Training NASNet with pretrianed weights but different input_shape about nasnet-keras HOT 2 CLOSED

I agree it is better if the code followed the usual Keras API conventions. So I changed the syntax for mobile and large.

nasnet.NASNetA can build a large number of models including mobile and large.
The default hyperparameters do not corresponds to large so that is why you get a shape mismatch. The model you created is not the same as the weights you have specified.

I think what you want is this:

import nasnet

input_shape = (img_height, img_width, 3)
model = nasnet.large(include_top=False, input_shape=input_shape, weights='imagenet')

Does this solve your issue @JoshuaC3 ?

from nasnet-keras.

@johannesu - thank you! - this worked nicely and it uses the Keras API as you mentioned.

from nasnet-keras.