papereading

model train feature -> ReconNet_dropout -> drop.feature cat[ drop.feature, init.feature] cat at dim=1 ReconNet_dropout dropout ​ shortcut = self.conv1(x) ​ shortcut.features = self.bn0(shortcut.features) ​ shortcut.features = self.act1(shortcut.features) ​ shortcut.features = self.dropout1(shortcut.features) eval metric ???predict_labels = torch.mean(outputs, dim=0, keepdim=True) uncertainty_scores = torch.var(outputs, dim=0, keepdim=True) uncertainty_scores = torch.mean(uncertainty_scores, dim=1)

跳过train直接val 注释以下两个地方： 150到156行： 125到170行 Pointcept is a powerful and flexible codebase for point cloud perception research. It is also an official implementation of the following paper: Masked Scene Contrast: A Scalable Framework for Unsupervised 3D Representation Learning Xiaoyang Wu, Xin Wen, Xihui Liu, Hengshuang Zhao IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2023 [ Pretrain ] [ MSC ] - [ arXiv ] [ Bib ] → here...

Background & Method few-shot learning(doesn’t really matter in OOD task) Everything you need to know about Few-Shot Learning prototype clustering Apply in Stratified Transformer dmlnet method train crit = nn.NLLLoss(ignore_index=-1) CE_loss = self.crit(pred, feed_dict[‘seg_label’] batch_data -> feed_dict train.py->main net_encoder = ModelBuilder.build_encoder( arch=cfg.MODEL.arch_encoder.lower(), fc_dim=cfg.MODEL.fc_dim, weights=cfg.MODEL.weights_encoder) net_decoder = ModelBuilder.build_decoder( arch=cfg.MODEL.arch_decoder.lower(), fc_dim=cfg.MODEL.fc_dim, num_class=cfg.DATASET.num_class, weights=cfg.MODEL.weights_decoder) Where is prototype? No prototype in Loss Prototype should be in classifier( probability procedure ) self.centers should be prototypes why prototype is 3?...

划分Patch：为了减少序列长度，按像素计算长度过长，将16x16的Patch作为一个元素，生成序列 Transformer vs. CNN: Transformer相比卷积缺少归纳偏置(inductive biases, 先验知识或者先验假设) CNN的inductive biases Locality 局部性 translation equivalence 平移不变性 Transformer在大规模数据集上预训练，可以达到归纳偏置d

sparse conv

层级式 Shift Window 在小窗口上计算self attention patch merging 类似于 max pooling transformer 目的：更好理解上下文 主要对比：ViT ​ 实现了多尺寸的特征

提出 REAL Related Work 传统 closed-set LIDAR语义分割 point-based methods: point-net / point-net++ 但是有性能限制 voxel-based methods: Cylinder3D，三维数据集SOTA，本文的base architecture Cylinder3D论文解读及代码略解 open-set 但是2D： uncertainty-based methods： generative model-based methods： Stochastic Gradient Langevin Dynamics (SGLD) Open-world Methodology REAL: Redundancy Classifier Framework Oseg: Open-set Semantic Segmentation 将模型从$\mathcal{M}_c$训练到$\mathcal{M}_o$ ，提出两个训练策略：Unknown Object Synthesis，Predictive Distribution Calibration Unknown Object Synthesis Predictive Distribution Calibration deep metric learning open-world 和 open-set 有何不同 不必太关注open-world 与 open-set 差异，但可以多留意生成式模型的后续研究 为什么uncertainty-based methods: the network predicts the novel classes as old classes with high confidence scores...

论文链接： Abstract 提出： Open World Object Detection 问题 模型能够在无明确监督的情形下将无标签物体识别为’unknown' 当逐步输入相应的标签时，能够学习到已识别的’unknown’类别中去 Open World Object Detector ORE 基于Contrastive clustering和Energy based unknown identification 发现： we find that identifying and characterising unknown instances helps to reduce confusion in an incremental object detection setting, where we achieve state-ofthe-art performance, with no extra methodological effort. 我们发现识别和表征未知实例有助于减少incremental object detection设置中的混淆，在这种设置中我们实现了SOTA，而无需额外的方法论。 Intro 问题对比： Open set learning / Incremental Learning / Open world object Detection 方法 ORE Faster R-CNN 作为detector...

OpenAI CLIP 官方文档 zero-shot 基于图像文本对 学习 语义信号 模型过大，为加速训练采用很多加速技术：How to Train Really Large Models on Many GPUs prompt engineering 作者提出在对比学习中再结合生成式的目标函数，或许可以扩大模型的灵活性，不受限于输入文本数据的标签 主要影响：摆脱了原有训练数据 固定类别 的范式 Question fine-tune / linear porbe logistic regression

特征 self-supervised learning 基于transformer BERT 与 GPT ​ BERT使用编码器而GPT使用解码器(Transformer) ->GPT2 ​ 提出zero-shot ->GPT3 ​ 用few-shot ​ 在做子任务时，不需要微调，不需计算重新更新梯度 Question： zero-shot 为什么大模型不容易过拟合？ (为什么：在大模型中不需要像较小模型一样选择较小的Batch_Size来使每次采样时数据中的噪音增多来避免过拟合)

阅读顺序