人工智能培训

搜索

论文代码开源:火炬手:PYTORCH的模型拟合库(Torchbearer: A Model Fitting Library for PyTorch)

[复制链接]
admin 发表于 2018-9-15 09:55:21 | 显示全部楼层 |阅读模式
admin 2018-9-15 09:55:21 614 0 显示全部楼层
人工智能论文代码开源:火炬手:PYTORCH的模型拟合库(Torchbearer: A Model Fitting Library for PyTorch)请注意该人工智能论文代码开源在github,大部分是python写的,框架可能是tensorflow或者pytorch。度量学习旨在学习与样本的语义含义一致的距离。通常通过学习每个样本的嵌入来解决该问题,使得相同类别的样本的嵌入是紧凑的,而不同类别的样本的嵌入在特征空间中展开。我们研究了从基于深度神经网络的分类器的第二层提取的特征,该分类器在softmax层之上用交叉熵损失训练。我们证明了具有不同温度值softmax函数的训练分类器导致具有不同紧凑程度的特征。利用这些见解,我们提出了一种“加热”策略来训练分类器增加温度,导致相应的嵌入在各种度量学习基准上实现最先进的性能。
Metric learning aims at learning a distance which is consistent with thesemantic meaning of the samples.The problem is generally solved by learning anembedding for each sample such that the embeddings of samples of the samecategory are compact while the embeddings of samples of different categoriesare spread-out in the feature space.We study the features extracted from thesecond last layer of a deep neural network based classifier trained with thecross entropy loss on top of the softmax layer.We show that trainingclassifiers with different temperature values of softmax function leads tofeatures with different levels of compactness.Leveraging these insights, wepropose a "heating-up" strategy to train a classifier with increasingtemperatures, leading the corresponding embeddings to achieve state-of-the-artperformance on a variety of metric learning benchmarks.论文代码开源:火炬手:PYTORCH的模型拟合库(Torchbearer: A Model Fitting Library for PyTorch) OhAHmH3yV383gYv8.jpg
URL地址:https://arxiv.org/abs/1809.04157v1     ----pdf下载地址:https://arxiv.org/pdf/1809.04157v1    ----         ----github下载地址:https://github.com/ColumbiaDVMM/Heated_Up_Softmax_Embedding    ----    论文代码开源:火炬手:PYTORCH的模型拟合库(Torchbearer: A Model Fitting Library for PyTorch)请注意该人工智能论文代码开源在github,大部分是python写的,框架可能是tensorflow或者pytorch,keras,至于具体是哪一个没有完全测试。
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

admin当前离线
管理员

查看:614 | 回复:0

快速回复 返回顶部 返回列表