• Docs >
  • Tutorial 13: Self-Supervised Contrastive Learning with SimCLR

Tutorial 13: Self-Supervised Contrastive Learning with SimCLR

  • Author: Phillip Lippe

  • License: CC BY-SA

  • Generated: 2022-05-03T02:43:21.313398

In this tutorial, we will take a closer look at self-supervised contrastive learning. Self-supervised learning, or also sometimes called unsupervised learning, describes the scenario where we have given input data, but no accompanying labels to train in a classical supervised way. However, this data still contains a lot of information from which we can learn: how are the images different from each other? What patterns are descriptive for certain images? Can we cluster the images? To get an insight into these questions, we will implement a popular, simple contrastive learning method, SimCLR, and apply it to the STL10 dataset. This notebook is part of a lecture series on Deep Learning at the University of Amsterdam. The full list of tutorials can be found at https://uvadlc-notebooks.rtfd.io.

Open in Open In Colab

Give us a ⭐ on Github | Check out the documentation | Join us on Slack


This notebook requires some packages besides pytorch-lightning.

! pip install --quiet "torch>=1.8" "torchmetrics>=0.7" "ipython[notebook]" "pytorch-lightning>=1.4" "torchvision" "setuptools==59.5.0" "seaborn" "matplotlib"

Methods for self-supervised learning try to learn as much as possible from the data alone, so it can quickly be finetuned for a specific classification task. The benefit of self-supervised learning is that a large dataset can often easily be obtained. For instance, if we want to train a vision model on semantic segmentation for autonomous driving, we can collect large amounts of data by simply installing a camera in a car, and driving through a city for an hour. In contrast, if we would want to do supervised learning, we would have to manually label all those images before training a model. This is extremely expensive, and would likely take a couple of months to manually label the same amount of data. Further, self-supervised learning can provide an alternative to transfer learning from models pretrained on ImageNet since we could pretrain a model on a specific dataset/situation, e.g. traffic scenarios for autonomous driving.

Within the last two years, a lot of new approaches have been proposed for self-supervised learning, in particular for images, that have resulted in great improvements over supervised models when few labels are available. The subfield that we will focus on in this tutorial is contrastive learning. Contrastive learning is motivated by the question mentioned above: how are images different from each other? Specifically, contrastive learning methods train a model to cluster an image and its slightly augmented version in latent space, while the distance to other images should be maximized. A very recent and simple method for this is SimCLR, which is visualized below (figure credit - Ting Chen et al. ).

simclr contrastive learning

The general setup is that we are given a dataset of images without any labels, and want to train a model on this data such that it can quickly adapt to any image recognition task afterward. During each training iteration, we sample a batch of images as usual. For each image, we create two versions by applying data augmentation techniques like cropping, Gaussian noise, blurring, etc. An example of such is shown on the left with the image of the dog. We will go into the details and effects of the chosen augmentation techniques later. On those images, we apply a CNN like ResNet and obtain as output a 1D feature vector on which we apply a small MLP. The output features of the two augmented images are then trained to be close to each other, while all other images in that batch should be as different as possible. This way, the model has to learn to recognize the content of the image that remains unchanged under the data augmentations, such as objects which we usually care about in supervised tasks.

We will now implement this framework ourselves and discuss further details along the way. Let’s first start with importing our standard libraries below:

import os
import urllib.request
from copy import deepcopy
from urllib.error import HTTPError

import matplotlib
import matplotlib.pyplot as plt
import pytorch_lightning as pl
import seaborn as sns
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.utils.data as data
import torchvision
from IPython.display import set_matplotlib_formats
from pytorch_lightning.callbacks import LearningRateMonitor, ModelCheckpoint
from torchvision import transforms
from torchvision.datasets import STL10
from tqdm.notebook import tqdm

%matplotlib inline
set_matplotlib_formats("svg", "pdf")  # For export
matplotlib.rcParams["lines.linewidth"] = 2.0

# Import tensorboard
%load_ext tensorboard

# Path to the folder where the datasets are/should be downloaded (e.g. CIFAR10)
DATASET_PATH = os.environ.get("PATH_DATASETS", "data/")
# Path to the folder where the pretrained models are saved
CHECKPOINT_PATH = os.environ.get("PATH_CHECKPOINT", "saved_models/ContrastiveLearning/")
# In this notebook, we use data loaders with heavier computational processing. It is recommended to use as many
# workers as possible in a data loader, which corresponds to the number of CPU cores
NUM_WORKERS = os.cpu_count()

