Rico贾若童的博客

[BEV] BEV Ideas and View Transformation

Training pipelines, multiview geometry, and how to lift image features into BEV

1. The Core Goal of a BEV Network A Bird’s Eye View (BEV) network aims to reconstruct, in real time using only onboard cameras, a local 3D map of comparable quality to what an offline reconstructi...

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

[BEV] BEV Introduction: Tesla's Challenges and Architecture

Why per-camera detection falls short, and how BEV solves it

1. Tesla’s Perception Challenges (2021) Before BEV, Tesla’s pipeline detected objects and lanes independently in each camera view and then tried to fuse the results. This created fundamental probl...

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

[BEV] BEV Introduction

Understanding L2 and L2+ Perception: Why the Real Challenge Is System Design, Not Just Better Detection When people talk about assisted driving, they often jump straight to compute power, percepti...

Posted by Rico's Nerd Cluster on April 2, 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

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

Rico's Nerd Cluster