glasses.models.classification.efficientnet package

Module contents

class glasses.models.classification.efficientnet.EfficientNet(encoder: torch.nn.modules.module.Module = <class 'glasses.models.classification.efficientnet.EfficientNetEncoder'>, head: torch.nn.modules.module.Module = <class 'glasses.models.classification.efficientnet.EfficientNetHead'>, *args, **kwargs)

Bases: glasses.models.classification.base.ClassificationModule

Implementation of EfficientNet proposed in EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

The basic architecture is similar to MobileNetV2 as was computed by using Progressive Neural Architecture Search .

The following table shows the basic architecture (EfficientNet-efficientnet_b0):

Then, the architecture is scaled up from -efficientnet_b0 to -efficientnet_b7 using compound scaling.

EfficientNet.efficientnet_b0()
EfficientNet.efficientnet_b1()
EfficientNet.efficientnet_b2()
EfficientNet.efficientnet_b3()
EfficientNet.efficientnet_b4()
EfficientNet.efficientnet_b5()
EfficientNet.efficientnet_b6()
EfficientNet.efficientnet_b7()
EfficientNet.efficientnet_b8()
EfficientNet.efficientnet_l2()

Examples

EfficientNet.efficientnet_b0(activation = nn.SELU)
# change number of classes (default is 1000 )
EfficientNet.efficientnet_b0(n_classes=100)
# pass a different block
EfficientNet.efficientnet_b0(block=...)
# store each feature
x = torch.rand((1, 3, 224, 224))
model = EfficientNet.efficientnet_b0()
# first call .features, this will activate the forward hooks and tells the model you'll like to get the features
model.encoder.features
model(torch.randn((1,3,224,224)))
# get the features from the encoder
features = model.encoder.features
print([x.shape for x in features])
# [torch.Size([1, 32, 112, 112]), torch.Size([1, 24, 56, 56]), torch.Size([1, 40, 28, 28]), torch.Size([1, 80, 14, 14])]

Parameters

• in_channels (int, optional) – Number of channels in the input Image (3 for RGB and 1 for Gray). Defaults to 3.

• n_classes (int, optional) – Number of classes. Defaults to 1000.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

default_depths: List[int] = [1, 2, 2, 3, 3, 4, 1]

default_widths: List[int] = [32, 16, 24, 40, 80, 112, 192, 320, 1280]

classmethod efficientnet_b0(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_b1(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_b2(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_b3(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_b4(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_b5(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_b6(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_b7(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_b8(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_l2(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod from_config(config, key, *args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

initialize()

models_config = {'efficientnet_b0': (1.0, 1.0, 0.2), 'efficientnet_b1': (1.0, 1.1, 0.2), 'efficientnet_b2': (1.1, 1.2, 0.3), 'efficientnet_b3': (1.2, 1.4, 0.3), 'efficientnet_b4': (1.4, 1.8, 0.4), 'efficientnet_b5': (1.6, 2.2, 0.4), 'efficientnet_b6': (1.8, 2.6, 0.5), 'efficientnet_b7': (2.0, 3.1, 0.5), 'efficientnet_b8': (2.2, 3.6, 0.5), 'efficientnet_l2': (4.3, 5.3, 0.5)}

training: bool

glasses.models.classification.efficientnet.EfficientNetBasicBlock

alias of glasses.models.classification.efficientnet.InvertedResidualBlock

class glasses.models.classification.efficientnet.EfficientNetEncoder(in_channels: int = 3, widths: List[int] = [32, 16, 24, 40, 80, 112, 192, 320, 1280], depths: List[int] = [1, 2, 2, 3, 3, 4, 1], strides: List[int] = [2, 1, 2, 2, 2, 1, 2, 1], expansions: List[int] = [1, 6, 6, 6, 6, 6, 6], kernel_sizes: List[int] = [3, 3, 5, 3, 5, 5, 3], se: List[bool] = [True, True, True, True, True, True, True], drop_rate: float = 0.2, stem: torch.nn.modules.module.Module = <class 'glasses.nn.blocks.ConvBnAct'>, activation: torch.nn.modules.module.Module = functools.partial(<class 'torch.nn.modules.activation.SiLU'>, inplace=True), **kwargs)

Bases: glasses.models.base.Encoder

EfficientNet encoder composed by multiple different layers with increasing features.

