Before following the steps for deploying your own model it is required that you already have an existing project that you want to run your experiments or create a new project following the steps mentioned in the Create project section.
For this tutorial we will go through a simple classification task titled
Medical Image Classification: Hello World Deep Learning Tutorial which classifies Xrays as either chest or abdomen. This project is hosted on our platform.
Creating a new model instance
Open the MD.ai tool and select your project by going to the
My Projects section and clicking on the
Once inside, click on the
Models tab and select
New Model. Enter a name for your model and add a description if you want. Next you need to choose the scope for your model. The model scope determines how batching of inputs is carried out -
- If your model takes as input an entire study/exam, choose the scope as
- If your model takes as input an entire volumetric series, choose the scope as
- If your model takes as input a single image at a time, choose the scope as
There are a bunch of settings that you can choose for your model. Check out the Model Settings page for more details.
Creating a model assets zip file
Creating a model instance isn't enough, you need to further create a model version in order to run your models. In order to do that you will first need to create a zip file containing the saved model and the required MD.ai interface code. The following file structure needs to be strictly followed -
Create a root folder named
model. Inside this folder, store your pretrained model file. Along with this, create a sub folder named
.mdai and all the MD.ai interface related build files need to be placed inside this folder. Zip the root folder and your model is then ready to be deployed.
Creating a model version
Now that your model assets are wrapped in a zip file, you are ready to deploy your model. Go back to your project and under your previously created model instance, click on the "NEW LABEL CLASS" button. Here you will have to map the label indices produced by your model to the corresponding labels you have created on our platform. Do this for all target labels that you might have. For eg. in this case we will map the class "Abdomen" to 0 and "Chest" to 1 because that's how our model was trained to use the classes.
Click on the "NEW MODEL VERSION" button and perform the following steps -
- Upload Model - Upload the model zip file that you created before.
- Set Model Args - Enter the values of any arguments that your model might take. (optional)
- Set Runtime Config - Select the runtine configuration for your model such as whether you want the model to run on a CPU or a GPU (and which type of GPU)
- Build Model - Press on "START BUILD" and you model will be prepared for deployment.
The "Build status" will go from initally "Queued" to "Working" to finally "Succeeded", when the model has been succesfully deployed on the cloud.
Build status stuck in Working
It might take a while for the model to build depending on the size of your model file, so it's completely normal if the "Build Status" stays in the mode "Working" for some time.