Fix failing onnx models

TrainingExtensions/onnx/src/python/aimet_onnx/adaround/adaround_optimizer.py

@@ -150,7 +150,7 @@ def _optimize_rounding(cls, module: ModuleInfo, quantized_input_name,
                                                                           out_data_torch.shape)
 
         attributes = read_attributes_for_op(module)
-        if len(attributes['pad']) > 2:
+        if len(attributes['pads']) > 2:
             logger.info("Skipping the Convolution layer because padding size of 4 is not supported for optimization")
             return
 

TrainingExtensions/onnx/src/python/aimet_onnx/batch_norm_fold.py

@@ -407,6 +407,9 @@ def get_input_output_channels(node: NodeProto, model: ModelProto) -> Tuple[int,
     """
     weight = ParamUtils.get_param(model, node, WEIGHT_INDEX)
     groups = get_node_attribute(node, "group")
+    # If group atttribute does not exist in the node,then default is 1
+    if not groups:
+        groups = 1
     if node.op_type == "Conv":
         num_in_channels = weight.dims[1] * groups
         num_out_channels = weight.dims[0]

TrainingExtensions/onnx/src/python/aimet_onnx/meta/connectedgraph.py

@@ -90,6 +90,7 @@ def __init__(self, model: ModelProto):
         # Maps output to consumer node
         self._input_to_node = {}
         self._get_input_to_node()
+        self._unnamed_op = 0
 
         self.starting_ops = []
         self._branch_count = 0
@@ -216,6 +217,9 @@ def _process_starting_ops(self, op_queue: List):
         """
         input_ops = self._get_input_ops()
         for node in input_ops:
+            if not node.name:
+                node.name = str(node.op_type) + '_unnamed_' + str(self._unnamed_op)
+                self._unnamed_op += 1
             node_name = node.name
             if node_name not in self._ops:
                 op = self._create_ir_op(node)
@@ -364,6 +368,9 @@ def fill_op_product_graph(self):
         # - Index 1 contains the parent node.
         while op_queue:
             child_node, parent_node, connecting_tensor_name = op_queue.pop()
+            if not child_node.name:
+                child_node.name = str(child_node.op_type) + '_unnamed_' + str(self._unnamed_op)
+                self._unnamed_op += 1
             # new module, create op/product and link to parent
             if child_node.name != parent_node.name:
                 self._create_op_if_not_exists(child_node)

TrainingExtensions/onnx/src/python/aimet_onnx/utils.py

@@ -200,7 +200,8 @@ def make_dummy_input(model: ModelProto, dynamic_size: int = 1) -> Dict[str, np.n
             else:
                 # Else, axis has a fixed dimension size stored in dim.dim_value
                 shape.append(dim.dim_value)
-        input_dict[name] = np.random.randn(*shape).astype(mapping.TENSOR_TYPE_TO_NP_TYPE[dtype])
+        if shape:
+            input_dict[name] = np.random.randn(*shape).astype(mapping.TENSOR_TYPE_TO_NP_TYPE[dtype])
     return input_dict