Model Config

Model config is the configuration to describe an inference workflow in AIAA. It is divided into the following sections (bold means required):

  1. Basic Information:

    • version: The version of AIAA, should be 3 or 2.

    • type: Following are the types of model currently supported in AIAA

      • segmentation

      • annotation

      • classification

      • deepgrow

      • others

      • pipeline

    • labels: The organs/subject of interest of this model.

    • description: A text to describe this model.


    Note that for compatibility reason, the type of DExtr3D model is “annotation”.

  2. Pre-Transforms:

    What transforms to do before the data is passed into the model. Each transform will have

    • name: Name of the transform. It can be a short name or full qualified name with class path (in case of bring your own transforms)

    • args: List of arguments passed during creation of transform object



    { "name": "LoadNifti", "args": { "fields": "image", "as_closest_canonical": "false" } }


    Transform should be either a callable object or derived from Clara Train transforms. Please refer to Bring your own Transforms for details.

  3. Inference:

    • image: The field to input images/volumes. (default: "image")

    • image_format: The shape format of the input image. This is required only if you don’t specify shape format in your pre-transforms.

    • name: Name of Inference. Built-Ins are:

      • TRTISInference: Triton/TRTIS Inference (strongly recommended) which supports simple and scanning window inference logic

      • TFInference: Native TensorFlow Inference which supports simple and scanning window inference logic

    • args: List of arguments passed during creation of Inference object

    • node_mapping: Map from model tensor names to field. (default: {"NV_MODEL_INPUT": "image", "NV_MODEL_OUTPUT": "model"})

    • additional_info: Additional info which will be passed to client in model info

    • trtis (required for Triton/TRTIS engine):

      This section contains Triton model configuration. Please refer to Triton Model Configuration to add any Triton specific settings

    • tf (TF model node mapping, is required in native mode):

      • input_nodes: Map from field to model input tensor names. Usually "image": "NV_MODEL_INPUT".

      • output_nodes: Map from field to model output tensor names. Usually "model": "NV_MODEL_OUTPUT".


    In case of custom Inference, you can provide fully qualified name with class path.

    Such custom Inference should be either a callable object or implements AIAA Inference interface.

  4. Post-Transforms:

    What transforms to do after predictions. Semantics are same as Pre-Transforms.

  5. Writer:

    What writer to use to write results out.


For PyTorch models with Triton, the platform of the Triton model config should be set to “pytorch_libtorch”.

For Triton related attributes can refer to Triton documentation.


If you want to upload TensorFlow ckpt model, you need to specify the tf part.


The config_aiaa.json in v1.x is not compatible with current release. Check Converting from previous TLT to modify your config file.

The following is an example of config_aiaa.json for the model clara_ct_seg_spleen_amp


{ "version": "3", "type": "segmentation", "labels": [ "spleen" ], "description": "A pre-trained model for volumetric (3D) segmentation of the spleen from CT image", "pre_transforms": [ { "name": "LoadNifti", "args": { "fields": "image" } }, { "name": "ConvertToChannelsFirst", "args": { "fields": "image" } }, { "name": "ScaleByResolution", "args": { "fields": "image", "target_resolution": [1.0, 1.0, 1.0] } }, { "name": "ScaleIntensityRange", "args": { "fields": "image", "a_min": -57, "a_max": 164, "b_min": 0.0, "b_max": 1.0, "clip": true } } ], "inference": { "image": "image", "name": "TRTISInference", "args": { "scanning_window": true, "batch_size": 1, "roi": [160, 160, 160] }, "trtis": { "platform": "tensorflow_graphdef", "max_batch_size": 1, "input": [ { "name": "NV_MODEL_INPUT", "data_type": "TYPE_FP32", "dims": [1, 160, 160, 160] } ], "output": [ { "name": "NV_MODEL_OUTPUT", "data_type": "TYPE_FP32", "dims": [2, 160, 160, 160] } ], "instance_group": [ { "count": 1, "kind": "KIND_AUTO" } ] }, "tf": { "input_nodes": { "image": "NV_MODEL_INPUT" }, "output_nodes": { "model": "NV_MODEL_OUTPUT" } } }, "post_transforms": [ { "name": "ArgmaxAcrossChannels", "args": { "fields": "model" } }, { "name": "FetchExtremePoints", "args": { "image_field": "image", "label_field": "model", "points": "points" } }, { "name": "CopyProperties", "args": { "fields": ["model"], "from_field": "image", "properties": ["affine"] } }, { "name": "RestoreOriginalShape", "args": { "field": "model", "src_field": "image", "is_label": true } } ], "writer": { "name": "WriteNifti", "args": { "field": "model", "dtype": "uint8" } } }

© Copyright 2020, NVIDIA. Last updated on Feb 2, 2023.