Misc fixes for adding keras2onnx tests

tf2onnx/onnx_opset/reduction.py

@@ -120,6 +120,9 @@ def version_6(cls, ctx, node, **kwargs):
             reduce_dim = [reduce_dim]
 
         if ctx.opset < 11:
+            inp_rank = ctx.get_rank(node.input[0])
+            if inp_rank is not None:
+                reduce_dim = [d + inp_rank if d < 0 else d for d in reduce_dim]
             utils.make_sure(all(i >= 0 for i in reduce_dim), "negative reduce axis is not supported in onnx for now")
 
         cast = ctx.make_node(op_type="Cast", inputs=[node.input[0]], attr={"to": onnx_pb.TensorProto.FLOAT})

tf2onnx/onnx_opset/tensor.py

@@ -856,7 +856,8 @@ def version_1(cls, ctx, node, **kwargs):
             node = GraphBuilder(ctx).make_slice(
                 kwargs, name=node.name, dtypes=out_dtypes, shapes=out_shapes, return_node=True)
         else:
-            node = ctx.make_node("Identity", [node.input[0]], name=node.name, dtypes=out_dtypes, shapes=out_shapes)
+            node = ctx.make_node("Identity", [node.input[0]], name=node.name, outputs=node.output,
+                                 dtypes=out_dtypes, shapes=out_shapes)
         nodes = [node]
         if needs_squeeze:
             # insert_new_node_on_output(self, op_type, output_name=None, name=None, inputs=None, domain=None, **kwargs)

tf2onnx/rewriter/dropout_rewriter.py

@@ -62,10 +62,10 @@ def rewrite_dropout(g, ops):
             ratio = input3.get_tensor_value()
 
             if input2.inputs[0].is_scalar():
-                data = input2.inputs[1]
+                data_output = input2.input[1]
                 scaling_constant = input2.inputs[0].get_tensor_value()
             elif input2.inputs[1].is_scalar():
-                data = input2.inputs[0]
+                data_output = input2.input[0]
                 scaling_constant = input2.inputs[1].get_tensor_value()
             else:
                 logger.warning("Could not find scaling constant for dropout pattern rooted at %s. "
@@ -89,12 +89,12 @@ def rewrite_dropout(g, ops):
             out_name = utils.port_name(op_name)
             new_node = g.make_node(
                 "Dropout",
-                inputs=[data.output[0]],
+                inputs=[data_output],
                 outputs=[out_name],
                 name=op_name,
                 attr={"ratio": ratio},
-                shapes=[g.get_shape(data.output[0])],
-                dtypes=[g.get_dtype(data.output[0])]
+                shapes=[g.get_shape(data_output)],
+                dtypes=[g.get_dtype(data_output)]
             )
             g.replace_all_inputs(outputs.output[0], new_node.output[0], ops=ops)
             for n in nodes_to_remove:

tf2onnx/rewriter/random_uniform.py

@@ -72,6 +72,7 @@ def create_onnx_random_uniform_op(g, tmax, tmin, ru_op, output, to_delete):
     dtype = g.get_dtype(output.output[0])
     op_name = utils.make_name("RandomUniform")
     shape_node = ru_op.inputs[0]
+    shape_node_output = ru_op.input[0]
     shape = g.get_shape(output.output[0])
     if shape_node.is_const():
         # if the tensorflow input (aka the shape) is const we can use the RandomUniform op
@@ -103,7 +104,7 @@ def create_onnx_random_uniform_op(g, tmax, tmin, ru_op, output, to_delete):
             to_delete.remove(shape_node)
             # create a fill op with the shape of the value of the input tensor
             zero = g.make_const(utils.make_name("zero"), np.zeros((), dtype=np.float32))
-            fill_node = g.make_node("Fill", inputs=[shape_node.output[0], zero.name],
+            fill_node = g.make_node("Fill", inputs=[shape_node_output, zero.name],
                                     shapes=[shape], dtypes=[dtype])
             func, _ = handler.tf_op.find_effective_op("Fill")
             func(g, fill_node)