[Matting] Update README and Fix Export Bug Of HumanMatting (#1730)

Author: wuyefeilin

contrib/Matting/README.md

@@ -200,10 +200,12 @@ python bg_replace.py \
     --bg_path path/to/your/background/image \
     --save_dir ./output/results
 ```
-If the model requires trimap information, pass the trimap path through '--trimap_path'.
+If the model requires trimap information, pass the trimap path through `--trimap_path`.
 
 If `--bg_path` is not provided, green background is used。
 
+**note：** `--image_path` must be a image path。
+
 You can directly download the provided model for background replacement.
 
 Run the following command to view more parameters.
@@ -231,7 +233,7 @@ python export.py --help
 python deploy/python/infer.py \
     --config output/export/deploy.yaml \
     --image_path data/PPM-100/val/fg/ \
-    --save_dir ouput/results
+    --save_dir output/results
 ```
 If the model requires trimap information, pass the trimap path through '--trimap_path'.
 

contrib/Matting/README_CN.md

@@ -208,6 +208,8 @@ python bg_replace.py \
 
 若不提供`--bg_path`, 则采用绿色作为背景。
 
+**注意：** `--image_path`必须是一张图片的具体路径。
+
 你可以直接下载我们提供的模型进行背景替换。
 
 更多参数信息请运行如下命令进行查看:
@@ -235,7 +237,7 @@ python export.py --help
 python deploy/python/infer.py \
     --config output/export/deploy.yaml \
     --image_path data/PPM-100/val/fg/ \
-    --save_dir ouput/results
+    --save_dir output/results
 ```
 如模型需要trimap信息，需要通过`--trimap_path`传入trimap路径。
 

contrib/Matting/configs/human_matting/human_matting_resnet34_vd.yml

@@ -39,7 +39,7 @@ model:
   type: HumanMatting
   backbone:
     type: ResNet34_vd
-    pretrained: https://bj.bcebos.com/paddleseg/dygraph/ResNet34_vd_pretrained.pdparams
+    pretrained: https://paddleseg.bj.bcebos.com/matting/models/ResNet34_vd_pretrained/model.pdparams
   pretrained: Null
   if_refine: True
 

contrib/Matting/deploy/human_matting_android_demo/README.md

@@ -1,7 +1,7 @@
 English | [简体中文](README_CN.md)
 
 # Human Matting Android Demo
-Based on [PaddleSeg](https://github.com/paddlepaddle/paddleseg/tree/develop)[MODNet](https://github.com/PaddlePaddle/PaddleSeg/tree/develop/contrib/Matting) algorithm to realise human matting（Android demo).
+Based on [PaddleSeg](https://github.com/paddlepaddle/paddleseg/tree/develop) [MODNet](https://github.com/PaddlePaddle/PaddleSeg/tree/develop/contrib/Matting) algorithm to realise human matting（Android demo).
 
 You can directly download and install the example project [apk](https://paddleseg.bj.bcebos.com/matting/models/deploy/app-debug.apk) to experience。
 

contrib/Matting/deploy/human_matting_android_demo/README_CN.md

@@ -26,7 +26,7 @@
 *注：此安卓demo基于[Paddle-Lite](https://paddlelite.paddlepaddle.org.cn/)开发，PaddleLite版本为2.8.0。*
 
 ### 2.3 预测
-* 在人像抠图Demo中，默认会载入一张人像图像，并会在图像下方给出CPU的预测结果和预测时延；
+* 在人像抠图Demo中，默认会载入一张人像图像，并会在图像下方给出CPU的预测结果和预测延时；
 * 在人像抠图Demo中，你还可以通过右上角的"打开本地相册"和"打开摄像头拍照"按钮分别从相册或相机中加载测试图像然后进行预测推理；
 
 *注意：demo中拍照时照片会自动压缩，想测试拍照原图效果，可使用手机相机拍照后从相册中打开进行预测。*
@@ -80,7 +80,7 @@ val_dataset:
   mode: val
   get_trimap: False
 ```
-上述修改中尤其注意short_size: 256这个字段，这个值直接决定我们最终的推理图像采用的尺寸大小。这个字段值设置太小会影响预测精度，设置太大会影响手机推理速度（甚至造成手机因性能问题无法完成推理而奔溃）。经过实际测试，对于hrnet18，该字段设置为256较好。
+上述修改中尤其注意short_size: 256这个字段，这个值直接决定我们最终的推理图像采用的尺寸大小。这个字段值设置太小会影响预测精度，设置太大会影响手机推理速度（甚至造成手机因性能问题无法完成推理而崩溃）。经过实际测试，对于hrnet18，该字段设置为256较好。
 
 完成配置文件修改后，采用下面的命令进行静态图导出：
 ``` shell

contrib/Matting/model/human_matting.py

@@ -162,12 +162,14 @@ def __init__(self,
 
     def forward(self, data):
         src = data['img']
-        src_h, src_w = src.shape[2:]
+        src_h, src_w = paddle.shape(src)[2:]
         if self.if_refine:
-            if (src_h % 4 != 0) or (src_w % 4) != 0:
-                raise ValueError(
-                    'The input image must have width and height that are divisible by 4'
-                )
+            # It is not need when exporting.
+            if isinstance(src_h, paddle.Tensor):
+                if (src_h % 4 != 0) or (src_w % 4) != 0:
+                    raise ValueError(
+                        'The input image must have width and height that are divisible by 4'
+                    )
 
         # Downsample src for backbone
         src_sm = F.interpolate(
@@ -366,7 +368,7 @@ def forward(self, src, pha, err, hid, tri):
             hid: (B, 32, Hc, Hc) coarse hidden encoding.
             tri: (B, 1, Hc, Hc) trimap prediction.
         '''
-        h_full, w_full = src.shape[2:]
+        h_full, w_full = paddle.shape(src)[2:]
         h_half, w_half = h_full // 2, w_full // 2
         h_quat, w_quat = h_full // 4, w_full // 4