DeepDetect Docker Usage

DeepDetect Docker images run the dede REST server. Use them when you want a long-running HTTP service, containerized serving, or a reproducible deployment environment. For local in-process work and the deepdetect CLI, use the Python wheel described in ../bindings/python/README.md.

The current recommended server backends are:

• torch for training and inference.
• tensorrt for optimized inference from exported or compatible models.

The retired caffe, caffe2, tf, and tensorflow backends are rejected by the server. Do not use old Docker examples that create services with "mllib":"caffe".

Images

Official images are published at docker.jolibrain.com:

docker pull docker.jolibrain.com/deepdetect_cpu
docker pull docker.jolibrain.com/deepdetect_gpu
docker pull docker.jolibrain.com/deepdetect_gpu_tensorrt

Use deepdetect_cpu for CPU-only serving, deepdetect_gpu for CUDA Torch serving, and deepdetect_gpu_tensorrt when you need TensorRT support.

To inspect the registry:

curl -X GET https://docker.jolibrain.com/v2/_catalog
curl -X GET https://docker.jolibrain.com/v2/deepdetect_cpu/tags/list

Run The Server

Start the CPU image:

docker run -d \
  --name deepdetect \
  -p 8080:8080 \
  docker.jolibrain.com/deepdetect_cpu

Check that the server is responding:

curl http://localhost:8080/info

The response should contain a status.code of 200 and a services list.

Stop and remove the container:

docker stop deepdetect
docker rm deepdetect

Run With GPUs

GPU images require Docker with NVIDIA Container Toolkit support. With recent Docker versions, use --gpus:

docker run -d \
  --name deepdetect \
  --gpus all \
  -p 8080:8080 \
  docker.jolibrain.com/deepdetect_gpu

For TensorRT:

docker run -d \
  --name deepdetect-trt \
  --gpus all \
  -p 8080:8080 \
  docker.jolibrain.com/deepdetect_gpu_tensorrt

If your Docker installation still uses the old nvidia-docker wrapper, the equivalent form is:

nvidia-docker run -d \
  --name deepdetect \
  -p 8080:8080 \
  docker.jolibrain.com/deepdetect_gpu

Mount Models And Data

Mount host directories for model repositories and datasets. The container runs as user dd, so the mounted paths must be readable by that user and writable when training or saving models.

mkdir -p models data

docker run -d \
  --name deepdetect \
  -p 8080:8080 \
  -v "$PWD/models:/models" \
  -v "$PWD/data:/data" \
  docker.jolibrain.com/deepdetect_cpu

Inside API calls, use the container paths (/models/..., /data/...), not the host paths.

Create A Torch Service

Create services from model repositories mounted into the container. This example assumes /models/resnet18 contains a compatible Torch model repository:

curl -X PUT "http://localhost:8080/services/imageserv" \
  -H "Content-Type: application/json" \
  -d '{
    "mllib": "torch",
    "description": "Torch image classification service",
    "type": "supervised",
    "model": {
      "repository": "/models/resnet18"
    },
    "parameters": {
      "input": {
        "connector": "image",
        "width": 224,
        "height": 224,
        "rgb": true,
        "scale": 0.0039
      },
      "mllib": {
        "template": "resnet18",
        "nclasses": 2,
        "gpu": false
      },
      "output": {}
    }
  }'

Run prediction:

curl -X POST "http://localhost:8080/predict" \
  -H "Content-Type: application/json" \
  -d '{
    "service": "imageserv",
    "parameters": {
      "input": {
        "width": 224,
        "height": 224
      },
      "output": {
        "best": 3
      }
    },
    "data": ["/data/image.jpg"]
  }'

For GPU inference, run a GPU image and set "gpu": true in the service parameters.mllib object.

Create A TensorRT Service

Use the TensorRT image with an exported or compatible model repository:

curl -X PUT "http://localhost:8080/services/yolox-trt" \
  -H "Content-Type: application/json" \
  -d '{
    "mllib": "tensorrt",
    "description": "TensorRT YOLOX inference service",
    "type": "supervised",
    "model": {
      "repository": "/models/yolox-trt"
    },
    "parameters": {
      "input": {
        "connector": "image",
        "width": 640,
        "height": 640,
        "rgb": true
      },
      "mllib": {
        "template": "yolox",
        "nclasses": 80,
        "datatype": "fp16",
        "maxBatchSize": 8,
        "gpuid": 0
      },
      "output": {}
    }
  }'

TensorRT can read an existing engine from the model repository or build and write one depending on readEngine, writeEngine, and maxBatchSize. See the TensorRT parameters in api.md.

Logs And Debugging

Follow server logs:

docker logs -f deepdetect

List running containers:

docker ps

Open a shell in the container:

docker exec -it deepdetect bash

Check loaded services:

curl "http://localhost:8080/info?status=true"

Swagger UI

With a running server, Swagger UI is served at:

http://localhost:8080/swagger/ui

If the server is behind a path prefix such as /api/deepdetect, pass the prefix as an extra server argument after the image name:

docker run -d \
  --name deepdetect \
  -p 8080:8080 \
  docker.jolibrain.com/deepdetect_cpu \
  -swagger_api_prefix api/deepdetect/

Build Images Locally

Build from the repository root, not from the docker/ directory:

export DOCKER_BUILDKIT=1

docker build \
  -t jolibrain/deepdetect_cpu \
  -f docker/cpu.Dockerfile \
  .

Build the GPU image:

export DOCKER_BUILDKIT=1

docker build \
  -t jolibrain/deepdetect_gpu \
  -f docker/gpu.Dockerfile \
  .

Build the TensorRT image:

export DOCKER_BUILDKIT=1

docker build \
  -t jolibrain/deepdetect_gpu_tensorrt \
  -f docker/gpu_tensorrt.Dockerfile \
  .

Useful build arguments:

• DEEPDETECT_DEFAULT_MODELS=false: skip downloading bundled legacy default models.
• USE_PREBUILT_TORCH=ON: use official prebuilt PyTorch/LibTorch packages.
• DEEPDETECT_GPU_VARIANT=legacy61: build a GPU image for older compute capabilities.
• PYTORCH_CUDA_INDEX=cu126: select the matching PyTorch CUDA wheel index for legacy GPU builds.
• DD_CUDA_VERSION and DD_CUDA_MAJOR_MINOR: override CUDA image versions for GPU builds.

More build details are in ../docker/README.md.