KeyError: 'multiclass_nms3_0.tmp_2.lod'

[wjplove8]

部署fasterrcnn模型时，预测出错
预测的代码

import sys
import numpy as np
from paddle_serving_client import Client
from paddle_serving_app.reader import *
import cv2

preprocess = DetectionSequential([
DetectionFile2Image(),
DetectionNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225], True),
DetectionResize(
(800, 1333), True, interpolation=cv2.INTER_LINEAR),
DetectionTranspose((2,0,1)),
DetectionPadStride(128)
])

postprocess = RCNNPostprocess("label_list.txt", "output")
client = Client()

client.load_client_config("serving_client/serving_client_conf.prototxt")
client.connect(['127.0.0.1:9292'])

im, im_info = preprocess(sys.argv[1])
fetch_map = client.predict(
feed={
"image": im,
"im_shape": np.array(list(im.shape[1:])).reshape(-1),
"scale_factor": im_info['scale_factor'],
},
#fetch=["save_infer_model/scale_0.tmp_1"],#multiclass_nms3_0.tmp_2
fetch=["multiclass_nms3_0.tmp_2"],
batch=False)
fetch_map["image"] = sys.argv[1]
postprocess(fetch_map)

serving_client_conf.prototxt:
feed_var {
name: "im_shape"
alias_name: "im_shape"
is_lod_tensor: false
feed_type: 1
shape: 2
}
feed_var {
name: "image"
alias_name: "image"
is_lod_tensor: false
feed_type: 1
shape: 3
}
feed_var {
name: "scale_factor"
alias_name: "scale_factor"
is_lod_tensor: false
feed_type: 1
shape: 2
}
fetch_var {
name: "concat_12.tmp_0"
alias_name: "concat_12.tmp_0"
is_lod_tensor: false
fetch_type: 1
shape: 6
}
fetch_var {
name: "multiclass_nms3_0.tmp_2"
alias_name: "multiclass_nms3_0.tmp_2"
is_lod_tensor: false
fetch_type: 2
}

client报错信息
Traceback (most recent call last):
File "test_client.py", line 47, in
postprocess(fetch_map)
File "/usr/local/lib/python3.6/site-packages/paddle_serving_app/reader/image_reader.py", line 427, in call
self.clsid2catid)
File "/usr/local/lib/python3.6/site-packages/paddle_serving_app/reader/image_reader.py", line 344, in _get_bbox_result
lod = [fetch_map[fetch_name + '.lod']]
KeyError: 'multiclass_nms3_0.tmp_2.lod'

server log:
W0216 10:07:18.747469 2563 infer.cpp:287] Succ reload version engine: 18446744073709551615
I0216 10:07:18.747480 2563 manager.h:131] Finish reload 1 workflow(s)
I0216 10:07:28.747581 2563 server.cpp:151] Begin reload framework...
W0216 10:07:28.747633 2563 infer.cpp:287] Succ reload version engine: 18446744073709551615
I0216 10:07:28.747643 2563 manager.h:131] Finish reload 1 workflow(s)
I0216 10:07:38.747742 2563 server.cpp:151] Begin reload framework...
W0216 10:07:38.747807 2563 infer.cpp:287] Succ reload version engine: 18446744073709551615
I0216 10:07:38.747819 2563 manager.h:131] Finish reload 1 workflow(s)
I0216 10:07:44.068684 2572 general_model_service.pb.cc:3319] (logid=0) remote_side=[127.0.0.1:46034]
I0216 10:07:44.068742 2572 general_model_service.pb.cc:3320] (logid=0) local_side=[127.0.0.1:9292]
I0216 10:07:44.068750 2572 general_model_service.pb.cc:3321] (logid=0) service_name=[GeneralModelService]
I0216 10:07:44.068750 2572 op_repository.h:68] RAW: Succ regist op: GeneralReaderOp
I0216 10:07:44.068750 2572 op_repository.h:68] RAW: Succ regist op: GeneralInferOp
I0216 10:07:44.068750 2572 op_repository.h:68] RAW: Succ regist op: GeneralResponseOp
I0216 10:07:44.079159 2572 op.cpp:164] (logid=0) general_reader_0_time=[10296]
I0216 10:07:44.326128 2572 op.cpp:164] (logid=0) general_infer_0_time=[246925]
I0216 10:07:44.326174 2572 op.cpp:164] (logid=0) general_response_0_time=[8]
I0216 10:07:44.326189 2572 service.cpp:263] (logid=0) workflow total time: 257421
I0216 10:07:44.326215 2572 general_model_service.pb.cc:3343] [serving]logid=0,cost=257.534ms.
I0216 10:07:48.747907 2563 server.cpp:151] Begin reload framework...
W0216 10:07:48.747962 2563 infer.cpp:287] Succ reload version engine: 18446744073709551615
I0216 10:07:48.747973 2563 manager.h:131] Finish reload 1 workflow(s)
I0216 10:07:58.748062 2563 server.cpp:151] Begin reload framework...
W0216 10:07:58.748112 2563 infer.cpp:287] Succ reload version engine: 18446744073709551615
I0216 10:07:58.748121 2563 manager.h:131] Finish reload 1 workflow(s)
I0216 10:08:08.748234 2563 server.cpp:151] Begin reload framework...
W0216 10:08:08.748302 2563 infer.cpp:287] Succ reload version engine: 18446744073709551615
I0216 10:08:08.748315 2563 manager.h:131] Finish reload 1 workflow(s)
I0216 10:08:18.748411 2563 server.cpp:151] Begin reload framework...
处理
fetch_var {
name: "multiclass_nms3_0.tmp_2"
alias_name: "multiclass_nms3_0.tmp_2"
is_lod_tensor: false
fetch_type: 2
}
修改为
fetch_var {
name: "multiclass_nms3_0.tmp_2"
alias_name: "multiclass_nms3_0.tmp_2"
is_lod_tensor: true
fetch_type: 2
}
也不行

