Difference between segmentation_models and segmentations_models_pytorch

[ouioui199]

Hello, I've been using segmentation_model developped by qubvel (thanks a lot!), written in tensorflow and keras. Recently, I've switched to Pytorch and been using segmentation_models_pytorch. When I do model.summary() on tensorflow version, the UNet with efficientnetb4 has 25M parameters, when in Pytorch version there is only 19M parameters. Why is this difference ? Thank you!

Here is the code in tensorflow:
model = sm.Unet(backbone_name = 'efficientnetb4',
input_shape = shape,
classes = 4,
activation = 'relu',
encoder_weights = 'imagenet')
model.summary()

Here is the code in Pytorch:
model = smp.Unet(encoder_name = 'efficientnet-b4'
encoder_weights = 'imagenet',
activation = nn.ReLU,
classes = 4,
decoder_attention_type = None)

[S-Lykles]

Hi, EfficientNetb4 for image classification is around 19M parameters, the fully connected layer is around 1.8M parameters which gets removed and then the decoder parameters are added. My guess is that tensorflow uses a different (larger) decoder and possibly doesnt remove the fully connected layer.

[ouioui199]

Hi, thanks for your answer. It seems to me that the difference of the number of parameters comes from the fact that skip connections are located differently from the tensorflow version to pytorch version. I come to this conclusion when checking the number of input channels of each decoder block.

[S-Lykles]

Hi I changed my answer a bit (before I saw your reply) because I realized the fully connected layer itself was not big enough to account for the difference. It is indeed the decoder part that is different. Did you notice a performance difference between smp and tensorflow?

[ouioui199]

No problem ;) From my experience, there is a performance difference between smp and tensorflow, so I ended up using the smp with efficientnet-B5 backbone (to match up with the number of parameters) and the result is better than tf with backbone efficientnet-B4