Attention: This MediaPipe Solutions Preview is an early release. Learn more.

mediapipe_model_maker.gesture_recognizer.GestureRecognizer

Stay organized with collections Save and categorize content based on your preferences.

GestureRecognizer for building hand gesture recognizer model.

Inherits From: Classifier, CustomModel

label_names A list of label names for the classes.
model_options options to create gesture recognizer model.
hparams The hyperparameters for training hand gesture recognizer model.

embedding_size Size of the input gesture embedding vector.

Methods

create

View source

Creates and trains a hand gesture recognizer with input datasets.

If a checkpoint file exists in the {options.hparams.export_dir}/checkpoint/ directory, the training process will load the weight from the checkpoint file for continual training.

Args
train_data Training data.
validation_data Validation data. If None, skips validation process.
options options for creating and training gesture recognizer model.

Returns
An instance of GestureRecognizer.

evaluate

View source

Evaluates the classifier with the provided evaluation dataset.

Args
data Evaluation dataset
batch_size Number of samples per evaluation step.

Returns
The loss value and accuracy.

export_labels

View source

Exports classification labels into a label file.

Args
export_dir The directory to save exported files.
label_filename File name to save labels model. The full export path is {export_dir}/{label_filename}.

export_model

View source

Converts the model to TFLite and exports as a model bundle file.

Saves a model bundle file and metadata json file to hparams.export_dir. The resulting model bundle file will contain necessary models for hand detection, canned gesture classification, and customized gesture classification. Only the model bundle file is needed for the downstream gesture recognition task. The metadata.json file is saved only to interpret the contents of the model bundle file.

The customized gesture model is in float without quantization. The model is lightweight and there is no need to balance performance and efficiency by quantization. The default score_thresholding is set to 0.5 as it can be adjusted during inference.

Args
model_name File name to save model bundle file. The full export path is {export_dir}/{model_name}.

export_tflite

View source

Converts the model to requested formats.

Args
export_dir The directory to save exported files.
tflite_filename File name to save TFLite model. The full export path is {export_dir}/{tflite_filename}.
quantization_config The configuration for model quantization.
preprocess A callable to preprocess the representative dataset for quantization. The callable takes three arguments in order: feature, label, and is_training.

summary

View source

Prints a summary of the model.