NNPlayground

Neural Network mucking about.

NN-framework

This is an attempt to create a framework for creating NNs for various purposes.

Getting started:

Create a virtual environment and activate it:

python3 -m venv venv
sounce venv/bin/activate

Install the requirements into the virtual environment.

pip install -r requirements.txt

Note: if new packages are added the requirements.txt can be updated with:

pip freeze > requirements.txt

How to train:

Note: see the sample_data.csv entry in [Output files](Output files:) for a python one-liner to create a random sample file.

• Train with 64 hidden units (default)

python train_model.py --data data/sample_data.csv

• Train with 128 hidden units

python train_model.py --data data/sample_data.csv --hidden-size 128

• Train with custom hyperparameters

python train_model.py \
  --data data/sample_data.csv \
  --hidden-size 128 \
  --num-layers 3 \
  --batch-size 64 \
  --learning-rate 0.0005

How to predict:

Note: see the test_samples.csv entry in [Output files](Output files:) for a python one-liner to create a test sample file from the training set.

• Predict a few samples from the training set:

python predict.py --data data/test_samples.csv

• Predict a single random input sample:

python test_predict_row.py

Input files:

• config.py Externalizes many of the hyperparameters for a Neural Network and various files that get generated.

• data/class_labels.txt is a list of the labels for the output classifications. The NN classifies into numeric indices which are translated to the labels specified in this file. EG index 0 = first label in the list, index 1 = second label, etc.

Output files:

• data/best_model.pth and data/best_model_config.json are both input and output files They are created when the model is trained and the model is used when predictions are made. the JSON file documents the hyperparameters used to create the model during training.

• data/confusion_matrix* various confusion matrix outputs. These give an idea of how well the model performs.

• data/sample_data.csv and test_sample.csv dummy random sample data created to test the framework.

  • the sample data is created with python -c "from dataset import generate_random_csv; generate_random_csv('data/sample_data.csv', 1500)"
  • test sample is created with python -c "import pandas as pd; df = pd.read_csv('data/sample_data.csv'); df.tail(5).to_csv('data/test_samples.csv', index=False)"

• data/classification_report.txt records a summary of how well the model performs it includes: the precision, recall, F1-score, and support.

• data/per_class_metrics.txt records the accuracy of each classification.

• data/top_misclassifications.txt records a deeper look at the top 10 misclassifications.

• data/best_model.onnx ONNX export of the best model

  • can be created with python -c "from utils import export_to_onnx; export_to_onnx('data/best_model')"

ONNX

C++ code to use the model created and exported to ONNX format.

tinkering

These are just playground files, tinkering with NNs in various ways.