Rico贾若童的博客

[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] PyTorch Functions

autograd Function, Convolution, Normalization, Sum, Torch Cache Emptying

Torch.autograd.Function torch.autograd.Function lets you define a custom op with explicit forward() and backward() passes. Common built-ins like MaxPool are implemented this way. It’s especially u...

Posted by Rico's Nerd Cluster on February 9, 2026

[ML] D-PCC Decoder Layers

isohedron, losses

Decoder We need multiple upsampling block, instead of a major one. Like instead of doing upsampling x3 -> x3 -> x3, we just do x27. Why? upsampling x27 in training might require large ...

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 features. Th...

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

[CUDA - 6] CUDA Functions

attomicAdd, pragma unroll

atomicAdd There’s no such thing as “atomic memory” in CUDA—atomicity is a property of an operation, not a memory type. atomicAdd is an instruction you apply to a memory location (typically in gl...

Posted by Rico's Nerd Cluster on January 21, 2026

[CUDA - 5] SIMT in CUDA

SIMD SIMT

SIMT and SIMD CUDA is a very good embodiment of SIMD (Single-Instruction-Multiple Data). SIMD is great for addressing embarassingly parallel problems, problems that are so “embarassingly” simple...

Posted by Rico's Nerd Cluster on January 19, 2026

[CUDA - 4] My First CUDA Kernel - Chamfer Distance

First CUDA Program, JIT Compile

On Chamfer Distance Chamfer distance measures how close two point clouds are by averaging nearest-neighbor distances in both directions. Given two point sets $P_1$ and $P_2$, For each point ...

Posted by Rico's Nerd Cluster on January 18, 2026

PyTorch Mixed Precision Training

torch.zeros

torch.zeros By default: 1 out = torch.zeros(b, c, m).to(features.device) # allocates on CPU, then transfer to device torch.zeros(...) creates a tensor with dtype torch.float32 unless you ...

Posted by Rico's Nerd Cluster on January 17, 2026

Rico's Nerd Cluster