'use client'; import React from 'react'; import styles from './MobileNetV3Diagram.module.css'; export default function MobileNetV3Diagram() { return (
MobileNetV3-Large Student
Destilado dos 5 Professores ResNet-18  ·  66.6 MB FP32  ·  1026 classes
Input Tensor
128 × 128 × 1 — grayscale silhouette
Conv2d Adaptado
in=1  |  16 filtros  |  3×3  |  stride 2
Blocos Inverted Residual + SE
Bottleneck ×3 — expansão 16→64
DWConv 3×3  |  SE squeeze 4→16  |  stride 2  |  h-swish
Bottleneck ×3 — expansão 24→88
DWConv 3×3  |  SE squeeze 4→24  |  stride 2  |  ReLU
Bottleneck ×6 — expansão 40→240
DWConv 5×5  |  SE squeeze 8→40  |  h-swish  |  stride 2
Bottleneck ×4 — expansão 80→400
DWConv 3×3  |  SE squeeze 16→80  |  h-swish  |  stride 1
Bottleneck ×2 — expansão 112→672
DWConv 5×5  |  SE squeeze 24→112  |  h-swish  |  stride 1
Bottleneck ×3 — expansão 160→960
DWConv 5×5  |  SE squeeze 32→160  |  h-swish  |  stride 2
Conv2d 1×1
960 → 1280  |  h-swish
Global Average Pool
AvgPool 7×7  →  1 × 1 × 1280
Flatten + Dropout
1280-dim  |  p=0.2
FC — Classifier Head
Linear(1280 → 1026)  |  1025 target species + 1 OOD rejection sink
↓
Destilado de 5 Professores ResNet-18 (230 MB) via Ltotal = (1−α)Lhard + αLsoft + γLOOD
I/O
Convolucional
Bottleneck + SE
Pooling / Flatten
); }