Classification algorithm based on TensorFlow 2.X.
-
The model is saved in savedmodel format for easy deployment
-
Use tensorflow_serving+docker to deploy CPU/GPU version (multi-model)
-
Including docker start command
-
Use the restful/grpc method to send prediction requests, one of which is one at a time in the restful method. grpc is sent in batches, and batch size adaptation does not need to be set.
-
Has preheated files
-
The preheat file is generated to reduce the delay of starting the model.
-
tf_sevring_grpc.py(grpc)
-
tf_sevring_restful.py----------------------(restful)
-
tf_sevring_warmup.py-----------------------(preheated files)
-
train.py-----------------------------------(training)
-
test.py------------------------------------(local test)
-
data.py,dadaset.py-------------------------(utils)
-
write_tfrecord.py,read_tfrecord.py---------(tfrecord)
-
json_data.py-------------------------------(label)
-
config.py----------------------------------(set config)
-
evaluate.py--------------------------------(eva)
-
test_eva.py--------------------------------(...)
- nets---------------------------------------(vgg,resnet,...)
- dataset------------------------------------(dataset)
- raw----------------------------------------(Store the original data set)
- test---------------------------------------(Store the picture to be recognized)
- saved_model--------------------------------(save model)
- class.json---------------------------------(label)
- tf_docker.txt------------------------------(docker command)
- Target Detection
基于tf2.x实现的多种分类算法,可进行多模型协同工作,进行预测,同时包括模型部署需要的文件,实现一个人完成单个分类项目。
header |
---|
模型保存为savedmodel格式方便部署 |
采用tensorflow_serving+docker方式部署CPU/GPU版(多模型) |
包括docker启动命令 |
采用restful/grpc方式发送预测请求,其中restful方式一次一张。grpc按batch发送,batch大小自适应不需要设置。 |
有预热文件 |
生成预热文件用于减少启动模型的延时。 |
存在的问题:restful方式耗时长,grpc方式图片转数组耗时较长。 json文件生成不稳定 |
文件功能说明: |
tf_sevring_grpc.py-------------------------grpc方式 |
tf_sevring_restful.py----------------------restful方式 |
tf_sevring_warmup.py-----------------------预热文件生成 |
train.py-----------------------------------模型训练 |
test.py------------------------------------本地循环加载模型预测 |
data.py,dadaset.py-------------------------一些函数 |
write_tfrecord.py,read_tfrecord.py---------数据集操作 |
json_data.py-------------------------------标签文件生成 |
config.py----------------------------------参数设置 |
evaluate.py--------------------------------测试集测试 |
test_eva.py--------------------------------单一模型针对可能性最高,出现次数最多两种方式的预测 |
blender------------------------------------存放blender脚本相关数据 |
nets---------------------------------------存放模型网络 |
dataset------------------------------------存放分割后的数据集和tfrecord文件 |
raw----------------------------------------存放原始数据集 |
test---------------------------------------存放待识别图片 |
saved_model--------------------------------存放训练后的模型 |
class.json---------------------------------标签信息 |
tf_docker.txt------------------------------启动docker命令及一些参数信息 |