Be awere that widths and strides also includes the width and stride for the steam in the first position.

Parameters

• in_channels (int, optional) – [description]. Defaults to 3.

• widths (List[int], optional) – [description]. Defaults to [32, 16, 24, 40, 80, 112, 192, 320, 1280].

• depths (List[int], optional) – [description]. Defaults to [1, 2, 2, 3, 3, 4, 1].

• strides (List[int], optional) – [description]. Defaults to [2, 1, 2, 2, 2, 1, 2, 1].

• expansions (List[int], optional) – [description]. Defaults to [1, 6, 6, 6, 6, 6, 6].

• kernel_sizes (List[int], optional) – [description]. Defaults to [3, 3, 5, 3, 5, 5, 3].

• se (List[bool], optional) – [description]. Defaults to [True, True, True, True, True, True, True].

• drop_rate (float, optional) – [description]. Defaults to 0.2.

• activation (nn.Module, optional) – [description]. Defaults to nn.SiLU.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

property features_widths

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

property stages

training: bool

class glasses.models.classification.efficientnet.EfficientNetHead(in_features: int, n_classes: int, drop_rate: float = 0.2)

Bases: torch.nn.modules.container.Sequential

This class represents the head of EfficientNet. It performs a global pooling, dropout and maps the output to the correct class by using a fully connected layer.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class glasses.models.classification.efficientnet.EfficientNetLite(encoder: torch.nn.modules.module.Module = <class 'glasses.models.classification.efficientnet.EfficientNetEncoder'>, head: torch.nn.modules.module.Module = <class 'glasses.models.classification.efficientnet.EfficientNetHead'>, *args, **kwargs)

Bases: glasses.models.classification.efficientnet.EfficientNet

Implementations of EfficientNetLite proposed in Higher accuracy on vision models with EfficientNet-Lite

Main differences from the EfficientNet implementation are:

• Removed squeeze-and-excitation networks since they are not well supported

• Replaced all swish activations with RELU6, which significantly improved the quality of post-training quantization (explained later)

• Fixed the stem and head while scaling models up in order to reduce the size and computations of scaled models

Examples

Create a default model

>>> EfficientNetLite.efficientnet_lite0()
>>> EfficientNetLite.efficientnet_lite1()
>>> EfficientNetLite.efficientnet_lite2()
>>> EfficientNetLite.efficientnet_lite3()
>>> EfficientNetLite.efficientnet_lite4()

Parameters

• in_channels (int, optional) – Number of channels in the input Image (3 for RGB and 1 for Gray). Defaults to 3.

• n_classes (int, optional) – Number of classes. Defaults to 1000.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

classmethod efficientnet_lite0(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_lite1(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_lite2(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_lite3(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod efficientnet_lite4(*args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

classmethod from_config(config, key, *args, **kwargs) → glasses.models.classification.efficientnet.EfficientNet

models_config = {'efficientnet_lite0': (1.0, 1.0, 0.2), 'efficientnet_lite1': (1.0, 1.1, 0.2), 'efficientnet_lite2': (1.1, 1.2, 0.3), 'efficientnet_lite3': (1.2, 1.4, 0.3), 'efficientnet_lite4': (1.4, 1.8, 0.3)}

training: bool

class glasses.models.classification.efficientnet.InvertedResidualBlock(in_features: int, out_features: int, stride: int = 1, expansion: int = 6, activation: torch.nn.modules.module.Module = <class 'torch.nn.modules.activation.SiLU'>, drop_rate: float = 0.2, se: bool = True, kernel_size: int = 3, **kwargs)

Bases: torch.nn.modules.module.Module

Inverted residual block proposed originally for MobileNetV2.

Parameters

• in_features (int) – Number of input features

• out_features (int) – Number of output features

• stride (int, optional) – Stide used in the depth convolution. Defaults to 1.

• expansion (int, optional) – The expansion ratio applied. Defaults to 6.

• activation (nn.Module, optional) – The activation funtion used. Defaults to nn.SiLU.

• drop_rate (float, optional) – If > 0, add a nn.Dropout2d at the end of the block. Defaults to 0.2.

• se (bool, optional) – If True, add a ChannelSE module after the depth convolution. Defaults to True.

• kernel_size (int, optional) – [description]. Defaults to 3.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(x: torch.Tensor) → torch.Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool