Inference

This section shows how to run inference on AWS Deep Learning Containers for Amazon Elastic Compute Cloud using PyTorch, and TensorFlow.

PyTorch Inference

Deep Learning Containers with PyTorch version 1.6 and later use TorchServe for inference calls. Deep Learning Containers with PyTorch version 1.5 and earlier use multi-model-server for inference calls.

PyTorch 1.6 and later

To run inference with PyTorch, this example uses a model pretrained on Imagenet from a public S3 bucket. Inference is served using TorchServe. For more information, see this blog on Deploying PyTorch inference with TorchServe.

For CPU instances:

$ docker run -itd --name torchserve -p 80:8080  -p 8081:8081 <your container image id> \
torchserve --start --ts-config /home/model-server/config.properties \
--models pytorch-densenet=https://torchserve.s3.amazonaws.com/mar_files/densenet161.mar

For GPU instances

$ nvidia-docker run -itd --name torchserve -p 80:8080  -p 8081:8081 <your container image id> \
torchserve --start --ts-config /home/model-server/config.properties \
--models pytorch-densenet=https://torchserve.s3.amazonaws.com/mar_files/densenet161.mar

If you have docker-ce version 19.03 or later, you can use the --gpus flag when you start Docker.

The configuration file is included in the container.

With your server started, you can now run inference from a different window by using the following.

$ curl -O https://s3.amazonaws.com/model-server/inputs/flower.jpg
curl -X POST http://127.0.0.1:80/predictions/pytorch-densenet -T flower.jpg

After you are done using your container, you can remove it using the following.

$ docker rm -f torchserve

PyTorch 1.5 and earlier

To run inference with PyTorch, this example uses a model pretrained on Imagenet from a public S3 bucket. Inference is served using multi-model-server, which can support any framework as the backend. For more information, see multi-model-server.

For CPU instances:

$ docker run -itd --name mms -p 80:8080  -p 8081:8081 <your container image id> \
multi-model-server --start --mms-config /home/model-server/config.properties \
--models densenet=https://dlc-samples.s3.amazonaws.com/pytorch/multi-model-server/densenet/densenet.mar

For GPU instances

$ nvidia-docker run -itd --name mms -p 80:8080  -p 8081:8081 <your container image id> \
multi-model-server --start --mms-config /home/model-server/config.properties \
--models densenet=https://dlc-samples.s3.amazonaws.com/pytorch/multi-model-server/densenet/densenet.mar

If you have docker-ce version 19.03 or later, you can use the --gpus flag when you start Docker.

The configuration file is included in the container.

With your server started, you can now run inference from a different window by using the following.

$ curl -O https://s3.amazonaws.com/model-server/inputs/flower.jpg
curl -X POST http://127.0.0.1/predictions/densenet -T flower.jpg

After you are done using your container, you can remove it using the following.

$ docker rm -f mms

TensorFlow Inference

To demonstrate how to use Deep Learning Containers for inference, this example uses a simple half plus two model with TensorFlow 2 Serving. We recommend using the Deep Learning Base AMI for TensorFlow 2. After you log into your instance run the following.

$ git clone -b r2.0 https://github.com/tensorflow/serving.git
$ cd serving

Use the commands here to start TensorFlow Serving with the Deep Learning Containers for this model. Unlike the Deep Learning Containers for training, model serving starts immediately upon running the container and runs as a background process.

Next, run inference with the Deep Learning Containers.

$ curl -d '{"instances": [1.0, 2.0, 5.0]}' -X POST http://127.0.0.1:8501/v1/models/saved_model_half_plus_two:predict

The output is similar to the following.

{
    "predictions": [2.5, 3.0, 4.5
    ]
}

$ docker attach <your docker container name>

Next steps

To learn about using custom entrypoints with Deep Learning Containers on Amazon ECS, see Custom Entrypoints.