View Code? Open in Web Editor NEW

Messing around with Databases Dimensionality Reduction and classification using Multi Layers Perceptron. (simple academic research)

Python 100.00%

hello-pca-mlp's Introduction

Hello PCA & MLP

Messing around with Databases Dimensionality Reduction and classification using Multi Layers Perceptron.

(simple academic research)

set up and run


git clone https://github.com/myreli/hello-pca-mlp.git

cd hello-pca-mlp

conda create -n pcamlp

conda activate pcamlp

conda install python=3.6.5

pip install -r requirements.txt

python script.py

requirements

Python > 3.6.x

Anaconda

about this sample

Using mnist dataset tutorial from tensorflow library

the MNIST database is a large database of handwritten digits, commonly used for ML.

Using matplotlib for graphic plotation

to graphically understand the MNIST dataset

Using PCA from sklearn.decomposition

to try it out dimensionality reduction

Using Multi Layer Perceptron from mlxtend.classifier

to understand dimensionality reduction, each reducted db has been exposed agains 3 MLPs, A (10 hidden layers and 0.005 learning curve), B (30 hidden layers and 0.05 learning curve) and C (60 hidden layers and 0.5 learning curve).

output

reading MNIST dataset from tensorflow sample...
Extracting MNIST_data/train-images-idx3-ubyte.gz
Extracting MNIST_data/train-labels-idx1-ubyte.gz
Extracting MNIST_data/t10k-images-idx3-ubyte.gz
Extracting MNIST_data/t10k-labels-idx1-ubyte.gz

ploting random MNIST sample to prove that dataset loaded properly and understand data...

applying PCA 1 to dataset...

applying PCA 2 to dataset...

apply MLPs and display accuracy to each...

------------------------------------------------

---
[RES] Full: 

[RES] MLP 1: 93.10%
[RES] MLP 2: 87.84%
[RES] MLP 3: 10.09%

---
[RES] PCA 1: 

[RES] MLP 1: 93.25%
[RES] MLP 2: 99.68%
[RES] MLP 3: 92.83%

---
[RES] PCA 2: 

[RES] MLP 1: 93.01%
[RES] MLP 2: 19.13%
[RES] MLP 3: 9.82%

------------------------------------------------

Execution time:  81.14863263649603

Recommend Projects