ImageNet training

This repository contains modification of pytorch example script¹ for ImageNet training.

Modified script is contained in the file main_stratified.py.

Modified script extracts an extra validation set from the training set before training and in each epoch it saves trained network outputs on training set and both validation sets.

Usage

Additional parameters of the modified script are:

• validation-size - size of subset of train data set aside for extra validation,
• output-folder - path to a folder in which training outputs will be stored,
• existing-val-split - path to a folder with files val_idx.npy and train_idx.npy specifying training/validation split of training set, None if there is no existing split,
• from-ptm - whether to load the model from library pretrainedmodels.

Output

Script produces following output into the output folder:

• train_idx.npy -indexes into the original train set used for training
• val_idx.npy -indexes into the original train set uset for validation 2
• val2_output_{epoch}.npy -output for validation set 2 in epoch
• val2_target.npy -targets for validation set 2
• val_output_{epoch}.npy -output for validation set in epoch
• val_target.npy -targets for validation set
• train_output_{epoch}.npy -output for training set in epoch
• train_target_{epoch}.npy -targets for training set in epoch
• valid_summary.txt - {epoch},{losses.avg:.4e},{top1.avg:.3f},{top5.avg:.3f},\t{losses.avg:.4e},{top1.avg:.3f},{top5.avg:.3f}\n - first three values correspond to training data, second three to validation data
• checkpoint.pth.tar -checkpoint rewritten every epoch
• model_best.pth.tar -checkpoint with best top1 validation accuracy

Footnotes

1. https://github.com/pytorch/examples/tree/master/imagenet ↩