Remove constant for Gemm and Add node for A if A is an constant in Gemm (#472)

* 1 . remove constant for Gemm
2. add node for A if A is an constant

* re-work on PR#472 to fit the convention

Author: edis219

onnx_coreml/_operators_nd.py

@@ -674,11 +674,21 @@ def _convert_gemm(builder, node, graph, err):
     transB = node.attrs.get('transB', False)
 
     A = node.inputs[0]
+    if A in node.input_tensors:
+        A_tensor = node.input_tensors[A]
+        builder.add_load_constant_nd(
+            name=node.name + A + "_const",
+            output_name='const_' + A,
+            constant_value=A_tensor,
+            shape=A_tensor.shape
+        )
+        A = 'const_'+A
+
     if alpha != 1.0:
         builder.add_load_constant_nd(
             name=node.name + '_load_alpha',
             output_name='alpha_for_'+A,
-            constant_value=alpha,
+            constant_value=np.array([alpha]),
             shape=[1]
         )
         builder.add_multiply_broadcastable(
@@ -715,7 +725,7 @@ def _convert_gemm(builder, node, graph, err):
             builder.add_load_constant_nd(
                 name=node.name + '_load_beta',
                 output_name='beta_for_'+B,
-                constant_value=beta,
+                constant_value=np.array([beta]),
                 shape=[1]
             )
             builder.add_multiply_broadcastable(

onnx_coreml/_transformers.py

@@ -284,7 +284,7 @@ def __call__(self, graph):  # type: (Graph) -> Graph
             if any([s == 0 for s in shape]):
                 continue
 
-            reshaped_tensor = tensor.reshape(shape)
+            reshaped_tensor = tensor.reshape(shape.astype(int))
 
             for child in node.children:
                 child.parents.remove(node)
@@ -760,6 +760,21 @@ def __call__(self, graph):  # type: (Graph) -> Graph
                     axes = node.attrs.get('axes', None)
                     output = np.squeeze(x, axis = tuple(axes)) 
                 transformation_performed = True
+            elif node.op_type == 'Gemm':
+                alpha = node.attrs.get('alpha', 1.0)
+                beta = node.attrs.get('beta', 1.0)
+                transA = node.attrs.get('transA', False)
+                transB = node.attrs.get('transB', False)
+
+                A_tensor = node.input_tensors[node.inputs[0]]
+                B_tensor = node.input_tensors[node.inputs[1]]
+                C_tensor = node.input_tensors[node.inputs[2]]
+
+                A_tensor = np.transpose(A_tensor) if transA else A_tensor
+                B_tensor = np.transpose(B_tensor) if transB else B_tensor
+
+                output = alpha * np.dot(A_tensor, B_tensor) + beta * C_tensor
+                transformation_performed = True
 
             if transformation_performed:
                 nodes_to_be_removed.append(node)