Hello,
Here: base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE,
include_top=True,
weights=‘imagenet’)
Why the input shape (160, 160, 3) is valid? If include_top=True, is the only valid input shape not (224, 224, 3)?
The include_top here is refering to the last layer of the network (mobilnet) here, it doesnt have to do with the input layer (first layer)!
1 Like
Ok, so can you please explain this (from the MobileNetV2 documentation):
input_shape: Optional shape tuple, only to be specified if include_top is False (otherwise the input shape has to be (224, 224, 3) (with "channels_last" data format) or (3, 224, 224) (with "channels_first" data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. (200, 200, 3) would be one valid value. Defaults to None. input_shape will be ignored if the input_tensor is provided.
Hello, @shaya_kahn,
You may try these 3 input_shape: (240, 240, 3), (160, 160, 3), (100, 100, 3) and, in the model summary, examine the output shape of the last layer before the Global Average Pooling. The last input shape won’t build a model, but the first two will give you two different summarys. You see the two different summarys, then you may think from there how they are valid.
Cheers,
Raymond
Edit: The above (240, 240, 3) was a typo and should be (224, 224, 3) instead.
1 Like
I see your point now, and to be frank I don’t know why this is happening, perhaps the model allows those shapes too…I also want to know why at this point!
1 Like
Hello, Raymond
I think I now understand how to find valid input shapes. From the examples you suggested, only the (160, 160, 3) works. The other options are: 96, 128, 192, 224. Correct?
1 Like
Yes, @shaya_kahn!
My (240, 240, 3) was a typo (I meant to say 224 instead of 240) 
, but it should be helpful to try 240, too.
These options are hard-coded in the implementation of MobileNetV2, and they correspond to different weight files, too:
-
Hard-coded:
-
Weight file name:
As for why these numbers are chosen, I believe you might have found the reason by examining the summaries
Cheers,
Raymond
1 Like
So Raymond @rmwkwok the Mobilnetv2 accepts those shapes too! Strange because in Tensorflow is mentioned:
which from what I understand if include_top=True it should only accept (224, 224, 3)!
This question should be for TF’s dev team. ^^
2 Likes
Well, if it comes down to that, we all cool then