引用nl · 2016-01
“[1] is a Google-authored demo running Inception-v3[2] (about twice as accurate as VGG) on Android phones.”
引用 Inception-v3 说明它比 VGG 准确度高约一倍
用分解卷积与强正则化高效扩展卷积网络,即 Inception-v3。
论文探讨如何尽可能高效地利用额外算力来扩大卷积网络,办法是适当分解卷积并采用较强的正则化。在 ILSVRC 2012 分类验证集上,单帧评估取得 21.2% top-1 和 5.6% top-5 错误率,网络计算量为 50 亿次乘加、参数不到 2500 万;用 4 个模型集成加多裁剪评估,验证集 top-5 错误率为 3.5%(测试集 3.6%),top-1 错误率为 17.3%,较之前最优有明显提升。
评论者用它说明 Inception-v3 比 VGG 更准确也更省参数、适合在手机上运行,并把它当作该架构细节的出处。
“[1] is a Google-authored demo running Inception-v3[2] (about twice as accurate as VGG) on Android phones.”
“Here is a paper if others are wondering:”
“See the Inception-v3 paper: Rethinking the Inception Architecture for Computer Vision”
“itself a 12x reduction over AlexNet (60M) and VGGNet (180M)! (source: Rethinking the Inception Architecture for Computer Vision)”