CODEWITHSUNDEEP
pythontensorflowtensorflow-lite
Tsmith
Tsmith

Reputation: 21

Where is code to perform the fake quantize in TensorFlow?

I would like to make a custom quantizer (not stardard 8 bit) in TensorFlow.

I've gone through the code in tensorflow\tensorflow\contrib\quantize\python and can see how the nodes are added, but I would like to modify how the tf.fake_quantize_with_min_max_vars function calculates that outputs.

I cannot seem to find the code that actually does the 32 bit accumulate and downsampling to 8 bit. Can anyone point me to where this code resides?

Upvotes: 1

Views: 161

Answers (1)

Pete Warden
Pete Warden

Reputation: 2878

The code that does the actual quantization of the values is in C++, in this function here: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/fake_quant_ops_functor.h#L79

It's not particularly easy to modify, since you'll need to rebuild TensorFlow to get the changes, but hopefully that gives you a start.

Upvotes: 1

Related Questions