update PR

Author: sdpython

tests/test_backend.py

@@ -69,6 +69,7 @@
     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
@@ -88,6 +89,7 @@
     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
@@ -3352,6 +3354,19 @@ 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):
+        x_val = np.random.random(size=[4, 5]).astype(np.float32) * 2048. - 1024.
+        def func(x):
+            ret = fake_quant_with_min_max_args(
+                x, min=-1024, max=1024, num_bits=8, narrow_range=False, name=None)
+            return tf.identity(ret, name=_TFOUTPUT)
+        self._run_test_case(func, [_OUTPUT], {_INPUT: x_val})
+
 
 if __name__ == '__main__':
+    cl = BackendTests()
+    cl.setUp()
+    cl.test_fakequant_with_min_max()
     unittest_main()

tests/test_quantization.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/__init__.py

@@ -0,0 +1,61 @@
+# 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 sys
+
+import numpy as np
+from onnx import onnx_pb
+from onnx.onnx_pb import TensorProto
+
+from tf2onnx import constants, utils
+from tf2onnx.graph_builder import GraphBuilder
+from tf2onnx.handler import tf_op
+from tf2onnx.onnx_opset import nn, math
+
+logger = logging.getLogger(__name__)
+
+
+# pylint: disable=unused-argument,missing-docstring,unused-variable,pointless-string-statement,invalid-name
+
+
+@tf_op("FakeQuantWithMinMaxVars")
+class FakeQuantWithMinMaxVars:
+    @classmethod
+    def version_11(cls, ctx, node, **kwargs):
+        # hack to make up for the missing onnx pack op
+        import pprint
+        pprint.pprint(node)
+        amin = node.get_attr("min").i
+        if axis < 0:
+            axis += len(ctx.get_shape(node.input[0])) + 1
+
+        inputs = []
+        dtype = None
+        # insert Unsqueeze on each input
+        for i, n in enumerate(node.inputs):
+            dtype = ctx.get_dtype(node.input[i])
+            shape = ctx.get_shape(node.input[i])
+            new_node = ctx.make_node("Unsqueeze", [node.input[i]], op_name_scope=node.name, attr={"axes": [axis]},
+                                     shapes=[shape], dtypes=[dtype])
+            output_name = new_node.output[0]
+            node.input[i] = output_name
+            inputs.append(output_name)
+
+        shapes = node.output_shapes
+        dtypes = node.output_dtypes
+        ctx.remove_node(node.name)
+        # concat all unqueezes
+        concat = ctx.make_node("Concat", inputs, op_name_scope=node.name, attr={"axis": axis},
+                               shapes=shapes, dtypes=dtypes)
+        ctx.replace_all_inputs(ctx.get_nodes(), node.output[0], concat.output[0])
+
+