Inference¶
The dh_segment.inference module implements the function related to the usage of a dhSegment model,
for instance to use a trained model to inference on new data.
Loading a model¶
|
Loads an exported dhSegment model |
-
class
dh_segment.inference.LoadedModel(model_base_dir, predict_mode='filename', num_parallel_predictions=2)¶ Loads an exported dhSegment model
- Parameters
model_base_dir – the model directory i.e. containing saved_model.{pb|pbtxt}. If not, it is assumed to be a TF exporter directory, and the latest export directory will be automatically selected.
predict_mode – defines the input/output format of the prediction output (see .predict())
num_parallel_predictions – limits the number of conccurent calls of predict to avoid Out-Of-Memory issues if predicting on GPU
-
predict(input_tensor, prediction_key=None)¶ Performs the prediction from the loaded model according to the prediction mode.
Prediction modes:
prediction_mode
input_tensor
Output prediction dictionnary
Comment
filename
Single filename string
labels, probs, original_shape
Loads the image, resizes it, and predicts
filename_original_shape
Single filename string
labels, probs
Loads the image, resizes it, predicts and scale the output to the original resolution of the file
image
Single input image [1,H,W,3] float32 (0..255)
labels, probs, original_shape
Resizes the image, and predicts
image_original_shape
Single input image [1,H,W,3] float32 (0..255)
labels, probs
Resizes the image, predicts, and scale the output to the original resolution of the input
image_resized
Single input image [1,H,W,3] float32 (0..255)
labels, probs
Predicts from the image input directly
- Parameters
input_tensor – a single input whose format should match the prediction mode
prediction_key – if not None, will returns the value of the corresponding key of the output dictionnary instead of the full dictionnary
- Returns
the prediction output
-
predict_with_tiles(filename, resized_size=None, tile_size=500, min_overlap=0.2, linear_interpolation=True)¶