wangcongcong123/mts

A pytorch MLP implementation for playing around with sklearn-like CSV datasets

MTS (MLP-Torch-Sklearn): Pytorch MLP implementation for Sklearn-like datasets classification and regression.

Features

• Regression/classification using sklearn-like (numeric csv) datasets
• Logging, model loading and saving, hyper-parameter tuning, easy model configuration.
• Well-deigned for pytorch-preferred users who just stepped to the world of deep learning (DL) and want to understand important DL concepts with some toy examples.

Notebooks

• demo - an example of MLP using iris dataset
• quickstart - read-to-go train and eval

Demo

The output will be like:

*******************eval on test set of iris*******************
         0.0       1.00      1.00      1.00         6
         1.0       1.00      1.00      1.00         3
         2.0       1.00      1.00      1.00         6

    accuracy                           1.00        15
   macro avg       1.00      1.00      1.00        15
weighted avg       1.00      1.00      1.00        15

** will add this to a Colab notebook

Quickstart

• Install

git clone https://github.com/wangcongcong123/mts.git
cd mts
pip install -r requirements.txt

• Obtain and split a data set. This downloads datasets from scikit-learn lib and convert them to csv format that this repo requires.

python obtain_split_data.py --dataset_name iris --test_size 0.1 
python obtain_split_data.py --dataset_name covtype --test_size 0.1 
python obtain_split_data.py --dataset_name digits --test_size 0.1 
python obtain_split_data.py --dataset_name boston --test_size 0.1

** try python obtain_split_data.py --help for the details of each parameter.

• Train and eval

python run.py --dataset_name iris --task cls --train_epochs 100 --train_batch_size 130 --device cpu --lr 0.03 --do_train --do_eval

python run.py --dataset_name covtype --task cls --train_epochs 120 --train_batch_size 128 --device cuda --lr 0.003 --do_train --do_eval

python run.py --dataset_name digits --task cls --train_epochs 100 --train_batch_size 128 --device cpu --lr 0.003 --do_train --do_eval

python run.py --dataset_name boston --task reg --train_epochs 100 --train_batch_size 128 --device cpu --lr 0.03 --do_train --do_eval

** try python run.py --help for the details of each parameter.

• Tracking the training process using tensorboard tensorboard dev upload --logdir runs.
  • here is an example of training using iris

Datasets available so far

• boston (regression)
• covtype (classification)
• digits (classification)
• iris (classification)

** There are many others, which encourages the community to incorporate them into this repository.