Self-labelling via simultaneous clustering and representation learning
Checkout our blogpost for a quick non-technical overview and an interactive visualization of our clusters.
Self-Label
This code is the official implementation of the ICLR 2020 paper Self-labelling via simultaneous clustering and representation learning.
Abstract
Combining clustering and representation learning is one of the most promising approaches for unsupervised learning of deep neural networks. However, doing so naively leads to ill posed learning problems with degenerate solutions. In this paper, we propose a novel and principled learning formulation that addresses these issues. The method is obtained by maximizing the information between labels and input data indices. We show that this criterion extends standard crossentropy minimization to an optimal transport problem, which we solve efficiently for millions of input images and thousands of labels using a fast variant of the Sinkhorn-Knopp algorithm. The resulting method is able to self-label visual data so as to train highly competitive image representations without manual labels. Our method achieves state of the art representation learning performance for AlexNet and ResNet-50 on SVHN, CIFAR-10, CIFAR-100 and ImageNet.
Results at a glance
NMI(%) | aNMI(%) | ARI(%) | LP Acc (%) | |
---|---|---|---|---|
AlexNet 1k | 50.5 | 12.2 | 2.7 | 42.1 |
AlexNet 10k | 66.4 | 4.7 | 4.7 | 43.8 |
R50 10x3k | 54.2 | 34.4 | 7.2 | 61.5 |
With better augmentations (all single crop)
Label-Acc | NMI(%) | aNMI(%) | ARI(%) | LP Acc (%) | model_weights | |
---|---|---|---|---|---|---|
Aug++ R18 1k (new) | 26.9 | 62.7 | 36.4 | 12.5 | 53.3 | here |
Aug++ R50 1k (new) | 30.5 | 65.7 | 42.0 | 16.2 | 63.5 | here |
Aug++ R50 10x3k (new) | 38.1 | 75.7 | 52.8 | 27.6 | 68.8 | here |
(MoCo-v2 + k-means**, K=3k) | 71.4 | 39.6 | 15.8 | 71.1 |
- "Aug++" refers to the better augmentations used in SimCLR, taken from the MoCo-v2 repo, but I still only trained for 280 epochs, with three lr-drops as in CMC.
- There are still further improvements to be made with a MLP or training 800 epochs (I train 280), as done in SimCLR, MoCov2 and SwAV.
- **MoCo-v2 uses 800 epochs, MLP and cos-lr-schedule. On MoCo-v2 I run k-means (K=3000) on the avg-pooled features (after the MLP-head it's pretty much the same performance) to obtain NMI, aNMI and ARI numbers.
- Models above use standard torchvision ResNet backbones so loading is now super easy:
import torch, torchvision
model = torchvision.models.resnet50(pretrained=False, num_classes=3000)
ckpt = torch.load('resnet50-10x3k_pp.pth')
model.load_state_dict(ckpt['state_dict'])
pseudolabels = ckpt['L']
- note on improvement potential: by just using "aug+": I get LP-accuracy of 67.2% after 200 epochs. MoCo-v2 with "aug+" only has 63.4% after 200 epochs.
Clusters that were discovered by our method
Sorted
Random
The edge-colors encode the true imagenet classes (which are not used for training). You can view all clusters here.
Requirements
- Python >3.6
- PyTorch > 1.0
- CUDA
- Numpy, SciPy
- also, see requirements.txt
- (optional:) TensorboardX
Running our code
Run the self-supervised training of an AlexNet with the command
$./scripts/alexnet.sh
or train a ResNet-50 with
$./scripts/resnet.sh
Note: you need to specify your dataset directory (it expects a format just like ImageNet with "train" and "val" folders). You also need to give the code enough GPUs to allow for storage of activations on the GPU. Otherwise you need to use the CPU variant which is significantly slower.
