Rico贾若童的博客

Computer Vision - Pose Estimate Pipeline

Building a 2D-to-6D Object Pose Pipeline with RF-DETR and FoundationPose I built a small pose-estimation pipeline that combines RF-DETR and FoundationPose. RF-DETR solves the 2D detection problem:...

Posted by Rico's Nerd Cluster on March 13, 2026

[CUDA - 6] CUDA Inference Speedups

inference mode, jit tracing, cuda graphs, tensor rt, optimize_for_inference

Introduction I built a RF-DETR + FoundationPose pipeline. the clean optimization stack is: Make inference measurable. Log per-stage latency: detector, register, track, publish, visualize. Us...

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

DETR

Embedding, Encoding An embedding is a learnable vector, part of an embedding table A, which is similar to a fully connected layer y = x @ W.T + b. However, an embedding usually does: 1 one_hot(in...

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

ViT

ViT in short: Image -> 16x16 patches [16 x 16 x 3] -> project to [256/768] token embedding TODO

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

Model Parameter vs VRAM

VRAM

VRAM = Video RAM. It is the memory on your GPU, separate from your normal computer RAM. The GPU uses VRAM to store: model weights input images / batches intermediate activations gradients...

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

[Point Cloud Compression] Draco

Introduction Draco is Google’s open-source library for compressing 3D geometric meshes and point clouds. It quantizes floating-point attributes (positions, normals, colors, texture coordinates) in...

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

[ML] HuggingFace Trainer

Introduction The Trainer class provides a feature-complete training loop for PyTorch, supporting distributed training on multiple GPUs/TPUs and mixed precision via NVIDIA Apex, AMD ROCm, and torch...

Posted by Rico's Nerd Cluster on February 28, 2026

[ML] Model Visualization

ONNX, Netron, Model Size

Netron Netron is a viewer for neural network models. The typical workflow is to export a PyTorch model to ONNX format and then drag the .onnx file into the Netron web app. What is ONNX? Open Neur...

Posted by Rico's Nerd Cluster on February 26, 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 23, 2026

PyTorch Mixed Precision Training

torch.zeros, GradScaler, GradCheck

Pytorch Setup 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 use_amp = True net = make_model(in_size, out_size, num_layers) opt = torch.optim.SGD(net.parameters(), lr=0.001) # if False, th...

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

Rico's Nerd Cluster