Jae sung Chung comments

Results 20 comments of


                                            Jae sung Chung

WARNING:tensorflow:AutoGraph could not transform <tensorflow_model_optimization.python.core.quantization.keras.quantizers.AllValuesQuantizer object at 0x7f92927f3f90> and will run it as-is.

@Xhark any thoughts on this?

model written with tf2.6.0 (not using keras) cannot find instructions to do quantization aware training, only keras model based quantization example can be found

Current QAT feature is based on the Keras API. See the following document as well: https://www.tensorflow.org/model_optimization/guide/quantization/training

Depthwise convolution min_max dimension

@Xhark could you take a look?

Set zero_debias=True for Quantization Aware Training

@Xhark could you take a look at this one?

Issue for the FAKE_QUANT result

@Xhark could you take a look at this question?

Weight in fully connected layers don't follow tensorflow quantization spec (zero-point!=0)

@thaink @teijeong @daverim for the visibility.

modify_model_interface removing input float tensor is not completed, memory allocator occupied this dummy tensor

Thanks for filing the feature request! Would you mind uploading an affected TFLite model file or sharing the reproducible steps?

Update TFLite concrete function conversion codes to have the associat…

@hmoraldo could you take a look? Thanks :)

Add zero point check for kTfLiteInt16 and scale check for multiple operators

@smpurkis could you share the motivation of this change? Why are these runtime check necessary?

[TFLite] Add int16x8 support for RELU_N1_TO_1 and PRELU operators

@alankelly could you take a look at this pull request? In quantized kernel implementation wise, it looks okay to me.