【论文笔记】OverFeat
原文见OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks:
- 新的网络结构:两种网络模型
- 多尺度应用
- classification, localization, detection
Classification
Model Design and Training
fast model
Layer | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
---|---|---|---|---|---|---|---|---|
Stage | conv + max | conv + max | conv | conv | conv + max | full | full | full |
channels | 96 | 256 | 512 | 1024 | 1024 | 3072 | 4096 | 1000 |
Filter size | 11x11 | 5x5 | 3x3 | 3x3 | 3x3 | - | - | - |
Conv. stride | 4x4 | 1x1 | 1x1 | 1x1 | 1x1 | - | - | - |
Pooling size | 2x2 | 2x2 | - | - | 2x2 | - | - | - |
Pooling stride | 2x2 | 2x2 | - | - | 2x2 | - | - | - |
Zero-Padding size | - | - | 1x1x1x1 | 1x1x1x1 | 1x1x1x1 | - | - | - |
Spatial input size | 231x231 | 28x28 | 12x12 | 12x12 | 12x12 | 6x6 | 1x1 | 1x1 |
accurate model
Layer | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |
---|---|---|---|---|---|---|---|---|---|
Stage | conv + max | conv + max | conv | conv | conv | conv + max | full | full | full |
channels | 96 | 256 | 512 | 512 | 1024 | 1024 | 4096 | 4096 | 1000 |
Filter size | 7x7 | 7x7 | 3x3 | 3x3 | 3x3 | 3x3 | - | - | - |
Conv. stride | 2x2 | 1x1 | 1x1 | 1x1 | 1x1 | 1x1 | - | - | - |
Pooling size | 3x3 | 2x2 | - | - | - | 3x3 | - | - | - |
Pooling stride | 3x3 | 2x2 | - | - | - | 3x3 | - | - | - |
Zero-Padding size | - | - | 1x1x1x1 | 1x1x1x1 | 1x1x1x1 | 1x1x1x1 | - | - | - |
Spatial input size | 221x221 | 36x36 | 15x15 | 15x15 | 15x15 | 15x15 | 5x5 | 1x1 | 1x1 |