mini-trainer 0.1.1

A mini Trainer for PyTorch ecosystem.

Mini Trainer for PyTorch

This is a mini Trainer for PyTorch ecosystem. Particularly suitable for research and experiments because of the following advantages:

• Fully transparent and retraceable training process
• Low code volume for easy debugging
• Meets the main requirements for model training and evaluating

Main features:

• Pipeline for model training and evaluating
• Checkpoint
• Earlystopping
• Logging based on json file

Installation

pip install mini-trainer

Quick Start

Below is two examples for starting using mini-trainer. First is classic image classification task and another is house price regression. Both of them are complete deep learning project, and you can learn how the basic usage and main APIs of this project.

MINST Classification

House Sale Price Prediction

Main Functions and APIs

Initialization: Trainer()

API	Type	Desc
model	nn.Module	A model object to train.
save_path	str	Path to save checkpoints/loss plot/log file/etc.
optimizer	torch.optim.optimizer	Optimizer class, default Adam
lr	float	Learning rate, default 1e-3
loss	callable	Loss function, default L1 loss
device	str	Device type, default "auto". ["auto", "cpu", "cuda"]
early_stopping	bool	Whether early stopping, default True
stop_patience	int	Stop patience, default 7
stop_mode	str	Stop mode. For example, if you use MSE to test you model, this argument should be "min" while this should be "max" for Accuracy. default "min"

Model Training: fit()

API	Type	Desc
train_dataloader	torch.utils.data.DataLoader	Training dataloader.
val_dataloader	torch.utils.data.DataLoader	Validation dataloader.
epochs	int	Number of epochs, default 50
prog_bar	bool	Whether display progress bar to monitor training process, default True

Predicting: predict()

API	Type	Desc
test_dataloader	torch.utils.data.DataLoader	Dataloader.

Result saving: log()

API	Type	Desc
log	dict	Anything you want to record to log file, saved as a dictionary. It is very useful for research experiments in which you can record experiment start time, version, key hyperparameters, etc.

Plot loss curve: plot_loss()

API	Type	Desc
save	bool	Whether to save loss plot.