Support ver 7 (#193)

* bug fix and some improvements

* support ver 7 and tile

* yapf

* support bias_add and improve coverage getter

doc/support_status.md

@@ -130,52 +130,52 @@ ______
 | -------------- |:------------------:|
 |abs|1, 6|
 |acos|7|
-|add|1, 6|
+|add|1, 6, 7|
 |add_n|1, 6|
 |arg_max|1|
 |arg_min|1|
 |asin|7|
 |atan|7|
 |avg_pool|1, 7|
-|batch_norm|1, 6|
-|bias_add|1, 6|
+|batch_norm|1, 6, 7|
+|bias_add|1, 6, 7|
 |cast|1, 6|
 |ceil|1, 6|
 |concat_v2|1, 4|
 |conv1_d|1|
 |conv2_d|1|
 |conv3_d|1|
 |cos|7|
-|equal|1|
+|equal|1, 7|
 |exp|1, 6|
 |expand_dims|1|
 |fill|1|
 |floor|1, 6|
-|fused_batch_norm|1, 6|
-|greater|1|
+|fused_batch_norm|1, 6, 7|
+|greater|1, 7|
 |identity|1|
-|less|1|
+|less|1, 7|
 |log|1, 6|
 |log_softmax|1|
-|logical_and|1|
+|logical_and|1, 7|
 |logical_not|1|
-|logical_or|1|
-|logical_xor|1|
+|logical_or|1, 7|
+|logical_xor|1, 7|
 |mat_mul|1|
 |max|1|
 |max_pool|1|
 |maximum|1, 6|
 |mean|1|
 |min|1|
 |minimum|1, 6|
-|mul|1, 6|
+|mul|1, 6, 7|
 |neg|1, 6|
 |pad|1, 2|
-|pow|1|
+|pow|1, 7|
 |prod|1|
 |random_standard_normal|1|
 |random_uniform|1|
-|real_div|1, 6|
+|real_div|1, 6, 7|
 |reciprocal|1, 6|
 |relu|1, 6|
 |reshape|1, 5|
@@ -190,10 +190,10 @@ ______
 |split_v|1, 2|
 |sqrt|1, 6|
 |squeeze|1|
-|sub|1, 6|
+|sub|1, 6, 7|
 |sum|1|
 |tan|7|
 |tanh|1, 6|
-|tile|1|
+|tile|6|
 |top_k_v2|1|
 |transpose|1|

onnx_tf/common/handler_helper.py

@@ -46,6 +46,13 @@ def get_frontend_coverage():
   :return: onnx_coverage: e.g. {'domain': {'ONNX_OP': [versions], ...}, ...}
   tf_coverage: e.g. {'domain': {'TF_OP': [versions], ...}, ...}
   """
+
+  def _update_coverage(coverage, domain, key, versions):
+    domain_coverage = coverage.setdefault(domain, {})
+    vers = domain_coverage.get(key, [])
+    vers.extend(versions)
+    domain_coverage[key] = sorted(list(set(vers)))
+
   tf_coverage = {}
   onnx_coverage = {}
   for handler in FrontendHandler.__subclasses__():
@@ -54,6 +61,6 @@ def get_frontend_coverage():
     versions = handler.get_versions()
     domain = handler.DOMAIN
     for tf_op in handler.TF_OP:
-      tf_coverage.setdefault(domain, {})[op_name_to_lower(tf_op)] = versions
-    onnx_coverage.setdefault(domain, {})[handler.ONNX_OP] = versions
+      _update_coverage(tf_coverage, domain, op_name_to_lower(tf_op), versions)
+    _update_coverage(onnx_coverage, domain, handler.ONNX_OP, versions)
   return onnx_coverage, tf_coverage

onnx_tf/handlers/frontend/add.py

@@ -15,3 +15,7 @@ def version_1(cls, node, **kwargs):
   @classmethod
   def version_6(cls, node, **kwargs):
     return cls.arithmetic_op(node, **kwargs)
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.arithmetic_op(node, **kwargs)

onnx_tf/handlers/frontend/argmax.py

@@ -18,5 +18,4 @@ def args_check(cls, node, **kwargs):
   @classmethod
   def version_1(cls, node, **kwargs):
     axis = np.asscalar(kwargs["consts"][node.inputs[1]])
-    return cls.make_node(
-        node, [node.inputs[0]], axis=axis, keepdims=0)
+    return cls.make_node(node, [node.inputs[0]], axis=axis, keepdims=0)

