Prediction Set-Up

To make accurate predictions on images, several parameters must be provided: - Class names: The model predicts class IDs, but to visualize results, the class names from the training dataset are needed. - Processing parameters: The model requires input data in a specific format. - Task-specific parameters: For instance, in the case of Detection, this includes IoU and Confidence thresholds.

SuperGradients manages all of these within its model.predict() method, but in certain scenarios, you might need to set these parameters explicitly first.

1. Training your model on a custom dataset

If you trained a model on a dataset that does not inherit from any of the SuperGradients datasets, you will need to set the processing parameters explicitly. To do this, use the model.set_dataset_processing_params() method. Once you've set the parameters, you can run model.predict().

2. Using pretrained weights or training on a SuperGradient's dataset

All necessary information is automatically saved during training within the model checkpoint, so you can run model.predict() without calling model.set_dataset_processing_params().

For more details about model.predict(), please refer to the related tutorial.

Set-up parameters

Class Names

This is straightforward as it corresponds to the list of classes used during training. For instance, if you're loading the weights of a model fine-tuned on a new dataset, use the classes from that dataset.

class_names = [
    "person",
    "bicycle",
    "car",
    "motorcycle",
    "airplane",
    "bus",
    ...
]

Ensure that the class order remains the same as during training.

Processing

Processing steps are necessary for making predictions. - Image preprocessing prepares the input data for the model by applying various transformations, such as resizing, normalization, and channel reordering. These transformations ensure the input data is compatible with the model. - Image postprocessing processes the model's output and converts it into a human-readable and interpretable format. This step may include tasks like converting class probabilities into class labels, applying non-maximum suppression to eliminate duplicate detections, and rescaling results to the original image size.

The super_gradients.training.processing module contains a wide range of Processing transformations responsible for both image preprocessing and postprocessing.

For example, DetectionCenterPadding applies center padding to the image while also handling the reverse transformation to remove padding from the prediction.

Multiple processing transformations can be combined using ComposeProcessing:

from super_gradients.training.processing import DetectionCenterPadding, StandardizeImage, NormalizeImage, ImagePermute, ComposeProcessing, DetectionLongestMaxSizeRescale

image_processor = ComposeProcessing(
    [
        DetectionLongestMaxSizeRescale(output_shape=(636, 636)),
        DetectionCenterPadding(output_shape=(640, 640), pad_value=114),
        StandardizeImage(max_value=255.0),
        ImagePermute(permutation=(2, 0, 1)),
    ]
)

Task Specific parameters

Detection

Default iou and conf values can be set, which will be used when calling model.predict(). - iou: IoU threshold for the non-maximum suppression (NMS) algorithm. If None, the default value associated with training is used. - conf: Confidence threshold. Predictions below this threshold are discarded. If None, the default value associated with training is used.

Saving your processing parameters to your model

After defining all parameters, call model.set_dataset_processing_params() and then use model.predict().

from super_gradients.common.object_names import Models
from super_gradients.training import models

model = models.get(Models.YOLO_NAS_L, checkpoint_path="/path/to/checkpoint")

model.set_dataset_processing_params(
    class_names=class_names,
    image_processor=image_processor,
    iou=0.35, conf=0.25,
)

IMAGES = [...]

images_predictions = model.predict(IMAGES)

For more information about the model.predict(), please check out the following tutorial.