My Torch Project

mainly for my own use - documentation not that great

Features

Training your model
Packaging your model
Training & interference through notebooks
A docker image is included

Usage

Use reinit.bash to reinit this scaffold to your needs
Install locally in development mode with pip3 install --user -e .
Dummy-test inference with python3 model.py

You can use make.bash to create a python package with the .pth neural net weights baked in.

After installing, using your trained model is as neat as this:

from my_torch_project.model import Detector1
detector = Detector1()
res = detector(img)

Files

your_project/
    model.py        : define your model here
                      define also a module wrapper for
                      weight loading and transformations
                      (say, for numerical values => labels, etc.)

    datamodel.py    : define data transformations, say:
                      tensor <=> image
                      define also data loaders (Dataset subclasses)

    eval.py         : define your evaluation metrics

    trainer.py      : define trainer (training scheme, losses, evaluation steps, etc.)


train/              : training aux dir

    train.py        : cli for running the trainer

    checkpoints/
                    : .pth weight files

    runs/
                    : tensorflow logs

data/
                    : any persistent data for distribution with this module
                      (for example, the final weights)

Extras

Like cool hardware & want to try external gpu for training? Please see egpu.md.

Copyright & License