[github-actions]

Message that will be displayed on users' first issue

[felixhjh]

您好，我看这边您应该是使用自己训练的模型对嘛？postprocess针对不同模型处理上并不一样，这边还需要自己去做一下模型的后处理工作。Paddle Serving提供的postprocess方法只针对，Paddle Serving示例代码，并不具有通用性。因此还需要开发者参照示例的postprocess进行一定的修改。除此之外，并不是每个模型都会返回lod信息，因此报错。

[wjplove8]

您好，我看这边您应该是使用自己训练的模型对嘛？postprocess针对不同模型处理上并不一样，这边还需要自己去做一下模型的后处理工作。Paddle Serving提供的postprocess方法只针对，Paddle Serving示例代码，并不具有通用性。因此还需要开发者参照示例的postprocess进行一定的修改。除此之外，并不是每个模型都会返回lod信息，因此报错。

谢谢解决了，文档还是不够全面呀。希望能增加更多系列文档。

[ChiTaoGIS]

我也是使用的fasterrcnn，遇到同样的问题，
def _get_bbox_result(self, fetch_map, fetch_name, clsid2catid): result = {} is_bbox_normalized = False output = fetch_map[fetch_name] lod = [fetch_map[fetch_name + '.lod']] lengths = self._offset_to_lengths(lod) np_data = np.array(output)
问题出现在这里，不是很清楚后续如何处理，希望可以支持一下～！

[zhiqiangohuo]

除此之外，并不是每个模型都会返回lod信息，因此报错。

使用自己训练的yolov3

feed_var {
  name: "im_shape"
  alias_name: "im_shape"
  is_lod_tensor: false
  feed_type: 1
  shape: 2
}
feed_var {
  name: "image"
  alias_name: "image"
  is_lod_tensor: false
  feed_type: 1
  shape: 3
  shape: 608
  shape: 608
}
feed_var {
  name: "scale_factor"
  alias_name: "scale_factor"
  is_lod_tensor: false
  feed_type: 1
  shape: 2
}
fetch_var {
  name: "multiclass_nms3_0.tmp_0"
  alias_name: "multiclass_nms3_0.tmp_0"
  is_lod_tensor: true
  fetch_type: 1
  shape: -1
}
fetch_var {
  name: "multiclass_nms3_0.tmp_2"
  alias_name: "multiclass_nms3_0.tmp_2"
  is_lod_tensor: false
  fetch_type: 2
}

遇见了同样的问题，想问一下是什么原因造成的？

[wjplove8]

