This repo provides the scripts to test a learned AlexNet's feature representation performance at the five different convolutional levels -- in parallel. The training lasts 36 epochs and should be finished in <1.5days.
$python eval_linear_probes.py
usage: eval_linear_probes.py [-h] [--data DATA] [--ckpt-dir DIR] [--device d]
[--modelpath MODELPATH] [--workers N]
[--epochs N] [--batch-size N]
[--learning-rate FLOAT] [--tencrops] [--evaluate]
[--img-size IMG_SIZE] [--crop-size CROP_SIZE]
[--imagenet-path IMAGENET_PATH]
AlexNet standard linear separability tests
optional arguments:
-h, --help show this help message and exit
--data DATA Dataset Imagenet or Places (default: Imagenet)
--ckpt-dir DIR path to checkpoints (default: ./test)
--device d GPU device
--modelpath MODELPATH
path to model
--workers N number of data loading workers (default: 6)
--epochs N number of epochs (default: 36)
--batch-size N batch size (default: 192)
--learning-rate FLOAT
initial learning rate (default: 0.01)
--tencrops flag to not use tencrops (default: on)
--evaluate flag to evaluate only (default: off)
--img-size IMG_SIZE imagesize (default: 256)
--crop-size CROP_SIZE
cropsize for CNN (default: 224)
--imagenet-path IMAGENET_PATH
path to imagenet folder, where train and val are
located
The settings follow the caffe code provided in Zhang et al., with optional tencrops enabled. Average pooling can be used, but max-pooling is faster and overall more common so it is used here.
If you use this code, please consider citing the following paper:
Yuki M. Asano, Christian Rupprecht and Andrea Vedaldi. "A critical analysis of self-supervision, or what we can learn from a single image." Proc. ICLR (2020)
@inproceedings{asano2020a,
title={A critical analysis of self-supervision, or what we can learn from a single image},
author={Asano, Yuki M. and Rupprecht, Christian and Vedaldi, Andrea},
booktitle={International Conference on Learning Representations (ICLR)},
year={2020},
}