Full documentation of the unsupervised training code main.py
:
usage: main.py [-h] [--epochs EPOCHS] [--batch-size BATCH_SIZE] [--lr LR]
[--lrdrop LRDROP] [--wd WD] [--dtype {f64,f32}] [--nopts NOPTS]
[--augs AUGS] [--paugs PAUGS] [--lamb LAMB] [--cpu]
[--arch ARCH] [--archspec {big,small}] [--ncl NCL] [--hc HC]
[--device DEVICE] [--modeldevice MODELDEVICE] [--exp EXP]
[--workers WORKERS] [--imagenet-path IMAGENET_PATH]
[--comment COMMENT] [--log-intv LOG_INTV] [--log-iter LOG_ITER]
PyTorch Implementation of Self-Label
optional arguments:
-h, --help show this help message and exit
--epochs EPOCHS number of epochs
--batch-size BATCH_SIZE
batch size (default: 256)
--lr LR initial learning rate (default: 0.05)
--lrdrop LRDROP multiply LR by 0.1 every (default: 150 epochs)
--wd WD weight decay pow (default: (-5)
--dtype {f64,f32} SK-algo dtype (default: f64)
--nopts NOPTS number of pseudo-opts (default: 100)
--augs AUGS augmentation level (default: 3)
--paugs PAUGS for pseudoopt: augmentation level (default: 3)
--lamb LAMB for pseudoopt: lambda (default:25)
--cpu use CPU variant (slow) (default: off)
--arch ARCH alexnet or resnet (default: alexnet)
--archspec {big,small}
alexnet variant (default:big)
--ncl NCL number of clusters per head (default: 3000)
--hc HC number of heads (default: 1)
--device DEVICE GPU devices to use for storage and model
--modeldevice MODELDEVICE
GPU numbers on which the CNN runs
--exp EXP path to experiment directory
--workers WORKERS number workers (default: 6)
--imagenet-path IMAGENET_PATH
path to folder that contains `train` and `val`
--comment COMMENT name for tensorboardX
--log-intv LOG_INTV save stuff every x epochs (default: 1)
--log-iter LOG_ITER log every x-th batch (default: 200)
Evaluation
Linear Evaluation
We provide the linear evaluation methods in this repo. Simply download the models via . ./scripts/download_models.sh
and then either run scripts/eval-alexnet.sh
or scripts/eval-resnet.sh
.
Pascal VOC
We follow the standard evaluation protocols for self-supervised visual representation learning.
- for Classification: we follow the PyTorch implementation of DeepCluster with frozen BatchNorm.
- for Segmentation: we follow the implmentation from the Colorization paper which uses the FCN repo. Note: requires the Caffe framework
- for Detection: we follow Krähenbühl et al.'s implementation based on the Faster RCNN. Note: requires the Caffe framework
Our extracted pseudolabels
As we show in the paper, the pseudolabels we generate from our training can be used to quickly train a neural network with regular cross-entropy. Moreover they seem to correctly group together similar images. Hence we provide the labels for everyone to use.
AlexNet
You can download the pseudolabels from our best (raw) AlexNet model with 10x3000 clusters here.
ResNet
You can download the pseudolabels from our best ResNet model with 10x3000 clusters here.
Trained models
You can also download our trained models by running
$./scripts/download_models.sh
Use them like this:
import torch
import models
d = torch.load('self-label_models/resnet-10x3k.pth')
m = models.resnet(num_classes = [3000]*10)
m.load_state_dict(d)
d = torch.load('self-label_models/alexnet-10x3k-wRot.pth')
m = models.alexnet(num_classes = [3000]*10)
m.load_state_dict(d)
Reference
If you use this code etc., please cite the following paper:
Yuki M. Asano, Christian Rupprecht and Andrea Vedaldi. "Self-labelling via simultaneous clustering and representation learning." Proc. ICLR (2020)
@inproceedings{asano2020self,
title={Self-labelling via simultaneous clustering and representation learning},
author={Asano, Yuki M. and Rupprecht, Christian and Vedaldi, Andrea},
booktitle={International Conference on Learning Representations (ICLR)},
year={2020},
}