image_classification/tf/src/models/efficientnetv2.py

# /*---------------------------------------------------------------------------------------------
#  * Copyright 2018 The TensorFlow Authors.
#  * Copyright (c) 2022-2023 STMicroelectronics.
#  * All rights reserved.
#  *
#  * This software is licensed under terms that can be found in the LICENSE file in
#  * the root directory of this software component.
#  * If no LICENSE file comes with this software, it is provided AS-IS.
#  *--------------------------------------------------------------------------------------------*/


import keras
from keras.applications import imagenet_utils
from keras import layers
from keras.applications.efficientnet_v2 import (EfficientNetV2B0, EfficientNetV2B1, EfficientNetV2B2, EfficientNetV2B3,
                                                EfficientNetV2S)


def get_efficientnetv2(input_shape: tuple, model_type: str = None, num_classes: int = None, dropout: float = None,
                       pretrained: bool = True, **kwargs) -> keras.Model:
    """
    Returns a transfer learning model based on efficient net v2 architecture pre-trained on imagenet or random.

    Args:
        input_shape (tuple): Shape of the input tensor.
        model_type (string): B0, B1, B2, B3, S. Default is None.
        num_classes (int): Number of output classes of the target use-case. Default is None.
        dropout (float, optional): The dropout rate for the custom classifier.
        pretrained_weights (str, optional): The pre-trained weights to use. Either "imagenet" or None.

    Returns:
        keras.Model: Transfer learning model based on efficient net v2 architecture.

    Raises:

    """
    
    # commented because no longer used later-on
#    if pretrained_weights:
#        training = False
#        # model is set in inference mode so that moving avg and var of any BN are kept untouched
#        # should help the convergence according to Keras tutorial
#    else:
#        training = True

    # fetch the backbone pre-trained on imagenet or random
    if model_type == "B0":
        backbone_func = EfficientNetV2B0
    elif model_type == "B1":
        backbone_func = EfficientNetV2B1
    elif model_type == "B2":
        backbone_func = EfficientNetV2B2
    elif model_type == "B3":
        backbone_func = EfficientNetV2B3
    elif model_type == "S":
        backbone_func = EfficientNetV2S

    if dropout:
      # Model loaded for training
      base_model = backbone_func(
          include_top=False,
          weights="imagenet" if pretrained else None,
          input_tensor=None,
          input_shape=input_shape,
          pooling="avg",
          classes=num_classes,
          include_preprocessing=False
      )
    
      # Create a new model on top
      x = layers.Dropout(rate=dropout, name="dropout")(base_model.output)
      if num_classes > 2:
          outputs = layers.Dense(num_classes, activation="softmax")(x)
      else:
          outputs = layers.Dense(1, activation="sigmoid")(x)
    
    else:
      # Model entirely loaded for other services than training => no dropout
      base_model = backbone_func(
          include_top=True,
          weights="imagenet" if pretrained else None,
          input_tensor=None,
          input_shape=input_shape,
          pooling="avg",
          classes=num_classes,
          include_preprocessing=False,
          classifier_activation="softmax"
      )
      outputs = base_model.output

    # Create the Keras model
    model = keras.Model(inputs=base_model.input, outputs=outputs, name="efficientnetv2"+model_type)

    return model