se_layer(ratio: int = 8, name: str | None = None, squeeze_activation: str | Callable = 'relu6', excite_activation: str | Callable = 'hard_sigmoid') -> keras.Layer
Squeeze & excite functional layer
Implements Squeeze and Excite block as in
Squeeze-and-Excitation Networks.
Parameters:
-
ratio
(Expansion ratio, default:
8
)
–
Expansion ratio. Defaults to 8.
-
name
(str | None, default:
None
)
–
Block name. Defaults to None.
-
squeeze_activation
(str | Callable, default:
'relu6'
)
–
Squeeze activation. Defaults to "relu6".
-
excite_activation
(str | Callable, default:
'hard_sigmoid'
)
–
Excite activation. Defaults to "hard_sigmoid".
Returns:
-
Layer
–
keras.Layer: Functional SE layer
Source code in neuralspot_edge/layers/squeeze_excite.py
| def se_layer(
ratio: int = 8,
name: str | None = None,
squeeze_activation: str | Callable = "relu6",
excite_activation: str | Callable = "hard_sigmoid",
) -> keras.Layer:
"""Squeeze & excite functional layer
Implements Squeeze and Excite block as in
[Squeeze-and-Excitation Networks](https://arxiv.org/pdf/1709.01507.pdf).
Args:
ratio (Expansion ratio, optional): Expansion ratio. Defaults to 8.
name (str|None, optional): Block name. Defaults to None.
squeeze_activation (str|Callable, optional): Squeeze activation. Defaults to "relu6".
excite_activation (str|Callable, optional): Excite activation. Defaults to "hard_sigmoid".
Returns:
keras.Layer: Functional SE layer
"""
def layer(x: keras.KerasTensor) -> keras.KerasTensor:
num_chan = x.shape[-1]
# Squeeze
name_pool = f"{name}.pool" if name else None
name_sq = f"{name}.sq" if name else None
name_sq_act = f"{name}.sq.act" if name else None
y = keras.layers.GlobalAveragePooling2D(name=name_pool, keepdims=True)(x)
y = keras.layers.Conv2D(filters=int(num_chan // ratio), kernel_size=(1, 1), use_bias=True, name=name_sq)(y)
y = keras.layers.Activation(squeeze_activation, name=name_sq_act)(y)
# Excite
name_ex = f"{name}.ex" if name else None
name_ex_act = f"{name}.ex.act" if name else None
name_ex_mul = f"{name}.ex.mul" if name else None
y = keras.layers.Conv2D(num_chan, kernel_size=(1, 1), use_bias=True, name=name_ex)(y)
y = keras.layers.Activation(excite_activation, name=name_ex_act)(y)
y = keras.layers.Multiply(name=name_ex_mul)([x, y])
return y
# END DEF
return layer
|