Home

Code: https://github.com/LEGO999/A-tutorial-for-few-shot-learning Prof. Shusen Wang at the Stevens Institute of Technology provided an informative tutorial for few-shot learning and metric learning. I took lecture notes and complemented this tutorial with additional materials and code built on PyTorch. Definition of few-shot learning Few-shot l...

Code: https://github.com/LEGO999/AD_robustness Deep Neural Networks (DNNs) are known to be prone to adversarial attacks. Adversarial attacks manipulate predictions of DNNs via adding crafted perturbations, as shown in Fig.1. Fig.1 An adversarial example generated by FGSM (source: Goodfellow et al., 2014) In essence, adversarial attacks opt...

Kullback-Leibler divergence (KL divergence) is a preferable alternative for cross-entropy with soft labels as the reason is explained in the previous post. The mathematical form of KL divergence is demonstrated in following equations: \(D_{KL} = \sum \mathbf{g}(x) \log\frac{\mathbf{g}(x)}{\mathbf{f}(x)}\) , where $\mathbf{f}(x)$ is predicted pro...

Shape Difference Both torch.cat() and torch.stack() will concatenate a sequence of tensors. torch.cat() will concatenate tensors along a given dimension rather than a new dimension. import torch concat_tensor = torch.cat([torch.rand(2,3) for _ in range(2)], dim=0) print(f'The shape of concatenated tensor is {concat_tensor.shape}') stack_tenso...

PyTorch tensors could be either CPU or GPU (CUDA). Numpy arrays are always in the memory (CPU). Pure CPU Numpy array and CPU tensor When a CPU tensor is built based on a Numpy tensor, the system will create a new copy in the memory. They won’t have influence on each other. import torch import torch.nn as nn import numpy as np # c...

After the operations like writer.add_scalar(), please don’t forget to use writer.flush() to write the buffer to the file.

Beginning of the .sh files #!/bin/sh # Must be put at the beginning of the .sh file Create directory mkdir $dir_name If the directory existed, commands in the following steps will be stopped and warning will be sent. Create directory with parents directories mkdir -p $dir_name If the directory existe...

Task of 3D Deep Learning classification segmentation reconstruction Representation Form rasterized form (regular) geometric form (irregular) Types multi-view volumetric part assembly point cloud mesh implicit (e.g. function space, occupancy network) Datasets ShapeNet PartNet SceneNet ScanNet KITTI Clas...