From ad4bd6f752d3263a82ffc61fd02ce7dfaa622a45 Mon Sep 17 00:00:00 2001 From: Roman Kazantsev Date: Fri, 24 Feb 2023 16:13:46 +0400 Subject: [PATCH] [TF FE] Simplify FakrQuantWithMinMaxVars translator and add layer test (#15927) Signed-off-by: Kazantsev, Roman --- .../src/op/fake_quant_min_max_vars.cpp | 37 ++++++------ .../test_tf_FakeQuantWithMinMaxVars.py | 60 +++++++++++++++++++ 2 files changed, 80 insertions(+), 17 deletions(-) create mode 100644 tests/layer_tests/tensorflow_tests/test_tf_FakeQuantWithMinMaxVars.py diff --git a/src/frontends/tensorflow_common/src/op/fake_quant_min_max_vars.cpp b/src/frontends/tensorflow_common/src/op/fake_quant_min_max_vars.cpp index aed41eea03e..fa5c819b88b 100644 --- a/src/frontends/tensorflow_common/src/op/fake_quant_min_max_vars.cpp +++ b/src/frontends/tensorflow_common/src/op/fake_quant_min_max_vars.cpp @@ -18,41 +18,44 @@ OutputVector translate_fake_quant_op(const NodeContext& node) { auto min = node.get_input(1); auto max = node.get_input(2); - auto narrow_range = node.get_attribute("narrow_range"); - auto num_bits = node.get_attribute("num_bits"); + // retrieve attributes + auto narrow_range = node.get_attribute("narrow_range", false); + auto num_bits = node.get_attribute("num_bits", 8); - size_t levels = static_cast(std::pow(2, num_bits) - int(narrow_range)); - auto min_less_max = make_shared(min, max); - auto minimum = make_shared(min_less_max, max, min); + size_t levels = static_cast(pow(2, num_bits)); + levels = narrow_range ? levels - 1 : levels; - auto zero = make_shared(min.get_element_type(), Shape{}, std::vector({0})); + // compute real min and max values + Output minimum = make_shared(min, max); + Output maximum = make_shared(min, max); + // adjust min and max so that min <= 0 + auto zero = make_shared(min.get_element_type(), Shape{}, 0); auto min_greater_zero = make_shared(minimum, zero); - auto max_minus_min = make_shared(maximum, minimum); + Output max_minus_min = make_shared(maximum, minimum); minimum = make_shared(min_greater_zero, max_minus_min, maximum); + // adjust min and max so that 0 <= max auto max_less_zero = make_shared(maximum, zero); auto min_minus_max = make_shared(minimum, maximum); minimum = make_shared(max_less_zero, zero, maximum); - auto float_range = make_shared(maximum, minimum); - auto quant_min_value = int(narrow_range); - auto quant_max_value = std::pow(2, num_bits) - 1; - auto value = static_cast(quant_max_value - quant_min_value); - auto int_range = make_shared(element::f32, Shape{}, std::vector({value})); - auto scale = make_shared(float_range, int_range); + // adjust min and max so that scale = (max - min) / (2^num_bits - 1), + // min_adj = scale * round(min / scale) and max_adj = max + min_adj - min + max_minus_min = make_shared(maximum, minimum); + auto const_levels = make_shared(element::f32, Shape{}, static_cast(levels - 1)); + auto scale = make_shared(max_minus_min, const_levels); auto descaled_min = make_shared(minimum, scale); auto rounded_descaled_min = make_shared(descaled_min, Round::RoundMode::HALF_TO_EVEN); auto min_adj = make_shared(scale, rounded_descaled_min); auto adjustment = make_shared(min_adj, minimum); auto max_adj = make_shared(maximum, adjustment); - auto res = make_shared(inputs, min_adj, max_adj, min_adj, max_adj, levels); - set_node_name(node.get_name(), res); - return {res}; + auto fake_quantize = make_shared(inputs, min_adj, max_adj, min_adj, max_adj, levels); + set_node_name(node.get_name(), fake_quantize); + return {fake_quantize}; } } // namespace op } // namespace tensorflow diff --git a/tests/layer_tests/tensorflow_tests/test_tf_FakeQuantWithMinMaxVars.py b/tests/layer_tests/tensorflow_tests/test_tf_FakeQuantWithMinMaxVars.py new file mode 100644 index 00000000000..73a5dcb832a --- /dev/null +++ b/tests/layer_tests/tensorflow_tests/test_tf_FakeQuantWithMinMaxVars.py @@ -0,0 +1,60 @@ +# Copyright (C) 2018-2023 Intel Corporation +# SPDX-License-Identifier: Apache-2.0 + +import numpy as np +import pytest +import tensorflow as tf +from common.tf_layer_test_class import CommonTFLayerTest + + +class TestFakeQuantWithMinMaxVars(CommonTFLayerTest): + def _prepare_input(self, inputs_info): + # generate elements so that the input tensor may contain repeating elements + assert 'inputs' in inputs_info, "Test error: inputs_info must contain `input`" + inputs_shape = inputs_info['inputs'] + inputs_data = {} + inputs_data['inputs'] = np.random.randint(-10, 10, inputs_shape).astype(np.float32) + return inputs_data + + def create_fake_quant_with_min_max_vars_net(self, inputs_shape, min_value, max_value, num_bits, narrow_range, + fake_quant_op): + tf.compat.v1.reset_default_graph() + with tf.compat.v1.Session() as sess: + inputs = tf.compat.v1.placeholder(tf.float32, inputs_shape, 'inputs') + min = tf.constant(min_value, dtype=tf.float32) + max = tf.constant(max_value, dtype=tf.float32) + fake_quant_op(inputs=inputs, min=min, max=max, num_bits=num_bits, + narrow_range=narrow_range) + tf.compat.v1.global_variables_initializer() + tf_net = sess.graph_def + + return tf_net, None + + test_basic = [ + # test FakeQuantWithMinMaxVars + dict(inputs_shape=[2, 6, 4], min_value=-3, max_value=4, num_bits=None, narrow_range=None, + fake_quant_op=tf.raw_ops.FakeQuantWithMinMaxVars), + dict(inputs_shape=[3, 2, 1, 5], min_value=-4, max_value=5, num_bits=14, narrow_range=True, + fake_quant_op=tf.raw_ops.FakeQuantWithMinMaxVars), + dict(inputs_shape=[3, 2, 4], min_value=2, max_value=4, num_bits=10, narrow_range=False, + fake_quant_op=tf.raw_ops.FakeQuantWithMinMaxVars), + dict(inputs_shape=[1, 2, 3], min_value=-6, max_value=-3, num_bits=8, narrow_range=True, + fake_quant_op=tf.raw_ops.FakeQuantWithMinMaxVars), + + # test FakeQuantWithMinMaxVarsPerChannel + pytest.param(dict(inputs_shape=[2, 6, 4], min_value=[-4, -3, -5, -8], max_value=[4, 7, 9, 5], num_bits=None, + narrow_range=None, + fake_quant_op=tf.raw_ops.FakeQuantWithMinMaxVarsPerChannel), + marks=pytest.mark.xfail(reason="104822")) + + ] + + @pytest.mark.parametrize("params", test_basic) + @pytest.mark.precommit_tf_fe + @pytest.mark.nightly + def test_fake_quant_with_min_max_vars_basic(self, params, ie_device, precision, ir_version, temp_dir, + use_new_frontend, + use_old_api): + self._test(*self.create_fake_quant_with_min_max_vars_net(**params), + ie_device, precision, ir_version, temp_dir=temp_dir, + use_new_frontend=use_new_frontend, use_old_api=use_old_api)