• 📒 Table of Contents
• 📍 Overview
• ✍️ Self Supervised Learning
• 🔎 Supported Methods
• 📦 Installation
• 💡 Tutorial
• 📊 Benchmarks
• 📜 References Used
• 💯 License
• 🤝 Collaborators

Welcome to the Self-Supervised Learning Library! This repository hosts a collection of tools and implementations for self-supervised learning. Self-supervised learning is a powerful paradigm that leverages unlabeled data to pre-trained models, which can then be fine-tuned on specific tasks with smaller labeled datasets. This library aims to provide researchers and practitioners with a comprehensive set of tools to experiment, learn, and apply self-supervised learning techniques effectively. This project was our assignment during the summer apprenticeship and final project in the newly established Intelligent and Learning System (ILS) laboratory at the University of Isfahan.

Self-supervised learning is a subfield of machine learning where models are trained to predict certain aspects of the input data without relying on manual labeling. This approach has gained significant attention due to its ability to leverage large amounts of unlabeled data, which is often easier to obtain than fully annotated datasets. This library provides implementations of various self-supervised techniques, allowing you to experiment with and apply these methods in your own projects.

Barlow Twins is a self-supervised learning method that aims to learn embeddings invariant to distortions of the input sample. It achieves this by applying two distinct sets of augmentations to the same input sample, resulting in two distorted views of the same image. The objective function measures the cross-correlation matrix between the outputs of two identical networks fed with these distorted sample versions, striving to make it as close to the identity matrix as possible. This causes the embedding vectors of the distorted sample versions to become similar while minimizing redundancy among the components of these vectors. Barlow Twins particularly benefits from utilizing high-dimensional output vectors.

BYOL (Bootstrap Your Own Latent) is one of the new approaches to self-supervised learning. Like other methods, BYOL aims to learn a representation that can be utilized for downstream tasks. It employs two neural networks for learning: the online and target networks. The online network is trained to predict the target network's representation of the same image from a different augmented view. Simultaneously, the target network is updated with a slow-moving average of the online network's parameters. While state-of-the-art methods rely on negative pairs, BYOL achieves a new state of the art without them. It directly minimizes the similarity between the representations of the same image from different augmented views (positive pair).

DINO (self-distillation with no labels) is a self-supervised learning method that directly predicts the output of a teacher network—constructed with a momentum encoder—by utilizing a standard cross-entropy loss. It is an innovative self-supervised learning algorithm developed by Facebook AI. Through the utilization of self-supervised learning with Transformers, DINO paves the way for creating machines that can comprehend images and videos at a much deeper level.

MoCo, short for Momentum Contrast, is a self-supervised learning algorithm that employs a contrastive loss. MoCo v2 represents an enhanced iteration of the original Momentum Contrast self-supervised learning algorithm. Motivated by the findings outlined in the SimCLR paper, the authors introduced several modifications in MoCo v1, which included replacing the 1-layer fully connected layer with a 2-layer MLP head featuring ReLU activation for the unsupervised training stage. Additionally, they incorporated blur augmentation and adopted a cosine learning rate schedule. These adjustments enabled MoCo to outperform the state-of-the-art SimCLR, even when utilizing a smaller batch size and fewer epochs.

MoCo v3, introduced in the paper "An Empirical Study of Training Self-Supervised Vision Transformers," represents another advancement in self-supervised learning. It builds upon the foundation of MoCo v1 / MoCo v2 and addresses the instability issue observed when employing ViT for self-supervised learning.

In contrast to MoCo v2, MoCo v3 adopts a different approach where the keys naturally coexist within the same batch. The memory queue (memory bank) is discarded, resulting in a setting similar to that of SimCLR. The encoder fq comprises a backbone (e.g., ResNet, ViT), a projection head, and an additional prediction head.

SimCLR (Simple Framework for Contrastive Learning of Representations) is a self-supervised technique used to learn image representations. The fundamental building blocks of contrastive self-supervised methods, such as SimCLR, are image transformations. Each image is transformed into multiple new images through randomly applied augmentations. The goal of the self-supervised model is to identify images that originate from the same original source among a set of negative examples. SimCLR operates on the principle of maximizing the similarity between positive pairs of augmented images while minimizing the similarity with negative pairs. The training process can be summarized as follows: Data Augmentation - SimCLR employs robust data augmentation techniques to generate multiple augmented versions of each input image.

