【AI模型部署】maskrcnn在tfserver部署以及调用时遇到问题：Servable not found for request “xx”、‘incompatible_shape_error‘

主要参考：

• 【maskrcnn版本】https://github.com/matterport/Mask_RCNN
• 【部署参考】https://github.com/bendangnuksung/mrcnn_serving_ready
• 【其他参考】https://github.com/huyhoang17/matterport-maskrcnn-with-tensorflow-serving

生成savedmodel

根据【部署参考】生成，自定义配置在附录中，

所用cooc模型地址： https://pan.baidu.com/s/14QeaR0shNdOaSvWNJnD7YQ 提取码: ibwg 

启动tf_server的docker

docker pull tensorflow/serving:1.15.0-gpu

docker pull tensorflow/serving:1.14.0-gpu

启动例子

根据自己参数修改

docker run --runtime=nvidia \
  -p 8500:8500 -p 8501:8501 \
  --name mask \
  -e MODEL_NAME=mask \
  -v /home/xxx/models/serving_model:/models/mask \
  tensorflow/serving:1.15.0-gpu \
  --per_process_gpu_memory_fraction=0.3 

参数解析：

 -v 将 savedmodel 本地位置挂载到容器下 /models/
 -e MODEL_NAME xx 要和挂载  /models/xx 名字相同
 --name 表示容器的名字
 --runtime=nvidia docker可调用gpu (安装 nvidia-docker2)
  tensorflow/serving:1.15.0-gpu  启动镜像的TAG 顺序很重要 
 --per_process_gpu_memory_fraction=0.3  (显存使用使用限制在本机显存*0.3)

遇到问题1

AbortionError(code=StatusCode.NOT_FOUND, details="Servable not found for request: Latest(“’mask”)

 这个问题是使用grpc与tfserver通信时，MODEL_NAME与request.model_spec.name不一致问题

我当时多打了一个 ‘ 找了很久

【参考】https://github.com/tensorflow/serving/issues/441

遇到问题2

grpc_message":"NodeDef mentions attr 'incompatible_shape_error'

 Check whether your GraphDef-interpreting binary is up to date with your GraphDef-generating binary

[[roi_align_classifier/Equal_3]]","grpc_status":3}"

【解决方法】

附录

【配置】https://github.com/bendangnuksung/mrcnn_serving_ready/blob/master/user_config.py

所用mask模型地址： https://pan.baidu.com/s/14QeaR0shNdOaSvWNJnD7YQ 提取码: ibwg

import os
# User Define parameters

# Make it True if you want to use the provided coco weights
is_coco = False

# keras model file path
H5_WEIGHT_PATH = './models/mask_rcnn_coco.h5'
MODEL_DIR = os.path.dirname(H5_WEIGHT_PATH)  # 去掉文件名，返回目录

# Path where the Frozen PB will be save
PATH_TO_SAVE_FROZEN_PB = './models/frozen_model/'

# Name for the Frozen PB name
FROZEN_NAME = 'mask_frozen_graph.pb'

# PATH where to save serving model
PATH_TO_SAVE_TENSORFLOW_SERVING_MODEL = './models/serving_model'

# Version of the serving model
VERSION_NUMBER = 1

# Number of classes that you have trained your model  
NUMBER_OF_CLASSES = 80  # coco is 80