Rico贾若童的博客

Deep Learning - Mixed Floating Point Training

FP16, BF16, Mixed Precision Training

Refresher: Floating Point Calculation A floating point is represented as sign bit | exponent | mantissa. 0 | 10000001 | 10000000000000000000000 represents 6 because: Sign bit 0 represents posi...

Posted by Rico's Nerd Cluster on February 14, 2026

[ML] Libraries For Point Cloud Compression

einops

einops Einops is a lightweight Python library that makes tensor reshaping, permutation, tiling, and reduction readable and explicit. It works with PyTorch, NumPy, and TensorFlow. rearrange — r...

Posted by Rico's Nerd Cluster on February 12, 2026

[ML] D-PCC Losses

Chamfer Loss, Density Loss, Point-Count Loss, Rate-Distortion Loss

Entropy Encoding What is the bottleneck? In learned compression the entropy bottleneck is a learned probability model $p(\hat{z})$ placed on the quantized latent features $\hat{z}$. It serves two...

Posted by Rico's Nerd Cluster on February 11, 2026

Entropy Bottleneck

Entropy Encoding, Range Encoding

Problem Setup Suppose one encoder output (latent vector) is z = [0.13, -1.82, 0.07]. Storing these as float32 costs $3 \times 32 = 96$ bits. Entropy coding requires discrete symbols, so we first q...

Posted by Rico's Nerd Cluster on February 8, 2026

[ML] D-PCC Decoder Layers

isohedron, losses

Decoder Overview The overall decoder is looking to achieve the below: Icosahedron vertices on a unit sphere become candidate directions — unit vectors pointing from a...

Posted by Rico's Nerd Cluster on February 6, 2026

[ML] D-PCC Encoder-Layers

sub-pixel convolution

D-PCC paper Terminology latent feature: a compressed internal representation that contains essential information needed to reconstruct the original data. Cardinality: the number of elements ...

Posted by Rico's Nerd Cluster on February 3, 2026

[ML] Point Cloud Transformer

Vector Attention

Terminology modulate In ML, modulation means changing one signal using another signal. So Attention $output=\alpha \odot v$ with elementwise product is modulation on value fea...

Posted by Rico's Nerd Cluster on February 2, 2026

[ML] OpenClaw

OpenClaw Video: https://www.youtube.com/watch?v=Otn-NbpQH1k Start Ollama: 1 2 3 4 5 6 ollama run http://localhost:11434/ ollama launch openclaw # on the control panel /model qwen2.5:7b...

Posted by Rico's Nerd Cluster on February 1, 2026

Nvdiffrast In FoundationPose-style tracking, nvdiffrast + kornia are the GPU crop/render pair builder before the pose networks. The pose network does not directly look at the whole camera frame. I...

Posted by Rico's Nerd Cluster on July 5, 2026

[CV] Rasterization

Barycentric Coordinates

Goal of Rasterization: Project 3D Triangles Down to 2D Triangles Let’s pretend a mini 3D mesh has two triangles: 1 2 3 4 faces = [ [0, 1, 2], # triangle 1 [3, 4, 5], # triangle 2 ] We...

Posted by Rico's Nerd Cluster on January 26, 2026

Rico's Nerd Cluster