The ge_layer implementation is not exactly the same as that in the paper

[zding047]

In the paper, when stride=2, the left branch (Figure 5 (c)) has two depth-wise layers. Their output channels are the same, which is expand_ratio * C. But in the codes, due to the chain rule, the second layer's output has expand_ratio * expand_ratio * C channels.

Also, the expansion is with respect to the input dimension, instead of the output dimension.

bisenetv2-tf2/model.py

Lines 14 to 37 in e018b3b

	def ge_layer(x_in, c, e=6, stride=1):
	x = layers.Conv2D(filters=c, kernel_size=(3,3), padding='same')(x_in)
	x = layers.BatchNormalization()(x)
	x = layers.Activation('relu')(x)
	if stride == 2:
	x = layers.DepthwiseConv2D(depth_multiplier=e, kernel_size=(3,3), strides=2, padding='same')(x)
	x = layers.BatchNormalization()(x)
	y = layers.DepthwiseConv2D(depth_multiplier=e, kernel_size=(3,3), strides=2, padding='same')(x_in)
	y = layers.BatchNormalization()(y)
	y = layers.Conv2D(filters=c, kernel_size=(1,1), padding='same')(y)
	y = layers.BatchNormalization()(y)
	else:
	y = x_in
	x = layers.DepthwiseConv2D(depth_multiplier=e, kernel_size=(3,3), padding='same')(x)
	x = layers.BatchNormalization()(x)
	x = layers.Conv2D(filters=c, kernel_size=(1,1), padding='same')(x)
	x = layers.BatchNormalization()(x)
	x = layers.Add()([x, y])
	x = layers.Activation('relu')(x)
	return x

[markus-k]

Thank you for reporting the issue! I have to admit that I'm quite out of the loop regarding ML, as this was part of my studies. I have created a pull request that should fix the double expension, feel free to have a look.

Also, the expansion is with respect to the input dimension, instead of the output dimension.
I'm not entirely sure what you mean by that, and how I'd go about fixing that. I'd be happy to accept a pull request :)