onnx_tf/handlers/frontend/argmin.py

@@ -18,5 +18,4 @@ def args_check(cls, node, **kwargs):
   @classmethod
   def version_1(cls, node, **kwargs):
     axis = np.asscalar(kwargs["consts"][node.inputs[1]])
-    return cls.make_node(
-        node, [node.inputs[0]], axis=axis, keepdims=0)
+    return cls.make_node(node, [node.inputs[0]], axis=axis, keepdims=0)

onnx_tf/handlers/frontend/batch_normalization.py

@@ -21,3 +21,9 @@ def version_6(cls, node, **kwargs):
         node,
         epsilon=node.attr.get("epsilon", 1e-5),
         is_test=node.attr.get("is_training", 0))
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.make_node(
+        node,
+        epsilon=node.attr.get("epsilon", 1e-5))

onnx_tf/handlers/frontend/bias_add.py

@@ -21,3 +21,7 @@ def version_6(cls, node, **kwargs):
     channel_first = data_format[1] == "C"
     axis = 1 if channel_first else -1
     return cls.arithmetic_op(node, axis=axis, **kwargs)
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.arithmetic_op(node, **kwargs)

onnx_tf/handlers/frontend/control_flow_mixin.py

@@ -4,7 +4,14 @@
 class LogicalMixin(object):
 
   @classmethod
-  def logical_op(cls, node, broadcast=1, **kwargs):
+  def logical_op(cls, node, **kwargs):
+    if cls.SINCE_VERSION <= 6:
+      return cls._limited_broadcast(node, **kwargs)
+    else:  # since_version >= 7
+      return cls._np_broadcast(node, **kwargs)
+
+  @classmethod
+  def _limited_broadcast(cls, node, broadcast=1, **kwargs):
     ex_kwargs = {}
     if broadcast == 1:
       ex_kwargs["broadcast"] = 1
@@ -14,16 +21,29 @@ def logical_op(cls, node, broadcast=1, **kwargs):
         ex_kwargs["axis"] = axis
     return cls.make_node(node, **ex_kwargs)
 
+  @classmethod
+  def _np_broadcast(cls, node, **kwargs):
+    return cls.make_node(node)
+
 
 class ComparisonMixin(object):
 
   @classmethod
-  def comparison_op(cls, node, broadcast=1, **kwargs):
-    ex_kwargs = {}
-    if broadcast == 1:
-      ex_kwargs["broadcast"] = 1
-      node_dict = kwargs["node_dict"]
-      axis = get_broadcast_axis(*[node_dict[x] for x in node.inputs[0:2]])
-      if axis is not None:
-        ex_kwargs["axis"] = axis
+  def comparison_op(cls, node, **kwargs):
+    if cls.SINCE_VERSION <= 6:
+      return cls._limited_broadcast(node, **kwargs)
+    else:  # since_version >= 7
+      return cls._np_broadcast(node, **kwargs)
+
+  @classmethod
+  def _limited_broadcast(cls, node, **kwargs):
+    ex_kwargs = {"broadcast": 1}
+    node_dict = kwargs["node_dict"]
+    axis = get_broadcast_axis(*[node_dict[x] for x in node.inputs[0:2]])
+    if axis is not None:
+      ex_kwargs["axis"] = axis
     return cls.make_node(node, **ex_kwargs)
+
+  @classmethod
+  def _np_broadcast(cls, node, **kwargs):
+    return cls.make_node(node)

onnx_tf/handlers/frontend/div.py

@@ -15,3 +15,7 @@ def version_1(cls, node, **kwargs):
   @classmethod
   def version_6(cls, node, **kwargs):
     return cls.arithmetic_op(node, **kwargs)
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.arithmetic_op(node, **kwargs)

onnx_tf/handlers/frontend/equal.py

@@ -11,3 +11,7 @@ class Equal(ComparisonMixin, FrontendHandler):
   @classmethod
   def version_1(cls, node, **kwargs):
     return cls.comparison_op(node, **kwargs)
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.comparison_op(node, **kwargs)