SimSiam is a self-supervised representation learning model that was proposed by Facebook AI Research (FAIR). It is a simple Siamese network designed to learn meaningful representations without requiring negative sample pairs, large batches, or momentum encoders.

SwAV, or Swapping Assignments Between Views, is a self-supervised learning approach that takes advantage of contrastive methods without requiring to compute pairwise comparisons. Specifically, it 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 as in contrastive learning. Simply put, SwAV uses a swapped prediction mechanism where we predict the cluster assignment of a view from the representation of another view.

CLIP (Contrastive Language-Image Pre-training) is a deep learning method that jointly trains images and text. This model consists of two neural networks: an image encoder and a text encoder. The image encoder takes an image as input and produces a vector of numbers as output, representing the content of the image. The text encoder takes a sentence as input and produces a vector of numbers as output, representing the meaning of the sentence. The goal of CLIP is to learn to make the output vectors of the image encoder and the text encoder similar for related images and sentences. For example, if an image of a cat and a sentence like "This is a cat" are given to CLIP, the model should produce similar output vectors for the image and the sentence. CLIP is trained using a technique called contrastive learning. In contrastive learning, the model is given a set of images and sentences and is asked to pair the related images and sentences together. The model learns by optimizing a loss function that measures the difference between the output vectors for related images and sentences.

ALBEF (ALign the image and text representations BEfore Fusing) is a deep learning method used to improve the accuracy of image and language processing tasks. This method is based on the idea of precisely aligning features extracted from different inputs before merging them into a single representation. ALBEF works by first extracting features from image and text inputs separately. This can be done using various deep learning models, such as convolutional neural networks for images or recurrent neural networks for text. Next, ALBEF aligns the extracted features in a shared space. This alignment ensures that features related to similar concepts are positioned in similar locations in the representation space. Various methods can be used for feature alignment, such as:

• Attention-based alignment: This method uses an attention mechanism to focus on relevant parts of each input and then aligns the corresponding features in a shared space.
• Metric-based alignment: This method uses a similarity metric to measure the similarity between features from different inputs and then aligns them in a shared space based on this similarity.

After aligning the features, ALBEF merges them into a single representation. This merging can be done in various ways, such as:

• Sum integration: This method sums the aligned features.
• Neural network-based integration: This method uses a neural network to learn how to merge the aligned features in a non-linear way.

Finally, ALBEF uses the merged single representation to make a prediction or decision. This prediction can be related to a classification, categorization, or question-answering task.

SLIP (Self-supervision meets Language-Image Pre-training) is an innovative method in machine learning within the realm of deep learning that offers a powerful combination of self-supervision and language-image pre-training. This approach allows models to simultaneously learn from large paired datasets of images and text. The main idea behind SLIP is to use self-supervision, utilizing the inherent signals in the data to create artificial labels instead of relying on expensive manual labeling. For example, the title of an image can serve as a label for that image. Additionally, with language-image pre-training, SLIP enables the model to understand the relationships between visual and linguistic concepts. This helps the model perform more complex tasks, such as generating image descriptions, answering questions about images, and even creating new images based on textual descriptions. Using the SLIP method offers numerous advantages, including efficient data utilization, improved performance on various tasks such as image classification, object detection, text generation, and image-to-text translation, as well as a deeper understanding of the connections between language and images. These features make SLIP highly useful and effective for many real-world applications.

VSE (Visual-Semantic Embedding) is a powerful method in machine learning used to establish a connection between visual and textual information. The main goal of VSE is to map images and text into a shared vector space such that images related to similar texts are close to each other. VSE helps models understand the complex relationship between visual and linguistic concepts, which is crucial for many real-world applications. VSE is utilized in a wide range of applications, including text-based image retrieval, image description generation, answering questions about images, and multi-modal recommendation systems. Additionally, VSE serves as the foundation for many advanced multi-modal models capable of processing and understanding various types of data, such as images, text, audio, and video. VSE works by first converting images and texts into dense vectors. For images, convolutional neural networks (CNNs) are used to extract visual features. For text, natural language models like BERT or GPT are used to generate word vectors. These vectors are then mapped into a shared vector space designed to bring related images and texts close together. A similarity function, such as cosine similarity or Euclidean distance, is used to measure the similarity between images and texts.

