mirror of
https://github.com/zoo-apps/papers.git
synced 2026-10-30 14:56:54 +00:00
27 lines
1.2 KiB
TeX
27 lines
1.2 KiB
TeX
\section{Implementation Details}
|
|
|
|
\label{sec:implementation}
|
|
% ══════════════════════════════════════════════════════════════════════════════
|
|
|
|
\subsection{Software and Hardware}
|
|
|
|
TaxoNet is implemented in PyTorch 2.2 with the \texttt{timm} library for
|
|
ViT models. Training was conducted on 8 NVIDIA A100 GPUs for
|
|
approximately 72 hours. Inference is deployed on NVIDIA T4 GPUs behind a
|
|
Kubernetes load balancer.
|
|
|
|
\subsection{Data Augmentation}
|
|
|
|
Training episodes apply random resized crop (scale 0.4--1.0), horizontal
|
|
flip, color jitter, random Gaussian blur, random grayscale (probability
|
|
0.1), and cutout with one $64 \times 64$ hole.
|
|
|
|
\subsection{Confidence Calibration}
|
|
|
|
We calibrate prediction confidence using temperature scaling
|
|
\citep{guo2017calibration}. Post-calibration, expected calibration error
|
|
(ECE) drops from 8.3\% to 2.1\%.
|
|
|
|
|
|
% ══════════════════════════════════════════════════════════════════════════════
|