heiha1232的博客

2021.4ubuntu 18.04 cuda 10.2 pytorch onnx→tensorrt使用tensorrt加载并运行onnx模型0. TensorRT介绍1. 安装tensorrt、pycuda1.1 安装tensorrt1.2 安装pycuda2. 将pytorch模型转化为onnx模型3. tensorrt调用onnx模型进行预测4. 注意事项0. TensorRT介绍https://docs.nvidia.com/deeplearning/tensorrt/developer-gu

Ubuntu 20.04 c++ libtorch gpu1. Install nvidia-driver, cuda, cudnn according to your system2. Download libtorch (gpu)2.1 chose suitable package based on your system2.2 extract package3. Execute c++ code3.1 Obtain "model.pt" form pytorch3.2 load "model.pt"

【论文阅读】DeepGCNs: Can GCNs Go as Deep as CNNs?1. Introduction2. Methodology2.1 图上的表征学习图卷积神经网络（Graph Convolutional Neural networks, GCNs）用于表征非欧几里得空间（non-Euclidean）的数据。但由于梯度消失问题，GCNs的层数通常不深，感知域范围有限，本文提出了一种训练更深GCNs的方法。1. IntroductionGCN现阶段被广泛应用于social networ

【论文阅读】SEMI-SUPERVISED CLASSIFICATION WITH GRAPH CONVOLUTIONAL NETWORKS1. Introduction2. Fast Approximate Convolutions on Graphs2.1 Spectral Graph Convolution2.2 Layer-wise Linear model3. Semi-supervised Node Classification3.1 ExampleRef.:本文在图结构数据上提出一种半监督学

Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting论文阅读问题描述：传统LSTM结构ConvLSTM结构问题描述：时空域序列预测问题：需要通过过去的雷达图预测未来的雷达图：其中，χ∈RP×M×N\chi\in R^{P\times M\times N}χ∈RP×M×N，表示面积范围 M×NM\times NM×N 的 PPP 个观测，需要通过过去的JJJ个观测预测未来KKK个观测。

多任务情况下各项loss权重平衡的问题Focal Loss的解决方案多任务情况下的loss：L=αl1+βl2+γl3L = \alpha l_1 + \beta l_2 + \gamma l_3L=αl1​+βl2​+γl3​其中，l1,l2,l3l_1,l_2,l_3l1​,l2​,l3​分别表示不同任务的loss项，需要对这些loss项进行同时优化，那么如何平衡这几个loss项？如何使...

Transforming AutoecodersHinton G E , Krizhevsky A , Wang S D . Transforming Auto-Encoders[C]// Artificial Neural Networks and Machine Learning - ICANN 2011 - 21st International Conference on Artifici...

Group Equivariant Convolutional NetworksCohen T, Welling M. Group equivariant convolutional networks[C]//International conference on machine learning. 2016: 2990-2999.https://github.com/tscohen/gcon...