# Setting the seed

# Ensure that all operations are deterministic on GPU (if used) for reproducibility
torch.backends.cudnn.determinstic = True
torch.backends.cudnn.benchmark = False

device = torch.device("cuda:0") if torch.cuda.is_available() else torch.device("cpu")
print("Device:", device)
print("Number of workers:", NUM_WORKERS)
/usr/lib/python3.8/site-packages/apex/pyprof/__init__.py:5: FutureWarning: pyprof will be removed by the end of June, 2022
  warnings.warn("pyprof will be removed by the end of June, 2022", FutureWarning)
WARNING:root:Bagua cannot detect bundled NCCL library, Bagua will try to use system NCCL instead. If you encounter any error, please run `import bagua_core; bagua_core.install_deps()` or the `bagua_install_deps.py` script to install bundled libraries.
/tmp/ipykernel_4616/1782126012.py:24: DeprecationWarning: `set_matplotlib_formats` is deprecated since IPython 7.23, directly use `matplotlib_inline.backend_inline.set_matplotlib_formats()`
  set_matplotlib_formats("svg", "pdf")  # For export
Global seed set to 42
Device: cuda:0
Number of workers: 12

As in many tutorials before, we provide pre-trained models. Note that those models are slightly larger as normal (~100MB overall) since we use the default ResNet-18 architecture. If you are running this notebook locally, make sure to have sufficient disk space available.

# Github URL where saved models are stored for this tutorial
base_url = "https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial17/"
# Files to download
pretrained_files = [
pretrained_files += [f"LogisticRegression_{size}.ckpt" for size in [10, 20, 50, 100, 200, 500]]
# Create checkpoint path if it doesn't exist yet
os.makedirs(CHECKPOINT_PATH, exist_ok=True)

# For each file, check whether it already exists. If not, try downloading it.
for file_name in pretrained_files:
    file_path = os.path.join(CHECKPOINT_PATH, file_name)
    if "/" in file_name:
        os.makedirs(file_path.rsplit("/", 1)[0], exist_ok=True)
    if not os.path.isfile(file_path):
        file_url = base_url + file_name
        print(f"Downloading {file_url}...")
            urllib.request.urlretrieve(file_url, file_path)
        except HTTPError as e:
                "Something went wrong. Please try to download the file from the GDrive folder, or contact the author with the full output including the following error:\n",
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial17/SimCLR.ckpt...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial17/ResNet.ckpt...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial17/tensorboards/SimCLR/events.out.tfevents.SimCLR...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial17/tensorboards/classification/ResNet/events.out.tfevents.ResNet...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial17/LogisticRegression_10.ckpt...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial17/LogisticRegression_20.ckpt...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial17/LogisticRegression_50.ckpt...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial17/LogisticRegression_100.ckpt...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial17/LogisticRegression_200.ckpt...
Downloading https://raw.githubusercontent.com/phlippe/saved_models/main/tutorial17/LogisticRegression_500.ckpt...


We will start our exploration of contrastive learning by discussing the effect of different data augmentation techniques, and how we can implement an efficient data loader for such. Next, we implement SimCLR with PyTorch Lightning, and finally train it on a large, unlabeled dataset.

Data Augmentation for Contrastive Learning

To allow efficient training, we need to prepare the data loading such that we sample two different, random augmentations for each image in the batch. The easiest way to do this is by creating a transformation that, when being called, applies a set of data augmentations to an image twice. This is implemented in the class ContrastiveTransformations below:

class ContrastiveTransformations:
    def __init__(self, base_transforms, n_views=2):
        self.base_transforms = base_transforms
        self.n_views = n_views

    def __call__(self, x):
        return [self.base_transforms(x) for i in range(self.n_views)]

The contrastive learning framework can easily be extended to have more positive examples by sampling more than two augmentations of the same image. However, the most efficient training is usually obtained by using only two.

Next, we can look at the specific augmentations we want to apply. The choice of the data augmentation to use is the most crucial hyperparameter in SimCLR since it directly affects how the latent space is structured, and what patterns might be learned from the data. Let’s first take a look at some of the most popular data augmentations (figure credit - Ting Chen and Geoffrey Hinton):


All of them can be used, but it turns out that two augmentations stand out in their importance: crop-and-resize, and color distortion. Interestingly, however, they only lead to strong performance if they have been used together as discussed by Ting Chen et al.  in their SimCLR paper. When performing randomly cropping and resizing, we can distinguish between two situations: (a) cropped image A provides a local view of cropped image B, or (b) cropped images C and D show neighboring views of the same image (figure credit - Ting Chen and Geoffrey Hinton).


