PyTorch implementation of SwAV https//arxiv.org/abs/2006.09882

Unsupervised Learning of Visual Features by Contrasting Cluster Assignments

This code provides a PyTorch implementation and pretrained models for SwAV (Swapping Assignments between Views), as described in the paper Unsupervised Learning of Visual Features by Contrasting Cluster Assignments.

SwAV is an efficient and simple method for pre-training convnets without using annotations. Similarly to contrastive approaches, SwAV learns representations by comparing transformations of an image, but unlike contrastive methods, it does not require to compute feature pairwise comparisons. It makes our framework more efficient since it does not require a large memory bank or an auxiliary momentum network. Specifically, our method simultaneously clusters the data while enforcing consistency between cluster assignments produced for different augmentations (or “views”) of the same image, instead of comparing features directly. Simply put, we use a “swapped” prediction mechanism where we predict the cluster assignment of a view from the representation of another view. Our method can be trained with large and small batches and can scale to unlimited amounts of data.

Model Zoo

We release our best ResNet-50 pre-trained with SwAV with the hope that other researchers might also benefit by replacing the ImageNet supervised network with SwAV backbone. To load the model, simply do:

import torch
model = torch.hub.load('facebookresearch/swav', 'resnet50')

We provide several baseline SwAV pre-trained models with ResNet-50 architecture in torchvision format. We also provide models pre-trained with DeepCluster-v2 and SeLa-v2 obtained by applying improvements from the self-supervised community to DeepCluster and SeLa (see details in the appendix of our paper).

Larger architectures

We provide SwAV models with ResNet-50 networks where we multiply the width by a factor ×2, ×4, and ×5.

Running times

We provide the running times for some of our runs:

Running SwAV unsupervised training

Requirements

• Python 3.6
• PyTorch install >= 1.4.0
• torchvision
• CUDA 10.1
• Apex with CUDA extension
• Other dependencies: opencv-python, scipy, pandas, numpy

Singlenode training

SwAV is very simple to implement and experiment with. Our implementation consists in a main_swav.py file from which are imported the dataset definition src/multicropdataset.py, the model architecture src/resnet50.py and some miscellaneous training utilities src/utils.py.

For example, to train SwAV baseline on a single node with 8 gpus for 400 epochs, run:

python -m torch.distributed.launch --nproc_per_node=8 main_swav.py \
--data_path /path/to/imagenet/train \
--epochs 400 \
--base_lr 0.6 \
--final_lr 0.0006 \
--warmup_epochs 0 \
--batch_size 32 \
--size_crops 224 96 \
--nmb_crops 2 6 \
--min_scale_crops 0.14 0.05 \
--max_scale_crops 1. 0.14 \
--use_fp16 true \
--freeze_prototypes_niters 5005 \
--queue_length 3840 \
--epoch_queue_starts 15

Multinode training

Distributed training is available via Slurm. We provide several SBATCH scripts to reproduce our SwAV models. For example, to train SwAV on 8 nodes and 64 GPUs with a batch size of 4096 for 800 epochs run:

sbatch ./scripts/swav_800ep_pretrain.sh

Note that you might need to remove the copyright header from the sbatch file to launch it.

Set up dist_url parameter: We refer the user to pytorch distributed documentation (env or file or tcp) for setting the distributed initialization method (parameter dist_url) correctly. In the provided sbatch files, we use the tcp init method (see * for example).

Evaluating models

Evaluate models: Linear classification on ImageNet

To train a supervised linear classifier on frozen features/weights on a single node with 8 gpus, run:

python -m torch.distributed.launch --nproc_per_node=8 eval_linear.py \
--data_path /path/to/imagenet \
--pretrained /path/to/checkpoints/swav_800ep_pretrain.pth.tar

The resulting linear classifier can be downloaded here.