去年就遇到了此问题，当时搞定了，今年又遇到了，来查，看到答案想抽自己俩耳光，写一下具体解决办法。
需要重写一下paddle_serving_app.reader模块的RCNNPostprocess类

class RCNNPostprocess(object):
def init(self, label_file, output_dir, resize_shape=None):
self.output_dir = output_dir
self.label_file = label_file
self.label_list = []
self.resize_shape = resize_shape
with open(label_file) as fin:
for line in fin:
self.label_list.append(line.strip())
self.clsid2catid = {i: i for i in range(len(self.label_list))}
self.catid2name = {i: name for i, name in enumerate(self.label_list)}

def _offset_to_lengths(self, lod):
    offset = lod[0]
    #lengths = [offset[i + 1] - offset[i] for i in range(len(offset) - 1)]
    lengths = [6 for i in range(len(offset) - 1)]
    return [lengths]

def _bbox2out(self, results, clsid2catid, is_bbox_normalized=False):
    xywh_res = []
    for t in results:
        bboxes = t['bbox'][0]
        lengths = t['bbox'][1]#[0]
        if bboxes.shape == (1, 1) or bboxes is None:
            continue

        k = 0
        for i in range(len(lengths)):
            num = lengths[i]
            if k==len(bboxes):
                break
            for j in range(num):               
                dt = bboxes[k]
                clsid, score, xmin, ymin, xmax, ymax = dt.tolist()
                catid = (clsid2catid[int(clsid)])
                catname=(self.catid2name[int(catid)])
                if is_bbox_normalized:
                    xmin, ymin, xmax, ymax = \
                        self.clip_bbox([xmin, ymin, xmax, ymax])
                    w = xmax - xmin
                    h = ymax - ymin
                    im_shape = t['im_shape'].tolist()
                    im_height, im_width = int(im_shape[0]), int(im_shape[1])
                    xmin *= im_width
                    ymin *= im_height
                    w *= im_width
                    h *= im_height
                else:
                    w = xmax - xmin + 1
                    h = ymax - ymin + 1

                bbox = [xmin, ymin, w, h]
                coco_res = {
                    'category_id': catid,
                    'cat_name':catname,
                    'bbox': bbox,
                    'score': score
                }
                xywh_res.append(coco_res)
                k += 1
    return xywh_res

def _get_bbox_result(self, fetch_map, fetch_name, clsid2catid):
    result = {}
    is_bbox_normalized = False
    output = fetch_map[fetch_name]
    
    # lod = [fetch_map[fetch_name + '.lod']]
    lod = [fetch_map[fetch_name ]]
    
    lengths = self._offset_to_lengths(lod)
    np_data = np.array(output)
    result['bbox'] = (np_data,[6] * (len(np_data)))#np.ones(10) * 6      np.ones(len(np_data.shape))*6
    result['im_id'] = np.array([[0]])

    bbox_results = self._bbox2out([result], clsid2catid, is_bbox_normalized)
    return bbox_results

def color_map(self, num_classes):
    color_map = num_classes * [0, 0, 0]
    for i in range(0, num_classes):
        j = 0
        lab = i
        while lab:
            color_map[i * 3] |= (((lab >> 0) & 1) << (7 - j))
            color_map[i * 3 + 1] |= (((lab >> 1) & 1) << (7 - j))
            color_map[i * 3 + 2] |= (((lab >> 2) & 1) << (7 - j))
            j += 1
            lab >>= 3
    color_map = np.array(color_map).reshape(-1, 3)
    return color_map

def draw_bbox(self, image, catid2name, bboxes, threshold, color_list):
    """
    draw bbox on image
    """
    draw = ImageDraw.Draw(image)

    for dt in np.array(bboxes):
        catid, bbox, score = dt['category_id'], dt['bbox'], dt['score']
        if score < threshold:
            continue

        xmin, ymin, w, h = bbox
        xmax = xmin + w
        ymax = ymin + h

        img_w, img_h = image.size
        if self.resize_shape is not None:
            xmin = xmin * img_w / self.resize_shape[0]
            xmax = xmax * img_w / self.resize_shape[0]
            ymin = ymin * img_h / self.resize_shape[1]
            ymax = ymax * img_h / self.resize_shape[1]

        color = tuple(color_list[catid])

        # draw bbox
        draw.line(
            [(xmin, ymin), (xmin, ymax), (xmax, ymax), (xmax, ymin),
             (xmin, ymin)],
            width=2,
            fill=color)

