Added support for FakeQuantWithMinMaxVars (#1301)

Signed-off-by: Tom Wildenhain <[email protected]>

Author: TomWildenhain-Microsoft

tests/test_backend.py

@@ -74,6 +74,7 @@
     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
+    fake_quant_with_min_max_vars = tf.quantization.fake_quant_with_min_max_vars
 elif LooseVersion(tf.__version__) >= "1.13":
     conv2d_backprop_input = tf.compat.v1.nn.conv2d_backprop_input
     conv3d_transpose = tf.compat.v1.nn.conv3d_transpose
@@ -95,6 +96,7 @@
     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
+    fake_quant_with_min_max_vars = tf.compat.v1.quantization.fake_quant_with_min_max_vars
 else:
     conv2d_backprop_input = tf.nn.conv2d_backprop_input
     conv3d_transpose = tf.nn.conv3d_transpose
@@ -4357,6 +4359,28 @@ def func_neg(x):
         except ValueError:
             pass
 
+    @check_opset_min_version(10)
+    @check_tf_min_version("1.14")
+    def test_fakequant_with_min_max_vars(self):
+        def func(x):
+            ret = fake_quant_with_min_max_vars(
+                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(9, "atan2")
     def test_atan2(self):
         # Test all possible pairs of pos, neg, zero for x and y.

tf2onnx/onnx_opset/quantize.py

@@ -24,14 +24,22 @@
 # pylint: disable=unused-argument,missing-docstring,unused-variable,pointless-string-statement,invalid-name
 
 
-@tf_op("FakeQuantWithMinMaxArgs")
+@tf_op(["FakeQuantWithMinMaxArgs", "FakeQuantWithMinMaxVars"])
 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
+        if node.type == "FakeQuantWithMinMaxVars":
+            utils.make_sure(node.inputs[1].is_scalar(), "%s node %s requires const scalar value for min",
+                            node.type, node.name)
+            utils.make_sure(node.inputs[2].is_scalar(), "%s node %s requires const scalar value for max",
+                            node.type, node.name)
+            amin = node.inputs[1].get_tensor_value()
+            amax = node.inputs[2].get_tensor_value()
+        else:
+            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
 
@@ -58,10 +66,10 @@ def version_10(cls, ctx, node, **kwargs):
         zero = np.array(-min_adj, dtype=np.uint8)
         make_sure(
             zero == -min_adj,
-            "Cannot convert FakeQuantWithMinMaxArgs with "
+            "Cannot convert %s node %s with "
             "min=%r max=%r numbits=%r because zero_scale=%r "
             "is outside uint8 boundary",
-            amin, amax, num_bits, -min_adj)
+            node.type, node.name, amin, amax, num_bits, -min_adj)
         zero_point = ctx.make_const(
             utils.make_name("{}_zpy".format(node.name)), zero)