[Solve]Encounter ValueError: Shapes (512,) and (3,) are incompatible while loading pre-trained weight about classification_models

[ytliang97]

I solve the problem.

Backend default is channel last. (B, H, W, C)

axis is an argument of BatchNorm and Concatenate.

axis = 1 refers to channel first (B, C, H, W)
axis = 3 refers to channel last (B, H, W, C)

I set channel last for input shape.

img_input = layers.Input(shape=(512, 512, 3), name='data')

But I set channel first for batchnorm and concatenate.

x = layers.BatchNormalization(name='bn_data', **{'axis': 1, 'momentum': 0.99, 'epsilon': 2e-5, 'center': True, 'scale': False})(img_input)