!3852 [华为大学][自研贡献][MindSpore][deeplabv3]-回归精度不达标修改

Merge pull request !3852 from 唐红艳/master

official/cv/deeplabv3/infer/Dockerfile

 ARG FROM_IMAGE_NAME
 FROM ${FROM_IMAGE_NAME}
 
+USER root
 COPY requirements.txt .
 RUN pip3 install -r requirements.txt

official/cv/deeplabv3/infer/convert/convert_om.sh

@@ -34,5 +34,6 @@ atc --framework=1 \
     --model=$model_path \
     --output=$output_model_name \
     --fusion_switch_file=fusion_switch.cfg \
+    --precision_mode=allow_fp32_to_fp16 \
     --op_select_implmode=high_precision \
     --soc_version=Ascend310

official/cv/deeplabv3/infer/data/config/deeplabv3.pipeline

@@ -3,52 +3,12 @@
         "stream_config": {
             "deviceId": "0"
         },
-        "mxpi_imagedecoder0": {
-            "props": {
-                "handleMethod": "opencv"
-            },
-            "factory": "mxpi_imagedecoder",
-            "next": "mxpi_imageresize0"
-        },
-        "mxpi_imageresize0": {
-            "props": {
-                "handleMethod": "opencv",
-                "resizeType": "Resizer_KeepAspectRatio_Long",
-                "scaleValue": "513"
-            },
-            "factory": "mxpi_imageresize",
-            "next": "opencv_normalize0"
-        },
-        "opencv_normalize0": {
-            "props": {
-                "alpha": "123.675, 116.28, 103.53",
-                "beta": "58.395, 57.120, 57.375",
-                "dataType": "FLOAT32"
-            },
-            "factory": "mxpi_imagenormalize",
-            "next": "mxpi_imageresize1"
-        },
-        "mxpi_imageresize1": {
-            "props": {
-                "handleMethod": "opencv",
-                "resizeType": "Resizer_OnlyPadding",
-                "scaleValue": "513",
-                "paddingType": "Padding_RightDown",
-                "paddingHeight": "513",
-                "paddingWidth": "513",
-                "paddingColorB": "0",
-                "paddingColorG": "0",
-                "paddingColorR": "0"
-            },
-            "factory": "mxpi_imageresize",
-            "next": "mxpi_tensorinfer0"
-        },
         "mxpi_tensorinfer0": {
             "props": {
-                "dataSource": "mxpi_imageresize1",
+                "dataSource": "appsrc0",
                 "modelPath": "../data/model/deeplabv3.om"
             },
-            "factory": "mxpi_modelinfer",
+            "factory": "mxpi_tensorinfer",
             "next": "mxpi_semanticsegpostprocessor0"
         },
         "mxpi_semanticsegpostprocessor0": {
@@ -73,7 +33,7 @@
                 "blocksize": "409600"
             },
             "factory": "appsrc",
-            "next": "mxpi_imagedecoder0"
+            "next": "mxpi_tensorinfer0"
         },
         "appsink0": {
             "props": {

official/cv/deeplabv3/infer/sdk/main.py

@@ -20,8 +20,9 @@ import json
 import os
 import cv2
 import numpy as np
-from StreamManagerApi import MxDataInput
 from StreamManagerApi import StreamManagerApi
+from StreamManagerApi import InProtobufVector, MxProtobufIn, MxDataInput
+import MxpiDataType_pb2 as MxpiDataType
 from get_dataset_colormap import label_to_color_image
 
 PIPELINE_PATH = "../data/config/deeplabv3.pipeline"
@@ -58,16 +59,35 @@ def _init_stream(pipeline_path):
         return stream_manager_api
 
 
-def _do_infer(stream_manager_api, data_input):
+def _do_infer(stream_manager_api, data_input, origin_img):
     stream_name = b'segmentation'
-    unique_id = stream_manager_api.SendDataWithUniqueId(
-        stream_name, 0, data_input)
+    vision_list = MxpiDataType.MxpiVisionList()
+    vision_vec = vision_list.visionVec.add()
+    vision_vec.visionInfo.format = 0
+    vision_vec.visionInfo.width = origin_img.shape[1]
+    vision_vec.visionInfo.height = origin_img.shape[0]
+    vision_vec.visionInfo.widthAligned = origin_img.shape[1]
+    vision_vec.visionInfo.heightAligned = origin_img.shape[0]
+
+    vision_vec.visionData.memType = 0
+
+    vision_vec.visionData.dataStr = data_input.tobytes()
+    vision_vec.visionData.dataSize = len(data_input)
+
+    protobuf = MxProtobufIn()
+    protobuf.key = b"appsrc0"
+    protobuf.type = b'MxTools.MxpiVisionList'
+    protobuf.protobuf = vision_list.SerializeToString()
+    protobuf_vec = InProtobufVector()
+
+    protobuf_vec.push_back(protobuf)
+    unique_id = stream_manager_api.SendProtobuf(stream_name, 0, protobuf_vec)
+
     if unique_id < 0:
         raise RuntimeError("Failed to send data to stream.")
 
-    timeout = 3000
-    infer_result = stream_manager_api.GetResultWithUniqueId(
-        stream_name, unique_id, timeout)
+    infer_result = stream_manager_api.GetResult(stream_name, unique_id)
+
     if infer_result.errorCode != 0:
         raise RuntimeError(
             "GetResultWithUniqueId error, errorCode=%d, errorMsg=%s" % (
@@ -80,6 +100,37 @@ def _do_infer(stream_manager_api, data_input):
     return np.frombuffer(data_str, dtype=np.uint8).reshape(shape)
 
 
+def resize_long(img, long_size=513):
+    h, w, _ = img.shape
+    if h > w:
+        new_h = long_size
+        new_w = int(1.0 * long_size * w / h)
+    else:
+        new_w = long_size
+        new_h = int(1.0 * long_size * h / w)
+    imo = cv2.resize(img, (new_w, new_h))
+    return imo
+
+
+def pre_process(img_, crop_size=513):
+    image_mean = [103.53, 116.28, 123.675]
+    image_std = [57.375, 57.120, 58.395]
+    # resize
+    img_ = resize_long(img_, crop_size)
+
+    # mean, std
+    image_mean = np.array(image_mean)
+    image_std = np.array(image_std)
+    img_ = (img_ - image_mean) / image_std
+
+    # pad to crop_size
+    pad_h = crop_size - img_.shape[0]
+    pad_w = crop_size - img_.shape[1]
+    if pad_h > 0 or pad_w > 0:
+        img_ = cv2.copyMakeBorder(img_, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=0)
+
+    return img_
+
 def main():
     args = _parse_args()
     # init stream manager
@@ -97,9 +148,15 @@ def main():
             img_path = os.path.join(args.data_root, img_path)
             msk_path = os.path.join(args.data_root, msk_path)
             msk_ = cv2.imread(msk_path, cv2.IMREAD_GRAYSCALE)
-            with open(img_path, 'rb') as f:
-                data_input.data = f.read()
-            each_array = _do_infer(stream_manager_api, data_input)
+            img_ = cv2.imread(img_path)
+            re_img = img_
+
+            crop_size = 513
+            img_ = pre_process(img_, crop_size)
+
+            data_input = img_.astype("float32")
+
+            each_array = _do_infer(stream_manager_api, data_input, re_img)
 
             hist += _cal_hist(
                 msk_.flatten(), each_array.flatten(), args.num_classes)
@@ -116,6 +173,5 @@ def main():
 
     stream_manager_api.DestroyAllStreams()
 
-
 if __name__ == '__main__':
     main()