SimVLM (Simple Visual Language Model Pretraining with Weak Supervision) is a powerful pre-training method in machine learning that allows models to simultaneously learn from large paired datasets of images and text. This method is based on the idea of using weak supervision, which utilizes the inherent signals in the data to create artificial labels instead of requiring expensive manual labeling. As mentioned, the main idea behind SimVLM is the use of weak supervision. SimVLM uses weak signals such as image titles, product descriptions, or auto-generated captions produced by text generation models. These weak signals act as artificial labels for training the model. By using image-text pairs, SimVLM enables the model to understand the relationship between visual and linguistic concepts. This helps the model perform more complex tasks, such as generating image descriptions, answering questions about images, and even creating new images based on textual descriptions.

UNITER (UNiversal Image-TExt Representation Learning) is an advanced model in the field of deep learning designed for learning unified representations of images and text. This model is developed to gain a deeper understanding of the relationship between visual and linguistic concepts and to improve performance across a wide range of multimodal tasks. UNITER maps images and text into shared vector spaces using a dual-branch transformer. This space is designed so that related images and texts are close to each other. It also utilizes weak supervision, such as image titles, product descriptions, or auto-generated captions produced by text generation models. These weak signals act as artificial labels for training the model. UNITER uses a multi-head attention mechanism to simultaneously focus on different parts of the image and text, modeling the relationships between them. Compared to other multimodal pre-training methods, UNITER performs better on complex tasks due to its use of a dual-branch transformer and multi-head attention.

You can install AK_SSL and its dependencies from PyPI with:

pip install AK-SSL

We strongly recommend that you install AK_SSL in a dedicated virtualenv, to avoid conflicting with your system packages

Using AK_SSL, you have the flexibility to leverage the most recent self-supervised learning techniques seamlessly, harnessing the complete capabilities of PyTorch. You can explore diverse backbones, models, and optimizer while benefiting from a user-friendly framework that has been purposefully crafted for ease of use.

You can easily import Trainer module from AK_SSL library and start utilizing it right away.

from AK_SSL.vision import Trainer

Now, let's initialize the self-supervised trainer with our chosen method, backbone, dataset, and other configurations.

trainer = Trainer(
    method="barlowtwins",           # training method as string (BarlowTwins, BYOL, DINO, MoCov2, MoCov3, SimCLR, SimSiam, SwAV)
    backbone=backbone,              # backbone architecture as torch.Module
    feature_size=feature_size,      # size of the extracted features as integer
    image_size=32,                  # dataset image size as integer
    save_dir="./save_for_report/",  # directory to save training checkpoints and Tensorboard logs as string
    checkpoint_interval=50,         # interval (in epochs) for saving checkpoints as integer
    reload_checkpoint=False,        # reload a previously saved checkpoint as boolean
    verbose=True,                   # enable verbose output for training progress as a boolean
    **kwargs                        # other arguments 
)

Note: The use of **kwargs can differ between methods, depending on the specific method, loss function, transformation, and other factors. If you are utilizing any of the objectives listed below, you must provide their arguments during the initialization of the Trainer class.

• SimCLR Transformation

    color_jitter_strength     # a float to Set the strength of color
    use_blur                  # a boolean to specify whether to apply blur augmentation
    mean                      # a float to specify the mean values for each channel
    std                       # a float to specify the standard deviation values for each channel
  

• BarlowTwins

  • Method

      projection_dim          # an integer to specify dimensionality of the projection head
      hidden_dim              # an integer to specify dimensionality of the hidden layers in the neural network
      moving_average_decay    # a float to specify decay rate for moving averages during training
    

  • Loss

      lambda_param            # a float to controlling the balance between the main loss and the orthogonality loss
    

• DINO Method

  • Method

      projection_dim          # an integer to specify dimensionality of the projection head
      hidden_dim              # an integer to specify dimensionality of the hidden layers in the projection head neural network
      bottleneck_dim          # an integer to specify dimensionality of the bottleneck layer in the student network
      temp_student            # a float to specify temperature parameter for the student's logits
      temp_teacher            # a float to specify temperature parameter for the teacher's logits
      norm_last_layer         # a boolean to specify whether to normalize the last layer of the network
      momentum_teacher        # a float to control momentum coefficient for updating the teacher network
      num_crops               # an integer to determines the number of augmentations applied to each input image
      use_bn_in_head          # a boolean to spcecify whether to use batch normalization in the projection head
    

  • Loss

      center_momentum        # a float to control momentum coefficient for updating the center of cluster assignments
    

• MoCo v2

  • Method

      projection_dim          # an integer to specify dimensionality of the projection head
      K                       # an integer to specify number of negative samples per positive sample in the contrastive loss
      m                       # a float to control momentum coefficient for updating the moving-average encoder
    

