To solve the issue #13690 in keras

[seriousran]

To solve the issue #13690 in keras

https://github.com/keras-team/keras/issues/13690

[taehoonlee]

Sorry for the late response, @seriousran. It is intended to designate require_flatten=True in NASNet for disabling NASNet with non-default input shapes. For example, the followings are intended behaviours.

• NASNetMobile(weights='imagenet', include_top=True): accept
• NASNetMobile(weights='imagenet', include_top=False): accept
• NASNetMobile(weights='imagenet', include_top=True, input_shape=(None, None, 3)): reject
• NASNetMobile(weights='imagenet', include_top=False, input_shape=(None, None, 3)): reject
• NASNetMobile(weights='imagenet', include_top=True, input_shape=(100, 100, 3)): reject
• NASNetMobile(weights='imagenet', include_top=False, input_shape=(100, 100, 3)): reject

The underlying fact is that the architecture of NASNet differs according to input_shape. Specifically, the line 513 causes structural changes, because the zero-padding in 2x down-sampling (conv, pool) perform differently according to whether input_shape is even or odd.

Thus, NASNet with weights='imagenet', include_top=any, input_shape=(None, None, 3) is meaningless because it may differ from the intended structure of the authors.

[seriousran]

@taehoonlee Thanks for detailed response :) Now I can understand NASNet architecture more precisely.