Torch Runner

A minimal wrapper that removes some of the overhead code in training pytorch models

Note: If you are looking for something more extensive, checkout Pytorch Lightning. This is mostly designed for my personal use.

Requirements

• torch
• tqdm

Installation

pip install torch-runner

Features

• seed all variables
• text logger
• early stopping
• save hyperparameters
• weights & biases support

Example

Checkout the examples folder which contains a jupyter notebook to train a resnet50 using torch_runner.

import torch
import torch_runner as T


class myTrainer(T.TrainerModule):
    def calc_metric(self, preds, target):
        ## Calc metrics such as accuracy etc.
    
    def loss_fct(self, preds, target):
        ## Calc loss
    
    def train_one_step(self, batch, batch_id):
        ## Get batch data from dataloader and perform one update
    
    def valid_one_step(self, batch, batch_id):
        ## Perform validation step

config = T.TrainerConfig()
model = myModel()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)

train_dataloader = ## pytorch dataloader
val_dataloader = ## pytorch dataloader

Trainer = myTrainer(model, optimizer, config)
Trainer.fit(train_dataloader, val_dataloader, epochs=10)