After installing Nvidia Driver and CUDA on a Nvidia Optimus laptop (following instructions on if-not-true-then-false), Fedora (34+) often gets stuck after several minutes poweron.

GIMP

less than 1 minute read

GIMP, notable software for image processing in linux

Useful Shell Commands

less than 1 minute read

Count number of fiels in a dir

ls | wc -l, all regular files/folders just in this dir
ls -a | grep ^\\. | wc -l, all hidden files/folders just in thid dir
tree -L 1, list tree just in this dir, and count number of files and folders respectively
find ./ -maxdepth 1 -type f,d | wc -l, but will match current dir and hidden files/folders

Explanations in Neural Networks

less than 1 minute read

Interpretability (of a DNN) is the ability to provide explanations in understandable terms to a human. F Doshi-Velez & B Kim, 2017。

Causality preliminary

2 minute read

This is an reading notes from an causality introduction website

basis for convex optimization

1 minute read

Notable Concepts

line / (closed) line segments: \theta * x1 + (1 - \theta) * x2, theta in R / in [0, 1]
affine set / convex set: \theta * x1 + (1 - \theta) * x2, theta in R / in [0, 1]
cone: theta * x, theta in R+ (nonnegative real values)
hyperplane / halfspace: a^T x = b or a^T (x - x0) = 0 / a^T x < b, a^T x > b or a^T (x - x0) < 0, a^T (x - x0) > 0
balls / ellipsoids: (x - xc)^T P (x - xc) <= r^2 , P = I or other kinds of positive definite solution.
polyhedron: the solution set of a finite number of linear equalities and inequalities
simplex: conv hull of k+1 affinely independent points, and its affine dimensionality is k. such as line segment, triangle (with interior), tetrahedron (with interior). As well as, unit simplex (x >= 0, 1^T * x <= 1), or probability simplex (x >= 0, 1^T * x = 1)
partial order, minimum (最小, for each s in S, s <= s0, or, S included in s0+K), minimal (极小, if s in S s.t. s<= s0, then s = s0, or, Intersection(so-K, S) = {x} )
supporting hyperplane

Common CV Tasks

less than 1 minute read

Common tasks in computer vision:

recognition
classification, image class labels
localization, per-object bounding box labels
detection, per-object class+bounding box labels
tracking
semantic segmentation, per-pixel class labels (all pixels, but do not distinct objects with the same class label)
instance segmentation, per-object mask and class labels (only objects)
panoptic segmentation, per-pixel class+instance labels (objects and background things)