Latency regularizer for a target hardware. about morph-net HOT 5 CLOSED

from morph-net.

i am trying using latency regulariser but it seems to stuck when i add the metric cost.

It will be very helpful if you could take a look
https://colab.research.google.com/drive/1W2jC3dLnt5Pl1RLNRk5gi5JILrfMHqbW?usp=sharing

from morph-net.

The problem occurs when trying to add the loss and the cost as metrics, a possible workaround is to define a function that returns the symbolic tensor you're interested in and use pass the function to the arguments in the compile method of the model.

from morph-net.

Thanx @ferova and @mn-robot .I was able to acheive the same.

from morph-net.

If you have the peak compute and memory bandwidth numbers for your hardware you can add them to the dictionary here with a new key e.g. "my_hardware"

morph-net/morph_net/network_regularizers/resource_function.py

Line 324 in 7ade4c1

def latency_function_factory(hardware, batch_size):

And then use "my_hardware" as the hardware value for the latency regularizer
…
--Yair
On Fri, Dec 4, 2020 at 7:59 AM 'Akshat Jaitly' via mn_robot < @.> wrote: I have a use case in which I have to tune a model according to my mobile hardware for faster inference and lower latency. How could I acheive this using morphnet. I need to pass the hardware values which are keys in peak_compute in resource function. py. These are some fixed constants Is there any ways I can make my model hardware aware and tune the model according to a specific hardware. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#196>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALRY4TZITZAEYPPPCZDU7VDSTEBPPANCNFSM4UNT6CEA . -- You received this message because you are subscribed to the Google Groups "mn_robot" group. To unsubscribe from this group and stop receiving emails from it, send an email to @. To view this discussion on the web visit https://groups.google.com/a/google.com/d/msgid/mn_robot/google-research/morph-net/issues/196%40github.com https://groups.google.com/a/google.com/d/msgid/mn_robot/google-research/morph-net/issues/196%40github.com?utm_medium=email&utm_source=footer .

Hi mn-robot!
How do we calculate the peak compute and memory bandwidth given the target device (mobile devices in my case) .

Thank You

from morph-net.