Fix the Doc and Error Format of `round_mode` in func `tf.quantization.quantize()`

[ILCSFNO]

Issue type

Documentation Bug

Have you reproduced the bug with TensorFlow Nightly?

Yes

Source

binary

TensorFlow version

tf v2.20.0-rc0-4-g72fbba3d20f 2.20.0

Custom code

Yes

OS platform and distribution

Linux Ubuntu 22.04

Mobile device

Linux Ubuntu 22.04

Python version

3.9

Bazel version

None

GCC/compiler version

None

CUDA/cuDNN version

None

GPU model and memory

None

Current behavior?

The doc of tf.quantization.quantize() shows its description as below:

tensorflow/tensorflow/python/ops/array_ops.py

Line 6133 in 5bc9d26

quantize.__doc__ = gen_array_ops.quantize_v2.__doc__

tensorflow/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td

Line 11981 in 0724da0

DefaultValuedOptionalAttr<TF_AnyStrAttrOf<["HALF_AWAY_FROM_ZERO", "HALF_TO_EVEN"]>, "\"HALF_AWAY_FROM_ZERO\"">:$round_mode,

And for repro below, which shows that Round mode 'HALF_TO_EVEN' only supported for mode 'SCALED'. I accept that it should raise error, but it is something that need to be noted in doc string.

Error raised through codes below, which I think is needed to be format also, from "b ut mode is '" to "but mode is '":

tensorflow/tensorflow/core/kernels/quantize_op.cc

Lines 95 to 104 in 0724da0

	if (round_mode_string == "HALF_AWAY_FROM_ZERO") {
	round_mode_ = ROUND_HALF_AWAY_FROM_ZERO;
	} else if (round_mode_string == "HALF_TO_EVEN") {
	OP_REQUIRES(ctx, mode_string == "SCALED",
	errors::InvalidArgument("Round mode 'HALF_TO_EVEN' "
	"only supported for mode 'SCALED', "
	"b ut mode is '" +
	mode_string + "'."));
	round_mode_ = ROUND_HALF_TO_EVEN;
	}

Suggestions

• Note the limitation that Round mode 'HALF_TO_EVEN' only supported for mode 'SCALED'
• Fix the error format to make the description alright

Standalone code to reproduce the issue

import tensorflow as tf
input_tensor = tf.random.uniform(shape=[1, 3, 256, 256], minval=0, maxval=1)
(quantized_tensor, output_min, output_max) = tf.quantization.quantize(input=input_tensor, min_range=0.0, max_range=1.0, T=tf.qint8, mode='MIN_COMBINED', round_mode='HALF_TO_EVEN', narrow_range=False, axis=None, ensure_minimum_range=0.01)
print('Input Tensor Shape:', input_tensor.shape)
print('Output Tensor Shape:', quantized_tensor.shape)

Relevant log output

InvalidArgumentError: {{function_node __wrapped__QuantizeV2_device_/job:localhost/replica:0/task:0/device:CPU:0}} Round mode 'HALF_TO_EVEN' only supported for mode 'SCALED', b  ut mode is 'MIN_COMBINED'. [Op:QuantizeV2] name:

[ILCSFNO]

PR opened here, thanks a lot!

• Fix the Doc and Error Format of round_mode in func tf.quantization.quantize() #99287

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No