1 (Flatten)   [null,1600]               0         
_________________________________________________________________ 
dense_Dense1 (Dense)         [null,128]                204928    
_________________________________________________________________ 
dense_Dense2 (Dense)         [null,10]                 1290      
================================================================= 
Total params: 215658 
Trainable params: 215658 
Non-trainable params: 0

Data:

• Dataset total size: 70000
• Dataset used size: 7000 (10%)
• Training data size: 6000
• Testing data size: 1000

Training

• Epochs: 20
• Batch size: 128
• Optimizer: Adam
• Loss function: Categorical Crossentropy .