onnx_tf/handlers/frontend/fill.py

@@ -18,5 +18,4 @@ def args_check(cls, node, **kwargs):
   @classmethod
   def version_1(cls, node, **kwargs):
     value = float(np.asscalar(kwargs["consts"][node.inputs[1]]))
-    return cls.make_node(
-        node, [node.inputs[0]], input_as_shape=1, value=value)
+    return cls.make_node(node, [node.inputs[0]], input_as_shape=1, value=value)

onnx_tf/handlers/frontend/greater.py

@@ -11,3 +11,7 @@ class Greater(ComparisonMixin, FrontendHandler):
   @classmethod
   def version_1(cls, node, **kwargs):
     return cls.comparison_op(node, **kwargs)
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.comparison_op(node, **kwargs)

onnx_tf/handlers/frontend/less.py

@@ -11,3 +11,7 @@ class Less(ComparisonMixin, FrontendHandler):
   @classmethod
   def version_1(cls, node, **kwargs):
     return cls.comparison_op(node, **kwargs)
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.comparison_op(node, **kwargs)

onnx_tf/handlers/frontend/logical_and.py

@@ -11,3 +11,7 @@ class LogicalAnd(LogicalMixin, FrontendHandler):
   @classmethod
   def version_1(cls, node, **kwargs):
     return cls.logical_op(node, **kwargs)
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.logical_op(node, **kwargs)

onnx_tf/handlers/frontend/logical_or.py

@@ -11,3 +11,7 @@ class LogicalOr(LogicalMixin, FrontendHandler):
   @classmethod
   def version_1(cls, node, **kwargs):
     return cls.logical_op(node, **kwargs)
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.logical_op(node, **kwargs)

onnx_tf/handlers/frontend/logical_xor.py

@@ -11,3 +11,7 @@ class LogicalXor(LogicalMixin, FrontendHandler):
   @classmethod
   def version_1(cls, node, **kwargs):
     return cls.logical_op(node, **kwargs)
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.logical_op(node, **kwargs)

onnx_tf/handlers/frontend/math_mixin.py

@@ -13,6 +13,13 @@ class ArithmeticMixin(object):
 
   @classmethod
   def arithmetic_op(cls, node, **kwargs):
+    if cls.SINCE_VERSION <= 6:
+      return cls._limited_broadcast(node, **kwargs)
+    else:  # since_version >= 7
+      return cls._np_broadcast(node, **kwargs)
+
+  @classmethod
+  def _limited_broadcast(cls, node, **kwargs):
     node_dict = kwargs["node_dict"]
     axis = kwargs.get(
         "axis", get_broadcast_axis(*[node_dict[x] for x in node.inputs[0:2]]))
@@ -21,6 +28,10 @@ def arithmetic_op(cls, node, **kwargs):
       ex_kwargs["axis"] = axis
     return cls.make_node(node, broadcast=1, **ex_kwargs)
 
+  @classmethod
+  def _np_broadcast(cls, node, **kwargs):
+    return cls.make_node(node)
+
 
 class ReductionMixin(object):
 

onnx_tf/handlers/frontend/multiply.py

@@ -15,3 +15,7 @@ def version_1(cls, node, **kwargs):
   @classmethod
   def version_6(cls, node, **kwargs):
     return cls.arithmetic_op(node, **kwargs)
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.arithmetic_op(node, **kwargs)

onnx_tf/handlers/frontend/pow.py

@@ -11,3 +11,7 @@ class Pow(BasicMathMixin, FrontendHandler):
   @classmethod
   def version_1(cls, node, **kwargs):
     return cls.basic_math_op(node, **kwargs)
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.basic_math_op(node, **kwargs)

onnx_tf/handlers/frontend/subtract.py

@@ -15,3 +15,7 @@ def version_1(cls, node, **kwargs):
   @classmethod
   def version_6(cls, node, **kwargs):
     return cls.arithmetic_op(node, **kwargs)
+
+  @classmethod
+  def version_7(cls, node, **kwargs):
+    return cls.arithmetic_op(node, **kwargs)

onnx_tf/handlers/frontend/tile.py

@@ -10,7 +10,7 @@
 class Tile(FrontendHandler):
 
   @classmethod
-  def version_1(cls, node, **kwargs):
+  def version_6(cls, node, **kwargs):
     data_type_cast_map = kwargs["data_type_cast_map"]
     data_type_cast_map[node.inputs[1]] = TensorProto.INT64
     return cls.make_node(node)