Convolution

• CNN Explainer
• 1x1卷积核
• 空洞卷积(Dilated/Atrous Convolution)
• 转置卷积、空洞卷积和可变形卷积
• 分组卷积(Group conv)与深度可分离卷积(Depthwise separable conv)
• 20种卷积
• Depthwise Separable Convolution原理
• [论文]Large Kernel Matters
• large_kernel_matters/scripts/model.py
• ICCV2019-3*3卷积+1*3卷积+3*1卷积
• ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks
• CVPR2021-Diverse branch block
• ICLR 2022-ODConv
• 动态卷积Dynamic Convolution

Attention

• CBAM: Convolutional Block Attention Module
• Triplet注意力机制
• SENet
• SENet
• A2-Nets: Double Attention Networks
• Non-local-Block以及其他注意力机制
• LVT | ViT轻量化

Basic

• 深度学习中的超参数调节（learning rate、epochs、batch-size...） - 知乎
• 【AutoML】优化方法可以进行自动搜索学习吗？ - 知乎
• 一文搞懂激活函数(Sigmoid/ReLU/LeakyReLU/PReLU/ELU) - 知乎
• 损失函数（Loss Function） - 知乎
• 深度学习中的标准化—Normalization Methods in Deep Learning
• pytorch优化器详解
• 激活函数glu
• Group Normalization - 知乎

DiffusionModel

• Diffusion Model详解：直观理解、数学原理、PyTorch 实现
• 通俗理解扩散模型
• DDIM凭什么加速DDPM的采样效率
• DDIM反转和音频扩散

INN

• 神经网络的可逆形式 - 知乎
• 可逆神经网络(Invertible Neural Networks, INN)详细解析

MTL

• 深度学习的多个loss如何平衡?
• 多任务权重自动学习论文介绍和代码实现
• Mikoto10032/AutomaticWeightedLoss: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, Auxiliary Tasks in Multi-task Learning
• AutomaticLossWeightingPyTorch/weighted_loss.py at master · AidasLiaudanskas/AutomaticLossWeightingPyTorch
• 多任务学习中各loss权重应该如何设计
• 【多任务】任务损失/梯度优化策略合集
• abcdump/Pytorch-PCGrad-GradVac-AMP-GradAccum: PyTorch 1.11 reimplementation of multi task gradient adaptation ideas: Gradient Surgery (PCGrad) and Gradient Vaccine

Loss

• xxgege/GAM: The official repo for CVPR2023 highlight paper "Gradient Norm Aware Minimization Seeks First-Order Flatness and Improves Generalization".
• 不可导AP-Loss
• A General Adaptive Robust Loss Function - 知乎
• 自适应隐写算法HiLL - 知乎
• 边缘引导loss在显著性检测中的应用
• TV Loss

Frequency

• 论文汇总|如何从频域视角提升图像生成质量 (Frequency GANs)
• 频域(DCT,小波变换)与CNN结合
• FcaNet/model/layer.py at master · dcdcvgroup/FcaNet
• CNN

OpticFlow

• CVPR 2022 Oral | 基于全局匹配的光流学习框架
• GMA/core at main · zacjiang/GMA
• mulns/AccFlow: Official code for paper "AccFlow: Backward Accumulation for Long-Range Optical Flow"
• 2023 optical flow整理
• open-mmlab/mmflow: OpenMMLab optical flow toolbox and benchmark
• RAFT算法 ECCV2020最佳论文
• RAFT: Recurrent All-Pairs Field Transforms for Optical Flow_CyrilSterling
• JiaRenChang/PSMNet: Pyramid Stereo Matching Network (CVPR2018)
• ECCV 2022-FlowFormer 光流估计新范式

UpSampling

• 反卷积(Transposed Convolution)详细推导
• 语义分割--Understand Convolution for Semantic Segmentation_understanding convolution for semantic segmentatio
• CVPR2019-DUpsampling数据相关型解码方式
• DUpsampling/models/dunet.py at master · haochange/DUpsampling

Meta Learning

• Meta-SGD原理讲解和代码实现
• 三张图理解「优化优化器的优化器」加上REINFORCE算法迭代求解 - 知乎
• Learning to learn by gradient descent by gradient descent | Papers With Code
• Meta Learning/Learning to learn - 知乎
• learnables/learn2learn: A PyTorch Library for Meta-learning Research
• odlgroup/odl: Operator Discretization Library https://odlgroup.github.io/odl/
• Meta-RL之Learning to Learn by gd by gd
• Learning to learn by gradient descent by gradient descent - PyTorch实践
• Meta-Learning之Meta-SGD

Others

• Self-Attention与CNN融合范式
• Self-Attention ConvLSTM for Spatiotemporal Prediction
• 时空序列预测