4 步交替 RPN/更快的 R-CNN 训练?- Tensorflow 对象检测模型 [英] 4-step Alternating RPN / Faster R-CNN Training? - Tensorflow Object Detection Models

问题描述

正在浏览最近发布的 tensorflow/models/../object_detection 模型，特别是更快的 r-cnn.

Been going through the recently released tensorflow/models/../object_detection models, particularly faster r-cnn.

论文中提到了 4 步交替训练，你会在这里

The paper mentions 4-step alternating training, where you would

1. 训练 RPN，然后冻结 RPN 层，
2. 训练 RCNN，然后冻结 RCNN 层，
3. 训练RPN，然后冻结RPN层
4. 训练 RCNN.

据我所知，在第 2 阶段 = RCNN，RPN 确实被冻结:

From what I gather, at stage 2 = RCNN, RPN is indeed frozen with:

if self._is_training:
    proposal_boxes = tf.stop_gradient(proposal_boxes) 

那么train RPN + freeze RPN Layers，然后是RCNN训练就讲完了，但是其他3个步骤在哪里执行呢?

So train RPN + freeze RPN Layers, followed by RCNN training is covered, but where are the other 3 steps performed?

我错过了什么吗?

推荐答案

我们在 TF 对象检测 API 中实现的 Faster R-CNN 与论文非常接近，但在一些方面有所不同.其中一个区别是我们端到端地训练模型，而不是论文中使用的交替训练.

Our implementation of Faster R-CNN in the TF Object Detection API follows the paper quite closely but differs in a few ways. And one of those differences is that we train the model end-to-end instead of the alternating training used in the paper.

您提到的 stop_gradient 实际上并没有冻结 RPN --- 它的作用是忽略梯度通过建议坐标的贡献，但继续允许梯度通过 RPN 特征.

The stop_gradient that you mention doesn't actually freeze the RPN --- what it does is it ignores the contribution of the gradient through the proposal coordinates, but continues to allow the gradient to pass through the RPN features.

希望这有帮助！

这篇关于4 步交替 RPN/更快的 R-CNN 训练?- Tensorflow 对象检测模型的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！