Evaluate models: Semi-supervised learning on ImageNet

To reproduce our results and fine-tune a network with 1% or 10% of ImageNet labels on a single node with 8 gpus, run:

• 10% labels

python -m torch.distributed.launch --nproc_per_node=8 eval_semisup.py \
--data_path /path/to/imagenet \
--pretrained /path/to/checkpoints/swav_800ep_pretrain.pth.tar \
--labels_perc "10" \
--lr 0.01 \
--lr_last_layer 0.2

• 1% labels

python -m torch.distributed.launch --nproc_per_node=8 eval_semisup.py \
--data_path /path/to/imagenet \
--pretrained /path/to/checkpoints/swav_800ep_pretrain.pth.tar \
--labels_perc "1" \
--lr 0.02 \
--lr_last_layer 5

Evaluate models: Transferring to Detection with DETR

DETR is a recent object detection framework that reaches competitive performance with Faster R-CNN while being conceptually simpler and trainable end-to-end. We evaluate our SwAV ResNet-50 backbone on object detection on COCO dataset using DETR framework with full fine-tuning. Here are the instructions for reproducing our experiments:

1. Install detr and prepare COCO dataset following these instructions.

2. Apply the changes highlighted in this gist to detr backbone file in order to load SwAV backbone instead of ImageNet supervised weights.

3. Launch training from detr repository with run_with_submitit.py.

python run_with_submitit.py --batch_size 4 --nodes 2 --lr_backbone 5e-5

Common Issues

For help or issues using SwAV, please submit a GitHub issue.

The loss does not decrease and is stuck at ln(nmb_prototypes) (8.006 for 3000 prototypes).

It sometimes happens that the system collapses at the beginning and does not manage to converge. We have found the following empirical workarounds to improve convergence and avoid collapsing at the beginning:

• use a lower epsilon value (--epsilon 0.03 instead of the default 0.05)
• carefully tune the hyper-parameters
• freeze the prototypes during first iterations (freeze_prototypes_niters argument)
• switch to hard assignment
• remove batch-normalization layer from the projection head
• reduce the difficulty of the problem (less crops or softer data augmentation)

We now analyze the collapsing problem: it happens when all examples are mapped to the same unique representation. In other words, the convnet always has the same output regardless of its input, it is a constant function. All examples gets the same cluster assignment because they are identical, and the only valid assignment that satisfy the equipartition constraint in this case is the uniform assignment (1/K where K is the number of prototypes). In turn, this uniform assignment is trivial to predict since it is the same for all examples. Reducing epsilon parameter (see Eq(3) of our paper) encourages the assignments Q to be less uniform (more peaked), which strongly helps avoiding collapsing. However, using a too low value for epsilon leads to numerical instability. If the loss goes to NaN, you can try using --improve_numerical_stability true in main_swav.py.

Training gets unstable when using the queue.

The queue is composed of feature representations from the previous batches. These lines discard the oldest feature representations from the queue and save the newest one (i.e. from the current batch) through a round-robin mechanism. This way, the assignment problem is performed on more samples: without the queue we assign B examples to num_prototypes clusters where B is the total batch size while with the queue we assign (B + queue_length) examples to num_prototypes clusters. This is especially useful when working with small batches because it improves the precision of the assignment.

If you start using the queue too early or if you use a too large queue, this can considerably disturb training: this is because the queue members are too inconsistent. After introducing the queue the loss should be lower than what it was without the queue. On the following loss curve (30 first epochs of this script) we introduced the queue at epoch 15. We observe that it made the loss go more down.

If when introducing the queue, the loss goes up and does not decrease afterwards you should stop your training and change the queue parameters. We recommend (i) using a smaller queue, (ii) starting the queue later in training.

Slow training.

If you experiment slow running times, it might be because of the Gaussian blur. We recommend operating Gaussian blur with PIL library instead of open-cv (pass --use_pil_blur true as argument). Note that we keep --use_pil_blur false in the scripts because all our experiments were performed with open-cv but we strongly recommend using PIL library instead.

License

See the LICENSE file for more details.

See also

SwAV-TF: A TensorFlow re-implementation.

Citation

If you find this repository useful in your research, please cite:

@article{caron2020unsupervised,
  title={Unsupervised Learning of Visual Features by Contrasting Cluster Assignments},
  author={Caron, Mathilde and Misra, Ishan and Mairal, Julien and Goyal, Priya and Bojanowski, Piotr and Joulin, Armand},
  journal={arXiv preprint arXiv:2006.09882},
  year={2020}
}