-[Inference and Visualization](#inference-and-visualization)
-[安装](#安装)
-[简介](#简介)
-[数据准备](#数据准备)
-[模型训练](#模型训练)
-[模型评估](#模型评估)
-[模型推断及可视化](#模型推断及可视化)
## Installation
## 安装
Running sample code in this directory requires PaddelPaddle Fluid v.1.3.0 and later. If the PaddlePaddle on your device is lower than this version, please follow the instructions in [installation document](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/install/index_en.html) and make an update.
Region Convolutional Neural Network (RCNN) models are two stages detector. According to proposals and feature extraction, obtain class and more precise proposals.
Now RCNN model contains two typical models: Faster RCNN and Mask RCNN.
1.Base conv layer. As a CNN objective dection, Faster RCNN extract feature maps using a basic convolutional network. The feature maps then can be shared by RPN and fc layers. This sampel uses [ResNet-50](https://arxiv.org/abs/1512.03385) as base conv layer.
2.Region Proposal Network (RPN). RPN generates proposals for detection。This block generates anchors by a set of size and ratio and classifies anchors into fore-ground and back-ground by softmax. Then refine anchors to obtain more precise proposals using box regression.
3. RoI Align. This layer takes feature maps and proposals as input. The proposals are mapped to feature maps and pooled to the same size. The output are sent to fc layers for classification and regression. RoIPool and RoIAlign are used separately to this layer and it can be set in roi\_func in config.py.
4.Detection layer. Using the output of roi pooling to compute the class and locatoin of each proposal in two fc layers.
Mask RCNN is a two stage model as well. At the first stage, it generates proposals from input images. At the second stage, it obtains class result, bbox and mask which is the result from segmentation branch on original Faster RCNN model. It decouples the relation between mask and classification.
**download the pre-trained model:** This sample provides Resnet-50 pre-trained model which is converted from Caffe. The model fuses the parameters in batch normalization layer. One can download pre-trained model as:
- Set ```export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7``` to specifiy 8 GPU to train.
- Set ```MASK_ON``` to choose Faster RCNN or Mask RCNN model.
- Set ```parallel``` to False to replace [fluid.ParallelExecutor](http://paddlepaddle.org/documentation/docs/zh/1.4/api_cn/fluid_cn.html#parallelexecutor) to [fluid.Executor](http://paddlepaddle.org/documentation/docs/zh/1.4/api_cn/fluid_cn.html#executor) when running the program in the Windows & GPU environment.
* NMS threshold=0.7. During training, pre\_nms=12000, post\_nms=2000; during test, pre\_nms=6000, post\_nms=1000.
* In generating proposal lables, fg\_fraction=0.25, fg\_thresh=0.5, bg\_thresh_hi=0.5, bg\_thresh\_lo=0.0.
* In rpn target assignment, rpn\_fg\_fraction=0.5, rpn\_positive\_overlap=0.7, rpn\_negative\_overlap=0.3.
**training strategy:**
**训练策略：**
* Use momentum optimizer with momentum=0.9.
* Weight decay is 0.0001.
* In first 500 iteration, the learning rate increases linearly from 0.00333 to 0.01. Then lr is decayed at 120000, 160000 iteration with multiplier 0.1, 0.01. The maximum iteration is 180000. Also, we released a 2x model which has 360000 iterations and lr is decayed at 240000, 320000. These configuration can be set by max_iter and lr_steps in config.py.
* Set the learning rate of bias to two times as global lr in non basic convolutional layers.
* In basic convolutional layers, parameters of affine layers and res body do not update.
Evaluation is to evaluate the performance of a trained model. This sample provides `eval_coco_map.py` which uses a COCO-specific mAP metric defined by [COCO committee](http://cocodataset.org/#detections-eval).
* Fluid mask no padding: Use RoIAlign. Images without padding.
* Fluid mask no padding: 使用RoIAlign，不对图像做填充处理
## Inference and Visualization
## 模型推断及可视化
Inference is used to get prediction score or image features based on trained models. `infer.py` is the main executor for inference, one can start infer step by:
模型推断可以获取图像中的物体及其对应的类别，`infer.py`是主要执行程序，调用示例如下：
```
python infer.py \
...
...
@@ -194,17 +195,17 @@ python infer.py \
--draw_threshold=0.6
```
Please set the model path and image path correctly. GPU device is used by default, you can set `--use_gpu=False` to switch to CPU device. And you can set `draw_threshold` to tune score threshold to control the number of output detection boxes.