While situation (a) requires the model to learn some sort of scale invariance to make crops A and B similar in latent space, situation (b) is more challenging since the model needs to recognize an object beyond its limited view. However, without color distortion, there is a loophole that the model can exploit, namely that different crops of the same image usually look very similar in color space. Consider the picture of the dog above. Simply from the color of the fur and the green color tone of the background, you can reason that two patches belong to the same image without actually recognizing the dog in the picture. In this case, the model might end up focusing only on the color histograms of the images, and ignore other more generalizable features. If, however, we distort the colors in the two patches randomly and independently of each other, the model cannot rely on this simple feature anymore. Hence, by combining random cropping and color distortions, the model can only match two patches by learning generalizable representations.

Overall, for our experiments, we apply a set of 5 transformations following the original SimCLR setup: random horizontal flip, crop-and-resize, color distortion, random grayscale, and gaussian blur. In comparison to the original implementation, we reduce the effect of the color jitter slightly (0.5 instead of 0.8 for brightness, contrast, and saturation, and 0.1 instead of 0.2 for hue). In our experiments, this setting obtained better performance and was faster and more stable to train. If, for instance, the brightness scale highly varies in a dataset, the original settings can be more beneficial since the model can’t rely on this information anymore to distinguish between images.

contrast_transforms = transforms.Compose(
        transforms.RandomApply([transforms.ColorJitter(brightness=0.5, contrast=0.5, saturation=0.5, hue=0.1)], p=0.8),
        transforms.Normalize((0.5,), (0.5,)),

After discussing the data augmentation techniques, we can now focus on the dataset. In this tutorial, we will use the STL10 dataset, which, similarly to CIFAR10, contains images of 10 classes: airplane, bird, car, cat, deer, dog, horse, monkey, ship, truck. However, the images have a higher resolution, namely 96\times 96 pixels, and we are only provided with 500 labeled images per class. Additionally, we have a much larger set of 100,000 unlabeled images which are similar to the training images but are sampled from a wider range of animals and vehicles. This makes the dataset ideal to showcase the benefits that self-supervised learning offers.

Luckily, the STL10 dataset is provided through torchvision. Keep in mind, however, that since this dataset is relatively large and has a considerably higher resolution than CIFAR10, it requires more disk space (~3GB) and takes a bit of time to download. For our initial discussion of self-supervised learning and SimCLR, we will create two data loaders with our contrastive transformations above: the unlabeled_data will be used to train our model via contrastive learning, and train_data_contrast will be used as a validation set in contrastive learning.

unlabeled_data = STL10(
    transform=ContrastiveTransformations(contrast_transforms, n_views=2),
train_data_contrast = STL10(
    transform=ContrastiveTransformations(contrast_transforms, n_views=2),
Downloading http://ai.stanford.edu/~acoates/stl10/stl10_binary.tar.gz to /__w/1/s/.datasets/stl10_binary.tar.gz
Extracting /__w/1/s/.datasets/stl10_binary.tar.gz to /__w/1/s/.datasets
Files already downloaded and verified

Finally, before starting with our implementation of SimCLR, let’s look at some example image pairs sampled with our augmentations:

# Visualize some examples
imgs = torch.stack([img for idx in range(NUM_IMAGES) for img in unlabeled_data[idx][0]], dim=0)
img_grid = torchvision.utils.make_grid(imgs, nrow=6, normalize=True, pad_value=0.9)
img_grid = img_grid.permute(1, 2, 0)

plt.figure(figsize=(10, 5))
plt.title("Augmented image examples of the STL10 dataset")
Global seed set to 42

We see the wide variety of our data augmentation, including randomly cropping, grayscaling, gaussian blur, and color distortion. Thus, it remains a challenging task for the model to match two, independently augmented patches of the same image.

