fix fake_quant

Author: sdpython

tests/test_backend.py

@@ -3359,23 +3359,37 @@ def func(base_matrix, diag, k):
     def test_fakequant_with_min_max(self):
         def func(x):
             ret = fake_quant_with_min_max_args(
-                x, min=-1024, max=1024, num_bits=8, narrow_range=False, name=None)
+                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})
-        self._run_test_case(func, [_OUTPUT], {_INPUT: 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 RuntimeError:
+            pass
 
 
 if __name__ == '__main__':
-    #cl = BackendTests()
-    #cl.setUp()
-    #cl.test_fakequant_with_min_max()
-    #import cProfile
-    #cProfile.run('unittest_main()', 'restats')
     unittest_main()
-    #import pstats
-    #from pstats import SortKey
-    #p = pstats.Stats('restats')
-    #p.sort_stats(SortKey.CUMULATIVE).print_stats()    

tf2onnx/onnx_opset/quantize.py

@@ -44,22 +44,8 @@ def version_11(cls, ctx, node, **kwargs):
                 "Unable to convert node FakeQuantWithMinMaxArgs with "
                 "num_bits=%r" % num_bits)
 
-        if 0 < amin < amax:
-            min_adj = 0
-            max_adj = amax - amin
-            scale = 1.
-        elif amin < amax < 0:
-            min_adj = amin - amax
-            max_adj = 0
-            scale = 1.
-        elif amin <= 0 <= amax:
-            scale = (amax - amin) / (2 ** num_bits - 1)
-            min_adj = scale * int(amin / scale)
-            max_adj = amax + min_adj - amin
-        else:
-            raise RuntimeError(
-                "Unable to convert node FakeQuantWithMinMaxArgs with "
-                "min=%f and max=%f" % (amin, amax))
+        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])
@@ -69,9 +55,15 @@ def version_11(cls, ctx, node, **kwargs):
         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)
+        if zero != -min_adj:
+            raise RuntimeError(
+                "Cannot convert FakeQuantWithMinMaxArgs with "
+                "min={} max={} numbits={} because zero_scale={} "
+                "is outside uint8 boundary".format(
+                    amin, amax, num_bits, -min_adj))
         zero_point = ctx.make_const(
-            utils.make_name("{}_zpy".format(node.name)),
-            np.array(min_adj, dtype=np.uint8))
+            utils.make_name("{}_zpy".format(node.name)), zero)
 
         new_node = ctx.make_node(
             "QuantizeLinear", [node.input[0], pb_scale.name, zero_point.name],
@@ -87,4 +79,3 @@ def version_11(cls, ctx, node, **kwargs):
             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])
-