        # draw label
        text = "{} {:.2f}".format(catid2name[catid], score)
        tw, th = draw.textsize(text)
        draw.rectangle(
            [(xmin + 1, ymin - th), (xmin + tw + 1, ymin)], fill=color)
        draw.text((xmin + 1, ymin - th), text, fill=(255, 255, 255))
    return image

def visualize(self, infer_img, bbox_results, catid2name, num_classes):
    image = Image.open(infer_img).convert('RGB')
    color_list = self.color_map(num_classes)
    image = self.draw_bbox(image, self.catid2name, bbox_results, 0.5,
                           color_list)
    image_path = os.path.split(infer_img)[-1]
    if not os.path.exists(self.output_dir):
        os.makedirs(self.output_dir)
    out_path = os.path.join(self.output_dir, image_path)
    image.save(out_path, quality=95)

def __call__(self, image_with_bbox, visualize=False):
    fetch_name = ""
    for key in image_with_bbox:
        if key == "image":
            continue
        if ".lod" in key or "im_shape" in key:
            continue
        fetch_name = key
    bbox_result = self._get_bbox_result(image_with_bbox, fetch_name,
                                        self.clsid2catid)
    if os.path.isdir(self.output_dir) is False:
        os.mkdir(self.output_dir)
    if visualize is False:
        return bbox_result
    self.visualize(image_with_bbox["image"], bbox_result, self.catid2name,
                   len(self.label_list))
    if os.path.isdir(self.output_dir) is False:
        os.mkdir(self.output_dir)
    bbox_file = os.path.join(self.output_dir, 'bbox.json')
    with open(bbox_file, 'w') as f:
        json.dump(bbox_result, f, indent=4)
    return bbox_result

def __repr__(self):
    return self.__class__.__name__ + "label_file: {1}, output_dir: {2}".format(
        self.label_file, self.output_dir)

我也是使用的fasterrcnn，遇到同样的问题， def _get_bbox_result(self, fetch_map, fetch_name, clsid2catid): result = {} is_bbox_normalized = False output = fetch_map[fetch_name] lod = [fetch_map[fetch_name + '.lod']] lengths = self._offset_to_lengths(lod) np_data = np.array(output) 问题出现在这里，不是很清楚后续如何处理，希望可以支持一下～！

[wjplove8]

去年就遇到了此问题，当时搞定了，今年又遇到了，来查，看到答案想抽自己俩耳光，写一下具体解决办法。 需要重写一下paddle_serving_app.reader模块的RCNNPostprocess类

class RCNNPostprocess(object): def init(self, label_file, output_dir, resize_shape=None): self.output_dir = output_dir self.label_file = label_file self.label_list = [] self.resize_shape = resize_shape with open(label_file) as fin: for line in fin: self.label_list.append(line.strip()) self.clsid2catid = {i: i for i in range(len(self.label_list))} self.catid2name = {i: name for i, name in enumerate(self.label_list)}

def _offset_to_lengths(self, lod):
    offset = lod[0]
    #lengths = [offset[i + 1] - offset[i] for i in range(len(offset) - 1)]
    lengths = [6 for i in range(len(offset) - 1)]
    return [lengths]

def _bbox2out(self, results, clsid2catid, is_bbox_normalized=False):
    xywh_res = []
    for t in results:
        bboxes = t['bbox'][0]
        lengths = t['bbox'][1]#[0]
        if bboxes.shape == (1, 1) or bboxes is None:
            continue

        k = 0
        for i in range(len(lengths)):
            num = lengths[i]
            if k==len(bboxes):
                break
            for j in range(num):               
                dt = bboxes[k]
                clsid, score, xmin, ymin, xmax, ymax = dt.tolist()
                catid = (clsid2catid[int(clsid)])
                catname=(self.catid2name[int(catid)])
                if is_bbox_normalized:
                    xmin, ymin, xmax, ymax = \
                        self.clip_bbox([xmin, ymin, xmax, ymax])
                    w = xmax - xmin
                    h = ymax - ymin
                    im_shape = t['im_shape'].tolist()
                    im_height, im_width = int(im_shape[0]), int(im_shape[1])
                    xmin *= im_width
                    ymin *= im_height
                    w *= im_width
                    h *= im_height
                else:
                    w = xmax - xmin + 1
                    h = ymax - ymin + 1

                bbox = [xmin, ymin, w, h]
                coco_res = {
                    'category_id': catid,
                    'cat_name':catname,
                    'bbox': bbox,
                    'score': score
                }
                xywh_res.append(coco_res)
                k += 1
    return xywh_res

def _get_bbox_result(self, fetch_map, fetch_name, clsid2catid):
    result = {}
    is_bbox_normalized = False
    output = fetch_map[fetch_name]
    
