print_qstats(): operation type issue with Sequential() model about qkeras HOT 6 CLOSED

@HaoranREN I've been looking into your issue and created a fix. You should hopefully have access to it soon.

The reason why you see unull_4_-1 for the first layer is because QKeras cannot find the input tensor to the first layer. This means QKeras doesn't know the the bit size used by the input tensor, and it therefore cannot determine what hardware operation would be needed to process that input. That's why the mode is null, and the bits of the inputs tensor is -1.

The reason why this is happening in the Sequential API and not the Functional API is because the Input Layer in the Sequential API is "hidden" when iterating model layers, and so the tool does not cache the input bits that would be passed in to the first convolution layer. This is just a quirk of Keras, so the fix allows code needs to handle both situations.

Thank you for bringing this to our attention,

Daniele

from qkeras.

@danielemoro please take a look. thanks

from qkeras.

Hi @HaoranREN,
Thank you for your input.
smult_4_8 means s:signed multi:multiplication op, 4:weight quantizer bit, 8:input quantizer bit
Regarding the undesired operation type for a Sequential model, we are fixing it right now.

from qkeras.

@lishanok Thanks a lot for your explanations 👍

from qkeras.

@danielemoro Thanks a lot for your response and it makes a lot of sense. I was also suspecting the Keras implementation, since like a said, it only happens to the first layer of the Sequential() model.

from qkeras.

a fix was merged. close it. feel free to reopen it if you have any questions.

from qkeras.