内容简介:If you are here it means you are interested in torchserve , the new tool to properly put models into production. So, without further due, let’s present today’s roadmap:
Deploy models and create custom handlers in Torchserve :rocket:
let’s put a model into production
Jun 11 ·6min read
All the code used in this article is here
If you are here it means you are interested in torchserve , the new tool to properly put models into production. So, without further due, let’s present today’s roadmap:
- Installation with Docker
- Export your model
- Define a handler
- Serve our model
To showcase torchserver, we will serve a fully trained ResNet34 to perform image classification.
Installation with Docker
Official doc here
The best way to install torchserve is with docker. You just need to pull the image.
You can use the following command to save the latest image.
docker pull pytorch/torchserve:latest
All the tags are available here
More about docker and torchserve here
Handlers
Official doc here
Handlers are the ones responsible to make a prediction using your model from one or more HTTP requests.
Default handlers
Torchserve supports the following default handlers
image_classifier object_detector text_classifier image_segmenter
But keep in mind that none of them supports batching requests!
Custom handlers
torchserve exposes a rich interface to do almost everything you want. An Handler
is just a class that must have three functions
- preprocess
- inference
- postprocess
You can create your own class or just subclass BaseHandler
. The main advantage of subclasssing BaseHandler
is to have the model loaded accessible at self.model
. The following snippet shows how to subclass BaseHandler
Going back to our image classification example. We need to
- get the images from each request and preprocess them
- get the prediction from the model
- send back a response
Preprocess
The .preprocess
function takes an array of requests. Assuming we are sending an image to the server, the serialized image can be accessed from the data
or body
field of the request. Thus, we can just iterate over all requests and preprocess individually each image. The full code is shown below.
self.transform
is our preprocess transformation, nothing fancy. This is a classic preprocessing step for models trained on ImageNet.
After we have preprocessed each image in each request we concatenate them to create a pytorch Tensor.
Inference
This step is very easy, we get the tensor from the .preprocess
function and we extract the prediction for each image.
Postprocess
Now we have our predictions for each image, we need to return something to the client. Torchserve always expects an array to be returned.
BaseHandler
also automatically opens a .json
file with the mapping index -> label
(we are going to see it later how to provide such file) and store it at self.mapping
. We can return an array of dictionaries with the label
and index
class for each prediction
Wrapping everything together, our glorious handler looks like
Since all the handling logic encapsulated in a class, you can easily unit test it!
Export your model
Official doc here
Torchserve expects a .mar
file to be provided. In a nutshell, the file is just your model and all the dependencies packed together. To create one need to first export our trained model.
Export the model
There are three ways to export your model for torchserve. The best way that I have found so far is to trace
the model and store the results. By doing so we do not need to add any additional files to torchserve.
Let’s see an example, we are going to deploy a fully trained ResNet34 model.
In order, we:
torch.jit.trace
Create the .mar file
Official doc here
You need to install torch-model-archiver
git clone https://github.com/pytorch/serve.git cd serve/model-archiver pip install .
Then, we are ready to create the .mar
file by using the following command
torch-model-archiver --model-name resnet34 \--version 1.0 \--serialized-file resnet34.pt \--extra-files ./index_to_name.json,./MyHandler.py \--handler my_handler.py \--export-path model-store -f
In order. The variable --model-name
defines the final name of our model. This is very important since it will be the namespace of the endpoint that will be responsible for its predictions. You can also specify a --version
. --serialized-file
points to the stored .pt
model we created before. --handler
is a python file where we call our custom handler. In general, it always looks like this:
It exposes a handle
function from which we call the methods in the custom handler. You can use the default names to use the default handled (e.g. --handler image_classifier
).
In --extra-files
you need to pass the path to all the files your handlers are using. In our case, we have to add the path to the .json
file with all the human-readable labels names and MyHandler.py
file in which we have the class definition for MyHandler.
One minor thing, if you pass an index_to_name.json
file, it will be automatically loaded into the handler and be accessible at self.mapping
.
--export-path
is where the .mar
file will be stored, I also added the -f
to overwrite everything in it.
If everything went smooth, you should see resnet34.mar
stored into ./model-store
.
Serve our model
This is an easy step, we can run the torchserve docker container with all the required parameters
docker run --rm -it \-p 3000:8080 -p 3001:8081 \-v $(pwd)/model-store:/home/model-server/model-store pytorch/torchserve:0.1-cpu \torchserve --start --model-store model-store --models resnet34=resnet34.mar
I am binding the container port 8080 and 8081 to 3000 and 3001 respectively (8080/8081 were already in used in my machine). Then, I am creating a volume from ./model-store
(where we stored the .mar
file) to the container default model-store
folder. Lastly, I am calling torchserve
by padding the model-store
path and a list of key-value pairs in which we specify the model name for each .mar
file.
At this point, torchserve has one endpoint /predictions/resnet34
to which we can get a prediction by sending an image. This can be done using curl
curl -X POST http://127.0.0.1:3000/predictions/resnet34 -T inputs/kitten.jpg
The response
{ "label": "tiger_cat", "index": 282 }
It worked!
Summary
To recap, in this article we have covered:
- torchserve installation with docker
- default and custom handlers
- model archive generation
- serving the final model with docker
All the code is here
If you like this article and pytorch, you may also be interested in these my other articles
Thank you for reading.
Francesco
以上就是本文的全部内容,希望本文的内容对大家的学习或者工作能带来一定的帮助,也希望大家多多支持 码农网
猜你喜欢:本站部分资源来源于网络,本站转载出于传递更多信息之目的,版权归原作者或者来源机构所有,如转载稿涉及版权问题,请联系我们。
JAVA核心技术(卷1)
Cay S. Horstmann、Gary Cornell / 杜永萍、邝劲筠、叶乃文 / 机械工业出版社 / 2008-6 / 98.00元
《JAVA核心技术(卷1):基础知识(原书第8版)》是《Java核心技术》的最新版,《Java核心技术》出版以来一直畅销不衰,深受读者青睐,每个新版本都尽可能快地跟上Java开发工具箱发展的步伐,而且每一版都重新改写了的部分内容,以便适应Java的最新特性。本版也不例外,它反遇了Java SE6的新特性。全书共14章,包括Java基本的程序结构、对象与类、继承、接口与内部类、图形程序设计、事件处理......一起来看看 《JAVA核心技术(卷1)》 这本书的介绍吧!