Understanding Model Packages Trained on Roboflow Platform¶

Models trained on the Roboflow Platform follow a custom RF standard where runtime behavior is determined by inference_config.json.

Overview¶

When you train a model on Roboflow Platform, the exported model package includes an inference_config.json file that controls:

Image preprocessing - Auto-orientation, static cropping, contrast adjustments, grayscale conversion
Network input - Resize modes, padding, color modes, normalization
Forward pass - Batch size configuration
Post-processing - NMS parameters, activation functions

This configuration is automatically generated during training based on your dataset settings and training parameters.

Package Structure¶

A typical Roboflow Platform model package contains:

model_package/
├── model_config.json       # Auto-generated metadata
├── weights.pt              # Model weights (or .onnx, .engine)
├── class_names.txt         # Class labels
└── inference_config.json   # Runtime behavior configuration

`inference_config.json` Structure¶

Complete Example¶

{
  "image_pre_processing": {
    "auto-orient": {
      "enabled": true
    },
    "static-crop": {
      "enabled": true,
      "x_min": 10,
      "y_min": 10,
      "x_max": 630,
      "y_max": 470
    },
    "contrast": {
      "enabled": false,
      "type": "Adaptive Equalization"
    },
    "grayscale": {
      "enabled": false
    }
  },
  "network_input": {
    "training_input_size": {
      "height": 640,
      "width": 640
    },
    "dataset_version_resize_dimensions": {
      "height": 480,
      "width": 640
    },
    "dynamic_spatial_size_supported": true,
    "dynamic_spatial_size_mode": {
      "type": "divisible-padding",
      "divisor": 32
    },
    "color_mode": "rgb",
    "resize_mode": "letterbox",
    "padding_value": 114,
    "input_channels": 3,
    "scaling_factor": 255.0,
    "normalization": [
      [0.485, 0.456, 0.406],
      [0.229, 0.224, 0.225]
    ]
  },
  "forward_pass": {
    "static_batch_size": null,
    "max_dynamic_batch_size": 8
  },
  "post_processing": {
    "type": "nms",
    "fused": false,
    "nms_parameters": {
      "max_detections": 300,
      "confidence_threshold": 0.25,
      "iou_threshold": 0.45,
      "class_agnostic": 0
    }
  },
  "class_names_operations": [
    {
      "type": "class_name_removal",
      "class_name": "background"
    }
  ]
}

Configuration Sections¶

1. Image Pre-Processing¶

Controls transformations applied before network input preparation:

auto-orient - Automatically correct image orientation based on EXIF data

{
  "enabled": true
}

static-crop - Apply fixed crop to all images

{
  "enabled": true,
  "x_min": 10,
  "y_min": 10,
  "x_max": 630,
  "y_max": 470
}

contrast - Apply contrast enhancement

{
  "enabled": true,
  "type": "Adaptive Equalization"  // or "Contrast Stretching", "Histogram Equalization"
}

grayscale - Convert images to grayscale

{
  "enabled": true
}

2. Network Input¶

Defines how images are prepared for the neural network:

training_input_size - Size the model was trained on

{
  "height": 640,
  "width": 640
}

dataset_version_resize_dimensions (optional) - The non-square resize dimensions applied to images by the Roboflow dataset version before training. When present and non-square (while training_input_size is square), preprocessing performs a two-step resize: first letterbox/center-crop to these dimensions, then bilinear-interpolate to the square training_input_size. This is used to reproduce the resize behavior used during training RF-DETR models on the Roboflow Platform. Ignored when resize_mode is "stretch".

{
  "height": 480,
  "width": 640
}

resize_mode - How to resize images to match training size - "letterbox" - Maintain aspect ratio with padding - "stretch" - Stretch to fit (may distort) - "center-crop" - Crop center region

dynamic_spatial_size_supported - Whether model supports variable input sizes

dynamic_spatial_size_mode - How to handle dynamic sizes

{
  "type": "divisible-padding",
  "divisor": 32
}

3. Forward Pass¶

Controls batch processing:

{
  "static_batch_size": null,
  "max_dynamic_batch_size": 8
}

4. Post-Processing¶

NMS (Object Detection)

{
  "type": "nms",
  "fused": false,
  "nms_parameters": {
    "max_detections": 300,
    "confidence_threshold": 0.25,
    "iou_threshold": 0.45,
    "class_agnostic": 0
  }
}

How It Works¶

When you load a Roboflow Platform model:

Model package is downloaded and cached
inference_config.json is parsed and validated
Preprocessing pipeline is configured automatically
Model is initialized with appropriate backend
Post-processing is configured based on settings