Module kerod.model.factory

None

None

View Source

from enum import Enum

import tensorflow as tf

from kerod.model.detr import DeTrResnet50, DeTrResnet50Pytorch

from kerod.model.faster_rcnn import (FasterRcnnFPNResnet50Caffe, FasterRcnnFPNResnet50Pytorch)

from kerod.model.smca_detr import SMCAR50Pytorch

from kerod.utils.training import (freeze_batch_normalization, freeze_layers_before)

class KerodModel(str, Enum):

    faster_rcnn_resnet50_pytorch = 'resnet50_pytorch'

    faster_rcnn_resnet50_caffe = 'resnet50_caffe'

    detr_resnet50 = 'detr_resnet50_pytorch'

    detr_resnet50_caffe = 'detr_resnet50'

    smca_r50 = 'smca_resnet50'

def build_model(num_classes: int,

                name: str = KerodModel.faster_rcnn_resnet50_pytorch.value) -> tf.keras.Model:

    """Build a localization model with all the `tf.keras.layers.BatchNormalization` frozen and

    all the layers before second residual block.

    Args:

        num_classes: Number of classes of your model. Do not include the background class.

        name: Target model that you wish to use: 'resnet50_pytorch', 'resnet50_caffe', 'detrresnet50_pytorch', 'smca_resnet50'.

    Returns:

        A `keras.Model` instance.

    Raises:

        NotImplementedError: If the provided isn't supported

    """

    if name == KerodModel.faster_rcnn_resnet50_pytorch:

        model = FasterRcnnFPNResnet50Pytorch(num_classes)

        freeze_batch_normalization(model.backbone)

        freeze_layers_before(model.backbone, 'resnet50/group0/block2/last_relu')

        return model

    elif name == KerodModel.faster_rcnn_resnet50_caffe:

        model = FasterRcnnFPNResnet50Caffe(num_classes)

        freeze_batch_normalization(model.backbone)

        freeze_layers_before(model.backbone, 'conv2_block3_out')

        return model

    elif name == KerodModel.detr_resnet50:

        model = DeTrResnet50Pytorch(num_classes)

        freeze_batch_normalization(model.backbone)

        freeze_layers_before(model.backbone, 'resnet50/group0/block2/last_relu')

        return model

    elif name == KerodModel.detr_resnet50_caffe:

        model = DeTrResnet50(num_classes)

        freeze_batch_normalization(model.backbone)

        freeze_layers_before(model.backbone, 'conv2_block3_out')

        return model

    elif name == KerodModel.smca_r50:

        model = SMCAR50Pytorch(num_classes)

        freeze_batch_normalization(model.backbone)

        freeze_layers_before(model.backbone, 'resnet50/group0/block2/last_relu')

        return model

    raise NotImplementedError(f'Name: {name} is not implemented.')

Functions

build_model

def build_model(
    num_classes: int,
    name: str = 'resnet50_pytorch'
) -> tensorflow.python.keras.engine.training.Model

Build a localization model with all the tf.keras.layers.BatchNormalization frozen and

all the layers before second residual block.

Parameters:

Name	Description
num_classes	Number of classes of your model. Do not include the background class.
name	Target model that you wish to use: 'resnet50_pytorch', 'resnet50_caffe', 'detrresnet50_pytorch', 'smca_resnet50'.

Returns:

Type	Description
None	A keras.Model instance.

Raises:

Type	Description
NotImplementedError	If the provided isn't supported

View Source

def build_model(num_classes: int,

                name: str = KerodModel.faster_rcnn_resnet50_pytorch.value) -> tf.keras.Model:

    """Build a localization model with all the `tf.keras.layers.BatchNormalization` frozen and

    all the layers before second residual block.

    Args:

        num_classes: Number of classes of your model. Do not include the background class.

        name: Target model that you wish to use: 'resnet50_pytorch', 'resnet50_caffe', 'detrresnet50_pytorch', 'smca_resnet50'.

    Returns:

        A `keras.Model` instance.

    Raises:

        NotImplementedError: If the provided isn't supported

    """

    if name == KerodModel.faster_rcnn_resnet50_pytorch:

        model = FasterRcnnFPNResnet50Pytorch(num_classes)

        freeze_batch_normalization(model.backbone)

        freeze_layers_before(model.backbone, 'resnet50/group0/block2/last_relu')

        return model

    elif name == KerodModel.faster_rcnn_resnet50_caffe:

        model = FasterRcnnFPNResnet50Caffe(num_classes)

        freeze_batch_normalization(model.backbone)

        freeze_layers_before(model.backbone, 'conv2_block3_out')

        return model

    elif name == KerodModel.detr_resnet50:

        model = DeTrResnet50Pytorch(num_classes)

        freeze_batch_normalization(model.backbone)

        freeze_layers_before(model.backbone, 'resnet50/group0/block2/last_relu')

        return model

    elif name == KerodModel.detr_resnet50_caffe:

        model = DeTrResnet50(num_classes)

        freeze_batch_normalization(model.backbone)

        freeze_layers_before(model.backbone, 'conv2_block3_out')

        return model

    elif name == KerodModel.smca_r50:

        model = SMCAR50Pytorch(num_classes)

        freeze_batch_normalization(model.backbone)

        freeze_layers_before(model.backbone, 'resnet50/group0/block2/last_relu')

        return model

    raise NotImplementedError(f'Name: {name} is not implemented.')

Classes

KerodModel

class KerodModel(
    /,
    *args,
    **kwargs
)

Ancestors (in MRO)

• builtins.str
• enum.Enum

Class variables

detr_resnet50

detr_resnet50_caffe

faster_rcnn_resnet50_caffe

faster_rcnn_resnet50_pytorch

name

smca_r50

value