is QAT still on the roadmap for keras3?

[matpalm]

is QAT still on the roadmap?

( #18930 from Feb has it as a "near future" release. )

have been porting code to keras3 ( as a way to move more to Jax ) & have options for post training quantisation but expect a non trivial benefit from QAT in a number of projects.

can see a path forward by partially porting pieces of https://www.tensorflow.org/model_optimization/api_docs/python/tfmot and/or https://github.com/google/aqt but will hold back if a QAT api is imminent ? ( additionally; might have some bandwidth to help if there are community contrib options? )

[matpalm]

bump

though i can jump through hoops to get tfmot working ( ish ) the fact it is keras2 compat only is a pain :/

if keras3 QAT is pending i can hold out, otherwise will have to consider dropping back to keras2 :( :( :(

[fchollet]

QAT is still on the roadmap but the work hasn't started yet, as we've had to prioritize other things... we can definitely re-prioritize it though. Most of the infra is already here in via the work @james77777778 did for float8 training.

[matpalm]

great! if there's any pointers you can give me @james77777778 i'd be happy to help out; should have a bit of bandwidth for this over the next few months.

[james77777778]

I'm happy to help with this but unfortunately I don't have enough bandwidth right now.

@matpalm I suggest starting by implementing these ops in a backend-agnostic way, as they are necessary for QAT:

• tf.quantization.fake_quant_with_min_max_args
• tf.quantization.fake_quant_with_min_max_vars
• tf.quantization.fake_quant_with_min_max_vars_per_channel

You should be able to implement them using ops.custom_gradient

[matpalm]

great thanks, i'll follow up in a PR

[james77777778]

I can help review the implementation if that sounds good to @fchollet!

[fchollet]

Awesome, thanks you two!