Merge pull request #1009 from xadupre/i500qu

Fixes #500, replaces operator FakeQuantWithMinMaxVars

Author: guschmue

tests/test_backend.py

@@ -70,6 +70,7 @@
     is_inf = tf.math.is_inf
     floormod = tf.math.floormod
     matrix_diag_part = tf.compat.v1.matrix_diag_part
+    fake_quant_with_min_max_args = tf.quantization.fake_quant_with_min_max_args
 elif LooseVersion(tf.__version__) >= "1.13":
     conv2d_backprop_input = tf.compat.v1.nn.conv2d_backprop_input
     multinomial = tf.compat.v1.random.multinomial
@@ -89,6 +90,7 @@
     is_inf = tf.math.is_inf
     floormod = tf.floormod
     matrix_diag_part = tf.compat.v1.matrix_diag_part
+    fake_quant_with_min_max_args = tf.compat.v1.quantization.fake_quant_with_min_max_args
 else:
     conv2d_backprop_input = tf.nn.conv2d_backprop_input
     multinomial = tf.multinomial
@@ -3353,6 +3355,42 @@ def func(base_matrix, diag, k):
 
         self._run_test_case(func, [_OUTPUT], {_INPUT: input_val, _INPUT1: diag_val, _INPUT2: k_val})
 
+    @check_opset_min_version(10)
+    @check_tf_min_version("1.14")
+    def test_fakequant_with_min_max(self):
+        def func(x):
+            ret = fake_quant_with_min_max_args(
+                x, min=-1024, max=1023, num_bits=8, narrow_range=False, name=None)
+            return tf.identity(ret, name=_TFOUTPUT)
+
+        x_val = np.random.random(size=[4, 3]).astype(np.float32) * 2048. - 1024.
+        x_val0 = np.abs(x_val)
+        self._run_test_case(func, [_OUTPUT], {_INPUT: x_val0}, rtol=1e-6, atol=1e-4)
+        self._run_test_case(func, [_OUTPUT], {_INPUT: x_val}, rtol=1e-6, atol=1e-4)
+
+        x_val = np.random.random(size=[4, 3]).astype(np.float32) * 2048. - 1024
+        x_val[0, 0] = -1024
+        x_val[0, 1] = -1023
+        x_val[0, 2] = 1024
+        x_val[1, 0] = 1023
+        x_val[1, 1] = 1025
+        x_val[1, 2] = -1025
+        self._run_test_case(func, [_OUTPUT], {_INPUT: x_val}, rtol=1e-6, atol=1e-4)
+
+    @check_opset_min_version(10)
+    @check_tf_min_version("1.14")
+    def test_fakequant_with_min_max_same_sign(self):
+        def func_neg(x):
+            ret = fake_quant_with_min_max_args(
+                x, min=-1024*3, max=-1024, num_bits=8, narrow_range=False, name=None)
+            return tf.identity(ret, name=_TFOUTPUT)
+
+        x_val = np.random.random(size=[4, 3]).astype(np.float32) * 2048. - 1024 * 3.
+        try:
+            self._run_test_case(func_neg, [_OUTPUT], {_INPUT: x_val}, rtol=1e-6, atol=1e-4)
+        except ValueError:
+            pass
+
     @check_opset_min_version(9, "atan2")
     def test_atan2(self):
         # Test all possible pairs of pos, neg, zero for x and y.

tf2onnx/onnx_opset/__init__.py

@@ -2,4 +2,4 @@
 # Licensed under the MIT license.
 """tf2onnx.onnx_opset module"""
 
-from . import common, controlflow, generator, logical, math, misc, nn, reduction, rnn, tensor, traditionalml
+from . import common, controlflow, generator, logical, math, misc, nn, quantize, reduction, rnn, tensor, traditionalml

tf2onnx/onnx_opset/quantize.py

@@ -0,0 +1,81 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT license.
+
+"""
+tensor
+"""
+
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+
+import logging
+
+import numpy as np
+from onnx.onnx_pb import TensorProto
+
+from tf2onnx import utils
+from tf2onnx.handler import tf_op
+from tf2onnx.utils import make_sure
+
+logger = logging.getLogger(__name__)
+
+
+# pylint: disable=unused-argument,missing-docstring,unused-variable,pointless-string-statement,invalid-name
+
+
+@tf_op("FakeQuantWithMinMaxArgs")
+class FakeQuantWithMinMaxArgs:
+    # see https://www.tensorflow.org/api_docs/cc/class/tensorflow/ops/fake-quant-with-min-max-args
+    @classmethod
+    def version_10(cls, ctx, node, **kwargs):
+        # hack to make up for the missing onnx pack op
+        amin = node.get_attr("min").f
+        amax = node.get_attr("max").f
+        narrow_range = node.get_attr("narrow_range").i
+        num_bits = node.get_attr("num_bits").i
+
+        make_sure(
+            not narrow_range,
+            "Unable to convert node FakeQuantWithMinMaxArgs with narrow_range=%r",
+            narrow_range)
+        make_sure(
+            num_bits == 8,
+            "Unable to convert node FakeQuantWithMinMaxArgs with "
+            "num_bits=%r", num_bits)
+
+        scale = (amax - amin) / (2 ** num_bits - 1)
+        min_adj = np.around(amin / scale)
+
+        dtype = ctx.get_dtype(node.input[0])
+        shape = ctx.get_shape(node.input[0])
+        axis = 1
+        idtype = TensorProto.UINT8
+
+        pb_scale = ctx.make_const(
+            utils.make_name("{}_scaley".format(node.name)),
+            np.array(scale, dtype=np.float32))
+        zero = np.array(-min_adj, dtype=np.uint8)
+        make_sure(
+            zero == -min_adj,
+            "Cannot convert FakeQuantWithMinMaxArgs with "
+            "min=%r max=%r numbits=%r because zero_scale=%r "
+            "is outside uint8 boundary",
+            amin, amax, num_bits, -min_adj)
+        zero_point = ctx.make_const(
+            utils.make_name("{}_zpy".format(node.name)), zero)
+
+        new_node = ctx.make_node(
+            "QuantizeLinear", [node.input[0], pb_scale.name, zero_point.name],
+            op_name_scope=node.name, attr={"axis": axis},
+            shapes=[shape], dtypes=[idtype])
+        output_name = new_node.output[0]
+        node.input[0] = output_name
+
+        ctx.remove_node(node.name)
+
+        last_node = ctx.make_node(
+            "DequantizeLinear", [new_node.output[0], pb_scale.name, zero_point.name],
+            op_name_scope=node.name, attr={"axis": axis},
+            shapes=[shape], dtypes=[dtype])
+        ctx.replace_all_inputs(ctx.get_nodes(), node.output[0], last_node.output[0])