Practical implementation of an astoundingly simple method for self-supervised learning that achieves a new state of the art (surpassing SimCLR) without contrastive learning and having to designate negative pairs.
This repository offers a module that one can easily wrap any image-based neural network (residual network, discriminator, policy network) to immediately start benefitting from unlabelled image data.
Update 1: There is now new evidence that batch normalization is key to making this technique work well
Update 2: A new paper has successfully replaced batch norm with group norm + weight standardization, refuting that batch statistics are needed for BYOL to work
Update 3: Finally, we have some analysis for why this works
Yannic Kilcher's excellent explanation
Now go save your organization from having to pay for labels :)
$ pip install byol-pytorchSimply plugin your neural network, specifying (1) the image dimensions as well as (2) the name (or index) of the hidden layer, whose output is used as the latent representation used for self-supervised training.
import torch
from byol_pytorch import BYOL
from torchvision import models
resnet = models.resnet50(pretrained=True)
learner = BYOL(
resnet,
image_size = 256,
hidden_layer = 'avgpool'
)
opt = torch.optim.Adam(learner.parameters(), lr=3e-4)
def sample_unlabelled_images():
return torch.randn(20, 3, 256, 256)
for _ in range(100):
images = sample_unlabelled_images()
loss = learner(images)
opt.zero_grad()
loss.backward()
opt.step()
learner.update_moving_average() # update moving average of target encoder
# save your improved network
torch.save(resnet.state_dict(), './improved-net.pt')That's pretty much it. After much training, the residual network should now perform better on its downstream supervised tasks.
A new paper from Kaiming He suggests that BYOL does not even need the target encoder to be an exponential moving average of the online encoder. I've decided to build in this option so that you can easily use that variant for training, simply by setting the use_momentum flag to False. You will no longer need to invoke update_moving_average if you go this route as shown in the example below.
import torch
from byol_pytorch import BYOL
from torchvision import models
resnet = models.resnet50(pretrained=True)
learner = BYOL(
resnet,
image_size = 256,
hidden_layer = 'avgpool',
use_momentum = False # turn off momentum in the target encoder
)
opt = torch.optim.Adam(learner.parameters(), lr=3e-4)
def sample_unlabelled_images():
return torch.randn(20, 3, 256, 256)
for _ in range(100):
images = sample_unlabelled_images()
loss = learner(images)
opt.zero_grad()
loss.backward()
opt.step()
# save your improved network
torch.save(resnet.state_dict(), './improved-net.pt')While the hyperparameters have already been set to what the paper has found optimal, you can change them with extra keyword arguments to the base wrapper class.
learner = BYOL(
resnet,
image_size = 256,
hidden_layer = 'avgpool',
projection_size = 256, # the projection size
projection_hidden_size = 4096, # the hidden dimension of the MLP for both the projection and prediction
moving_average_decay = 0.99 # the moving average decay factor for the target encoder, already set at what paper recommends
)By default, this library will use the augmentations from the SimCLR paper (which is also used in the BYOL paper). However, if you would like to specify your own augmentation pipeline, you can simply pass in your own custom augmentation function with the augment_fn keyword.
augment_fn = nn.Sequential(
kornia.augmentation.RandomHorizontalFlip()
)
learner = BYOL(
resnet,
image_size = 256,
hidden_layer = -2,
augment_fn = augment_fn
)In the paper, they seem to assure that one of the augmentations have a higher gaussian blur probability than the other. You can also adjust this to your heart's delight.
augment_fn = nn.Sequential(
kornia.augmentation.RandomHorizontalFlip()
)
augment_fn2 = nn.Sequential(
kornia.augmentation.RandomHorizontalFlip(),
kornia.filters.GaussianBlur2d((3, 3), (1.5, 1.5))
)
learner = BYOL(
resnet,
image_size = 256,
hidden_layer = -2,
augment_fn = augment_fn,
augment_fn2 = augment_fn2,
)To fetch the embeddings or the projections, you simply have to pass in a return_embeddings = True flag to the BYOL learner instance
import torch
from byol_pytorch import BYOL
from torchvision import models
resnet = models.resnet50(pretrained=True)
learner = BYOL(
resnet,
image_size = 256,
hidden_layer = 'avgpool'
)
imgs = torch.randn(2, 3, 256, 256)
projection, embedding = learner(imgs, return_embedding = True)The repository now offers distributed training with 🤗 Huggingface Accelerate. You just have to pass in your own Dataset into the imported BYOLTrainer
First setup the configuration for distributed training by invoking the accelerate CLI
$ accelerate configThen craft your training script as shown below, say in ./train.py
from torchvision import models
from byol_pytorch import (
BYOL,
BYOLTrainer,
MockDataset
)
resnet = models.resnet50(pretrained = True)
dataset = MockDataset(256, 10000)
trainer = BYOLTrainer(
resnet,
dataset = dataset,
image_size = 256,
hidden_layer = 'avgpool',
learning_rate = 3e-4,
num_train_steps = 100_000,
batch_size = 16,
checkpoint_every = 1000 # improved model will be saved periodically to ./checkpoints folder
)
trainer()Then use the accelerate CLI again to launch the script
$ accelerate launch ./train.pyIf your downstream task involves segmentation, please look at the following repository, which extends BYOL to 'pixel'-level learning.
https://github.com/lucidrains/pixel-level-contrastive-learning
@misc{grill2020bootstrap,
title = {Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning},
author = {Jean-Bastien Grill and Florian Strub and Florent Altché and Corentin Tallec and Pierre H. Richemond and Elena Buchatskaya and Carl Doersch and Bernardo Avila Pires and Zhaohan Daniel Guo and Mohammad Gheshlaghi Azar and Bilal Piot and Koray Kavukcuoglu and Rémi Munos and Michal Valko},
year = {2020},
eprint = {2006.07733},
archivePrefix = {arXiv},
primaryClass = {cs.LG}
}@misc{chen2020exploring,
title={Exploring Simple Siamese Representation Learning},
author={Xinlei Chen and Kaiming He},
year={2020},
eprint={2011.10566},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