  • Loss

      temperature             # a float to control the temperature for the contrastive loss function
    

• MoCo v3

  • Method

      projection_dim          # an integer to specify dimensionality of the projection head
      hidden_dim              # an integer to specify dimensionality of the hidden layers in the projection head neural network
      moving_average_decay    # a float to specify decay rate for moving averages during training
    

  • Loss

      temperature             # a float to control the temperature for the contrastive loss function
    

• SimCLR

  • Method

      projection_dim          # an integer to specify dimensionality of the projection head
      projection_num_layers   # an integer to specify the number of layers in the projection head (1: SimCLR v1, 2: SimCLR v2)
      projection_batch_norm   # a boolean to indicate whether to use batch normalization in the projection head
    

  • Loss

      temperature             # a float to control the temperature for the contrastive loss function
    

• SimSiam

  • Method

      projection_dim          # an integer to specify dimensionality of the projection head
    

  • Loss

      eps                     # a float to control the stability of the loss function
    

• SwAV

  • Method

      projection_dim          # an integer to specify dimensionality of the projection head
      hidden_dim              # an integer to specify dimensionality of the hidden layers in the projection head neural network
      epsilon                 # a float to control numerical stability in the algorithm
      sinkhorn_iterations     # an integer to specify the number of iterations in the Sinkhorn-Knopp algorithm
      num_prototypes          # an integer to specify the number of prototypes or clusters for contrastive learning
      queue_length            # an integer to specify rhe length of the queue for maintaining negative samples
      use_the_queue           # a boolean to indicate whether to use the queue for negative samples
      num_crops               # an integer to determines the number of augmentations applied to each input image
    

  • Loss

      temperature             # a float to control the temperature for the contrastive loss function
    

You can easily import Trainer module from AK_SSL library and start utilizing it right away.

from AK_SSL.multimodal import Trainer

Now, let's initialize the self-supervised trainer with our chosen method, backbone, dataset, and other configurations.

trainer = Trainer(
    method="clip",                  # training method as string (CLIP, ALBEF, SLIP, SimVLM, UNITER, VSE)
    image_encoder=img_encoder,      # vision model to extract image features as nn.Module
    text_encoder=txt_encoder,       # text model to extract text features as nn.Module
    mixed_precision_training=True,  # whether to use mixed precision training or not as boolean
    save_dir="./save_for_report/",  # directory to save training checkpoints and Tensorboard logs as string
    checkpoint_interval=50,         # interval (in epochs) for saving checkpoints as integer
    reload_checkpoint=False,        # reload a previously saved checkpoint as boolean
    verbose=True,                   # enable verbose output for training progress as a boolean
    **kwargs                        # other arguments 
)

Note: The use of **kwargs can differ between methods, depending on the specific method, loss function, transformation, and other factors. If you are utilizing any of the objectives listed below, you must provide their arguments during the initialization of the Trainer class.

• CLIP

    image_feature_dim         # Dimension of the image features as integer
    text_feature_dim          # Dimension of the text features as integer
    embed_dim                 # Dimension of the embeddings as integer
    init_tau                  # Initial value of tau as float
    init_b                    # Initial value of b as float
  

• ALBEF

    mlm_probability           # Masked language modeling probability as float
    embed_dim                 # Dimension of the embeddings as integer
    vision_width              # Vision encoder output width as integer
    temp                      # Temperature parameter as float
    queue_size                # Queue size as integer
    momentum                  # Momentum parameter as float
  

• SimVLM

    transformer_encoder       # Transformer encoder for vision and text embeddings as nn.Module
    transformer_decoder       # Transformer decoder for embeddings as nn.Module
    vocab_size                # Size of the vocabulary as integer
    feature_dim               # Dimension of the features as integer
    max_seq_len               # Maximum sequence length as integer
    max_trunc_txt_len         # Maximum truncated text length as integer
    prefix_txt_len            # Prefix text length as integer
    target_txt_len            # Target text length as integer
    pad_idx                   # Padding index as integer
    image_resolution          # Image resolution as integer
    patch_size                # Patch size as integer
    num_channels              # Number of channels as integer
  

• SLIP

    mlp_dim                   # Dimension of the MLP as integer
    vision_feature_dim        # Dimension of the vision features as integer
    transformer_feature_dim   # Dimension of the transformer features as integer
    embed_dim                 # Dimension of the embeddings as integer
  