    # lod = [fetch_map[fetch_name + '.lod']]
    lod = [fetch_map[fetch_name ]]
    
    lengths = self._offset_to_lengths(lod)
    np_data = np.array(output)
    result['bbox'] = (np_data,[6] * (len(np_data)))#np.ones(10) * 6      np.ones(len(np_data.shape))*6
    result['im_id'] = np.array([[0]])

    bbox_results = self._bbox2out([result], clsid2catid, is_bbox_normalized)
    return bbox_results

def color_map(self, num_classes):
    color_map = num_classes * [0, 0, 0]
    for i in range(0, num_classes):
        j = 0
        lab = i
        while lab:
            color_map[i * 3] |= (((lab >> 0) & 1) << (7 - j))
            color_map[i * 3 + 1] |= (((lab >> 1) & 1) << (7 - j))
            color_map[i * 3 + 2] |= (((lab >> 2) & 1) << (7 - j))
            j += 1
            lab >>= 3
    color_map = np.array(color_map).reshape(-1, 3)
    return color_map

def draw_bbox(self, image, catid2name, bboxes, threshold, color_list):
    """
    draw bbox on image
    """
    draw = ImageDraw.Draw(image)

    for dt in np.array(bboxes):
        catid, bbox, score = dt['category_id'], dt['bbox'], dt['score']
        if score < threshold:
            continue

        xmin, ymin, w, h = bbox
        xmax = xmin + w
        ymax = ymin + h

        img_w, img_h = image.size
        if self.resize_shape is not None:
            xmin = xmin * img_w / self.resize_shape[0]
            xmax = xmax * img_w / self.resize_shape[0]
            ymin = ymin * img_h / self.resize_shape[1]
            ymax = ymax * img_h / self.resize_shape[1]

        color = tuple(color_list[catid])

        # draw bbox
        draw.line(
            [(xmin, ymin), (xmin, ymax), (xmax, ymax), (xmax, ymin),
             (xmin, ymin)],
            width=2,
            fill=color)

        # draw label
        text = "{} {:.2f}".format(catid2name[catid], score)
        tw, th = draw.textsize(text)
        draw.rectangle(
            [(xmin + 1, ymin - th), (xmin + tw + 1, ymin)], fill=color)
        draw.text((xmin + 1, ymin - th), text, fill=(255, 255, 255))
    return image

def visualize(self, infer_img, bbox_results, catid2name, num_classes):
    image = Image.open(infer_img).convert('RGB')
    color_list = self.color_map(num_classes)
    image = self.draw_bbox(image, self.catid2name, bbox_results, 0.5,
                           color_list)
    image_path = os.path.split(infer_img)[-1]
    if not os.path.exists(self.output_dir):
        os.makedirs(self.output_dir)
    out_path = os.path.join(self.output_dir, image_path)
    image.save(out_path, quality=95)

def __call__(self, image_with_bbox, visualize=False):
    fetch_name = ""
    for key in image_with_bbox:
        if key == "image":
            continue
        if ".lod" in key or "im_shape" in key:
            continue
        fetch_name = key
    bbox_result = self._get_bbox_result(image_with_bbox, fetch_name,
                                        self.clsid2catid)
    if os.path.isdir(self.output_dir) is False:
        os.mkdir(self.output_dir)
    if visualize is False:
        return bbox_result
    self.visualize(image_with_bbox["image"], bbox_result, self.catid2name,
                   len(self.label_list))
    if os.path.isdir(self.output_dir) is False:
        os.mkdir(self.output_dir)
    bbox_file = os.path.join(self.output_dir, 'bbox.json')
    with open(bbox_file, 'w') as f:
        json.dump(bbox_result, f, indent=4)
    return bbox_result

def __repr__(self):
    return self.__class__.__name__ + "label_file: {1}, output_dir: {2}".format(
        self.label_file, self.output_dir)

我也是使用的fasterrcnn，遇到同样的问题， def _get_bbox_result(self, fetch_map, fetch_name, clsid2catid): result = {} is_bbox_normalized = False output = fetch_map[fetch_name] lod = [fetch_map[fetch_name + '.lod']] lengths = self._offset_to_lengths(lod) np_data = np.array(output) 问题出现在这里，不是很清楚后续如何处理，希望可以支持一下～！

其实就是官方写的后处理计算不能兼容lod数据和定长数据两种情况。我没有找到带lod数据的模型，不能完善代码，这只是应急处理，希望官方后续更新一下。