SimCLR implementation

Using the data loader pipeline above, we can now implement SimCLR. At each iteration, we get for every image x two differently augmented versions, which we refer to as \tilde{x}_i and \tilde{x}_j. Both of these images are encoded into a one-dimensional feature vector, between which we want to maximize similarity which minimizes it to all other images in the batch. The encoder network is split into two parts: a base encoder network f(\cdot), and a projection head g(\cdot). The base network is usually a deep CNN as we have seen in e.g. Tutorial 5 before, and is responsible for extracting a representation vector from the augmented data examples. In our experiments, we will use the common ResNet-18 architecture as f(\cdot), and refer to the output as f(\tilde{x}_i)=h_i. The projection head g(\cdot) maps the representation h into a space where we apply the contrastive loss, i.e., compare similarities between vectors. It is often chosen to be a small MLP with non-linearities, and for simplicity, we follow the original SimCLR paper setup by defining it as a two-layer MLP with ReLU activation in the hidden layer. Note that in the follow-up paper, SimCLRv2, the authors mention that larger/wider MLPs can boost the performance considerably. This is why we apply an MLP with four times larger hidden dimensions, but deeper MLPs showed to overfit on the given dataset. The general setup is visualized below (figure credit - Ting Chen et al. ):


After finishing the training with contrastive learning, we will remove the projection head g(\cdot), and use f(\cdot) as a pretrained feature extractor. The representations z that come out of the projection head g(\cdot) have been shown to perform worse than those of the base network f(\cdot) when finetuning the network for a new task. This is likely because the representations z are trained to become invariant to many features like the color that can be important for downstream tasks. Thus, g(\cdot) is only needed for the contrastive learning stage.

Now that the architecture is described, let’s take a closer look at how we train the model. As mentioned before, we want to maximize the similarity between the representations of the two augmented versions of the same image, i.e., z_i and z_j in the figure above, while minimizing it to all other examples in the batch. SimCLR thereby applies the InfoNCE loss, originally proposed by Aaron van den Oord et al.  for contrastive learning. In short, the InfoNCE loss compares the similarity of z_i and z_j to the similarity of z_i to any other representation in the batch by performing a softmax over the similarity values. The loss can be formally written as:

\ell_{i,j}=-\log \frac{\exp(\text{sim}(z_i,z_j)/\tau)}{\sum_{k=1}^{2N}\mathbb{1}_{[k\neq i]}\exp(\text{sim}(z_i,z_k)/\tau)}=-\text{sim}(z_i,z_j)/\tau+\log\left[\sum_{k=1}^{2N}\mathbb{1}_{[k\neq i]}\exp(\text{sim}(z_i,z_k)/\tau)\right]

The function \text{sim} is a similarity metric, and the hyperparameter \tau is called temperature determining how peaked the distribution is. Since many similarity metrics are bounded, the temperature parameter allows us to balance the influence of many dissimilar image patches versus one similar patch. The similarity metric that is used in SimCLR is cosine similarity, as defined below:

\text{sim}(z_i,z_j) = \frac{z_i^\top \cdot z_j}{||z_i||\cdot||z_j||}

The maximum cosine similarity possible is 1, while the minimum is -1. In general, we will see that the features of two different images will converge to a cosine similarity around zero since the minimum, -1, would require z_i and z_j to be in the exact opposite direction in all feature dimensions, which does not allow for great flexibility.

Finally, now that we have discussed all details, let’s implement SimCLR below as a PyTorch Lightning module:

class SimCLR(pl.LightningModule):
    def __init__(self, hidden_dim, lr, temperature, weight_decay, max_epochs=500):
        assert self.hparams.temperature > 0.0, "The temperature must be a positive float!"
        # Base model f(.)
        self.convnet = torchvision.models.resnet18(
            pretrained=False, num_classes=4 * hidden_dim
        )  # num_classes is the output size of the last linear layer
        # The MLP for g(.) consists of Linear->ReLU->Linear
        self.convnet.fc = nn.Sequential(
            self.convnet.fc,  # Linear(ResNet output, 4*hidden_dim)
            nn.Linear(4 * hidden_dim, hidden_dim),

    def configure_optimizers(self):
        optimizer = optim.AdamW(self.parameters(), lr=self.hparams.lr, weight_decay=self.hparams.weight_decay)
        lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(
            optimizer, T_max=self.hparams.max_epochs, eta_min=self.hparams.lr / 50
        return [optimizer], [lr_scheduler]

    def info_nce_loss(self, batch, mode="train"):
        imgs, _ = batch
        imgs = torch.cat(imgs, dim=0)