• UNITER

    pooler                          # pooler as nn.Module
    encoder                         # transformer encoder as nn.Module
    num_answer                      # number of answer classes as integer
    hidden_size                     # hidden size as integer
    attention_probs_dropout_prob    # dropout rate as float
    initializer_range               # initializer range as float
  

• VSE

    margin                   # Margin for contrastive loss as float
  

Then, we'll train the self-supervised model using the specified parameters.

  trainer.train(
      dataset=train_dataset,          # training dataset as torch.utils.data.Dataset               
      batch_size=256,          # the number of training examples used in each iteration as integer
      start_epoch=1,           # the starting epoch for training as integer (if 'reload_checkpoint' parameter was True, start epoch equals to the latest checkpoint epoch)
      epochs=100,              # the total number of training epochs as integer
      optimizer="Adam",        # the optimization algorithm used for training as string (Adam, SGD, or AdamW)
      weight_decay=1e-6,       # a regularization term to prevent overfitting by penalizing large weights as float
      learning_rate=1e-3,      # the learning rate for the optimizer as float
)

Then, we'll train the self-supervised model using the specified parameters.

  trainer.train(
      dataset=train_dataset,           # the training data set as torch.utils.data.Dataset             
      batch_size=256,          # the number of training examples used in each iteration as integer
      start_epoch=1,           # the starting epoch for training as integer (if 'reload_checkpoint' parameter was True, start epoch equals to the latest checkpoint epoch)
      epochs=100,              # the total number of training epochs as integer
      optimizer="Adam",        # the optimization algorithm used for training as string (Adam, SGD, or AdamW)
      weight_decay=1e-6,       # a regularization term to prevent overfitting by penalizing large weights as float
      learning_rate=1e-3,      # the learning rate for the optimizer as float
)

This evaluation assesses how well the pre-trained model performs on a dataset, specifically for tasks related to linear evaluation.

trainer.evaluate(
    train_dataset=train_dataset,      # to specify the training dataset as torch.utils.data.Dataset
    test_dataset=test_dataset,        # to specify the testing dataset as torch.utils.data.Dataset
    eval_method="linear",             # the evaluation method to use as string (linear or finetune)
    top_k=1,                          # the number of top-k predictions to consider during evaluation as integer
    epochs=100,                       # the number of evaluation epochs as integer
    optimizer='Adam',                 # the optimization algorithm used during evaluation as string (Adam, SGD, or AdamW)
    weight_decay=1e-6,                # a regularization term applied during evaluation to prevent overfitting as float
    learning_rate=1e-3,               # the learning rate for the optimizer during evaluation as float
    batch_size=256,                   # the batch size used for evaluation in integer
    fine_tuning_data_proportion=1,    # the proportion of training data to use during evaluation as float in range of (0.0, 1]
)

In case you want to use the pre-trained network in your own downstream task, you need to define a downstream task model. This model should include the self-supervised model backbone as one of its components. Here's an example of how to define a simple downstream model class:

  class DownstreamNet(nn.Module):
      def __init__(self, backbone, **kwargs):
          super().__init__()
          self.backbone = backbone
  
          # You can define your downstream task model here
  
      def forward(self, x):
          x = self.backbone(x)
          # ...
  
  
  downstream_model = DownstreamNet(trainer.get_backbone())

To load a previous checkpoint into the network, you can do as below.

path = 'YOUR CHECKPOINT PATH'
trainer.load_checkpoint(path)

To save model backbone, you can do as below.

trainer.save_backbone()

That's it! You've successfully trained and evaluate a self-supervised model using the AK_SSL Python library. You can further customize and experiment with different self-supervised methods, backbones, and hyperparameters to suit your specific tasks. You can find the description of Trainer class and its function using help built in fuction in python.

We executed models and obtained results on the CIFAR10 dataset, with plans to expand our experimentation to other datasets. Please note that hyperparameters were not optimized for maximum accuracy.

In the development of this project, we have drawn inspiration and utilized code, libraries, and resources from various sources. We would like to acknowledge and express our gratitude to the following references and their respective authors:

• Lightly Library
• PYSSL Library
• SimCLR Implementation
• All original codes of supported methods

These references have played a crucial role in enhancing the functionality and quality of our project. We extend our thanks to the authors and contributors of these resources for their valuable work.

This project is licensed under the MIT License.

By:

• Kian Majlessi
• Audrina Ebrahimi

Thanks to Dr. Peyman Adibi and Dr. Hossein Karshenas, for their invaluable guidance and support throughout this project.