Fast Neural Architecture Search of Compact Semantic Segmentation Models via Auxiliary Cells

[junxnone]

Reference

• 10/2018 Fast Neural Architecture Search of Compact Semantic Segmentation Models via Auxiliary Cells
• 04/2019 Template-Based Automatic Search of Compact Semantic Segmentation Architectures
• Code - nas-segm-pytorch
• 论文笔记：Fast Neural Architecture Search of Compact Semantic Segmentation Models via Auxiliary Cells
• 论文笔记11-基于辅助单元语义分割模型搜索

Brief

• Encoder-Decoder
• PPO Algos
• inner optimisation - Sampled Architecture
• outer optimisation - Controller
• Cell & Auxiliary Cell - 相同结构，但是不同 weights
• Polyak Averaging & Knowledge Distillation - 加速收敛

Arch

Search Space

Sample Cells Operation

conv 1 × 1, conv 3 × 3, separable conv 3 × 3, separable conv 5 × 5, global average pooling followed by upsampling and conv 1 × 1, conv 3 × 3 with dilation rate 3, conv 3 × 3 with dilation rate 12, separable conv 3 × 3 with dilation rate 3, separable conv 5 × 5 with dilation rate 6, skip-connection, zero-operation that effectively nullifies the path

Auxiliary Cell

• 结构同 Cell
• 用于产生 Auxiliary Classifier 辅助输出
• 中间监督
• 仅在 辅助输出上计算 Segmentation Loss

Polyak averaging

Knowledge Distillation

Training

• first Stage - 固定 Encoder Weights ，training Decoder
• Second Stage - Training 整个 Model

Search Result

[junxnone]

825