        # Encode all images
        feats = self.convnet(imgs)
        # Calculate cosine similarity
        cos_sim = F.cosine_similarity(feats[:, None, :], feats[None, :, :], dim=-1)
        # Mask out cosine similarity to itself
        self_mask = torch.eye(cos_sim.shape[0], dtype=torch.bool, device=cos_sim.device)
        cos_sim.masked_fill_(self_mask, -9e15)
        # Find positive example -> batch_size//2 away from the original example
        pos_mask = self_mask.roll(shifts=cos_sim.shape[0] // 2, dims=0)
        # InfoNCE loss
        cos_sim = cos_sim / self.hparams.temperature
        nll = -cos_sim[pos_mask] + torch.logsumexp(cos_sim, dim=-1)
        nll = nll.mean()

        # Logging loss
        self.log(mode + "_loss", nll)
        # Get ranking position of positive example
        comb_sim = torch.cat(
            [cos_sim[pos_mask][:, None], cos_sim.masked_fill(pos_mask, -9e15)],  # First position positive example
        sim_argsort = comb_sim.argsort(dim=-1, descending=True).argmin(dim=-1)
        # Logging ranking metrics
        self.log(mode + "_acc_top1", (sim_argsort == 0).float().mean())
        self.log(mode + "_acc_top5", (sim_argsort < 5).float().mean())
        self.log(mode + "_acc_mean_pos", 1 + sim_argsort.float().mean())

        return nll

    def training_step(self, batch, batch_idx):
        return self.info_nce_loss(batch, mode="train")

    def validation_step(self, batch, batch_idx):
        self.info_nce_loss(batch, mode="val")

Alternatively to performing the validation on the contrastive learning loss as well, we could also take a simple, small downstream task, and track the performance of the base network f(\cdot) on that. However, in this tutorial, we will restrict ourselves to the STL10 dataset where we use the task of image classification on STL10 as our test task.


Now that we have implemented SimCLR and the data loading pipeline, we are ready to train the model. We will use the same training function setup as usual. For saving the best model checkpoint, we track the metric val_acc_top5, which describes how often the correct image patch is within the top-5 most similar examples in the batch. This is usually less noisy than the top-1 metric, making it a better metric to choose the best model from.

def train_simclr(batch_size, max_epochs=500, **kwargs):
    trainer = pl.Trainer(
        default_root_dir=os.path.join(CHECKPOINT_PATH, "SimCLR"),
        gpus=1 if str(device) == "cuda:0" else 0,
            ModelCheckpoint(save_weights_only=True, mode="max", monitor="val_acc_top5"),
    trainer.logger._default_hp_metric = None  # Optional logging argument that we don't need

    # Check whether pretrained model exists. If yes, load it and skip training
    pretrained_filename = os.path.join(CHECKPOINT_PATH, "SimCLR.ckpt")
    if os.path.isfile(pretrained_filename):
        print(f"Found pretrained model at {pretrained_filename}, loading...")
        # Automatically loads the model with the saved hyperparameters
        model = SimCLR.load_from_checkpoint(pretrained_filename)
        train_loader = data.DataLoader(
        val_loader = data.DataLoader(
        pl.seed_everything(42)  # To be reproducable
        model = SimCLR(max_epochs=max_epochs, **kwargs)
        trainer.fit(model, train_loader, val_loader)
        # Load best checkpoint after training
        model = SimCLR.load_from_checkpoint(trainer.checkpoint_callback.best_model_path)

    return model

A common observation in contrastive learning is that the larger the batch size, the better the models perform. A larger batch size allows us to compare each image to more negative examples, leading to overall smoother loss gradients. However, in our case, we experienced that a batch size of 256 was sufficient to get good results.

simclr_model = train_simclr(
    batch_size=256, hidden_dim=128, lr=5e-4, temperature=0.07, weight_decay=1e-4, max_epochs=500
/home/AzDevOps_azpcontainer/.local/lib/python3.8/site-packages/pytorch_lightning/trainer/connectors/callback_connector.py:96: LightningDeprecationWarning: Setting `Trainer(progress_bar_refresh_rate=1)` is deprecated in v1.5 and will be removed in v1.7. Please pass `pytorch_lightning.callbacks.progress.TQDMProgressBar` with `refresh_rate` directly to the Trainer's `callbacks` argument instead. Or, to disable the progress bar pass `enable_progress_bar = False` to the Trainer.
GPU available: True, used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs
Found pretrained model at saved_models/ContrastiveLearning/SimCLR.ckpt, loading...

To get an intuition of how training with contrastive learning behaves, we can take a look at the TensorBoard below:

%tensorboard --logdir ../saved_models/tutorial17/tensorboards/SimCLR/