Category: 01. Image classification

Introduction BigTransfer (also known as BiT) is a state-of-the-art transfer learning method for image classification. Transfer of pre-trained representations improves sample efficiency and simplifies hyperparameter tuning when training deep neural networks for vision. BiT revisit the paradigm of pre-training on large supervised datasets and fine-tuning the model on a target task. The importance of appropriately…

Setup Introduction In this notebook, we will utilize multi-backend Keras 3.0 to implement the GCViT: Global Context Vision Transformer paper, presented at ICML 2023 by A Hatamizadeh et al. The, we will fine-tune the model on the Flower dataset for image classification task, leveraging the official ImageNet pre-trained weights. A highlight of this notebook is its compatibility…

Introduction Vision Transformers (ViTs) have sparked a wave of research at the intersection of Transformers and Computer Vision (CV). ViTs can simultaneously model long- and short-range dependencies, thanks to the Multi-Head Self-Attention mechanism in the Transformer block. Many researchers believe that the success of ViTs are purely due to the attention layer, and they seldom think…

Introduction In the academic paper An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale, the authors mention that Vision Transformers (ViT) are data-hungry. Therefore, pretraining a ViT on a large-sized dataset like JFT300M and fine-tuning it on medium-sized datasets (like ImageNet) is the only way to beat state-of-the-art Convolutional Neural Network models. The…

This example implements Swin Transformer: Hierarchical Vision Transformer using Shifted Windows by Liu et al. for image classification, and demonstrates it on the CIFAR-100 dataset. Swin Transformer (Shifted Window Transformer) can serve as a general-purpose backbone for computer vision. Swin Transformer is a hierarchical Transformer whose representations are computed with shifted windows. The shifted window scheme brings greater efficiency by…

Introduction Semi-supervised learning Semi-supervised learning is a machine learning paradigm that deals with partially labeled datasets. When applying deep learning in the real world, one usually has to gather a large dataset to make it work well. However, while the cost of labeling scales linearly with the dataset size (labeling each example takes a constant time),…

Introduction The Reptile algorithm was developed by OpenAI to perform model-agnostic meta-learning. Specifically, this algorithm was designed to quickly learn to perform new tasks with minimal training (few-shot learning). The algorithm works by performing Stochastic Gradient Descent using the difference between weights trained on a mini-batch of never-seen-before data and the model weights prior to training over…

Introduction This example implements the Perceiver: General Perception with Iterative Attention model by Andrew Jaegle et al. for image classification, and demonstrates it on the CIFAR-100 dataset. The Perceiver model leverages an asymmetric attention mechanism to iteratively distill inputs into a tight latent bottleneck, allowing it to scale to handle very large inputs. In other words: let’s…

Introduction Convolution has been the basis of most modern neural networks for computer vision. A convolution kernel is spatial-agnostic and channel-specific. Because of this, it isn’t able to adapt to different visual patterns with respect to different spatial locations. Along with location-related problems, the receptive field of convolution creates challenges with regard to capturing long-range…

Introduction This example implements the EANet model for image classification, and demonstrates it on the CIFAR-100 dataset. EANet introduces a novel attention mechanism named external attention, based on two external, small, learnable, and shared memories, which can be implemented easily by simply using two cascaded linear layers and two normalization layers. It conveniently replaces self-attention as used in…