[one-cmds] Allow to specialize shape of input/output tensors during ONNX-Circle conversion

compiler/circle-operator/driver/Driver.cpp

@@ -42,6 +42,7 @@ int entry(int argc, char **argv)
     "circle-operator allows users to retrieve operator information from a Circle model file"};
   arser.add_argument("--name").nargs(0).help("Dump operators name in circle file");
   arser.add_argument("--code").nargs(0).help("Dump operators code in circle file");
+  arser.add_argument("--shapes").nargs(0).help("Dump shapes");
   arser.add_argument("--output_path").help("Save output to file (default output is console)");
   arser.add_argument("circle").help("Circle file to dump");
 
@@ -59,6 +60,7 @@ int entry(int argc, char **argv)
   cirops::DumpOption option;
   option.names = arser["--name"];
   option.codes = arser["--code"];
+  option.shapes = arser["--shapes"];
 
   std::ofstream oFstream;
   std::ostream *oStream = &std::cout;

compiler/circle-operator/src/Dump.cpp

@@ -24,6 +24,29 @@
 namespace
 {
 
+// TODO handle multiple outputs case
+const circle::Tensor *get_output_tensor(mio::circle::Reader &reader, const circle::Operator *op)
+{
+  const auto tensors = reader.tensors();
+  const auto output_tensors = reader.outputs(op);
+  const auto output = output_tensors.at(0);
+  return tensors->Get(output);
+}
+
+void dump_shape(std::ostream &os, const ::flatbuffers::Vector<int32_t> *shape)
+{
+  os << "[";
+  for (uint32_t i = 0; i < shape->size(); ++i)
+  {
+    os << shape->Get(i);
+    if (i < shape->size() - 1)
+    {
+      os << ",";
+    }
+  }
+  os << "]";
+}
+
 void dump_ops(std::ostream &os, mio::circle::Reader &reader, const cirops::DumpOption &option)
 {
   auto ops = reader.operators();
@@ -44,19 +67,19 @@ void dump_ops(std::ostream &os, mio::circle::Reader &reader, const cirops::DumpO
       const auto op_name = reader.opcode_name(op);
       os << op_name;
     }
+
     if (option.names)
     {
-      // TODO multiple outputs?
-      const auto tensors = reader.tensors();
-      const auto output_tensors = reader.outputs(op);
-      const auto output = output_tensors.at(0);
-      const auto tensor = tensors->Get(output);
-      const std::string name = mio::circle::tensor_name(tensor);
-      if (option.codes)
-      {
-        os << ",";
-      }
-      os << name;
+      const std::string name = mio::circle::tensor_name(get_output_tensor(reader, op));
+      os << (option.codes ? "," : "") << name;
+    }
+
+    if (option.shapes)
+    {
+      os << (option.codes || option.names ? "," : "");
+      auto const out_tensor = get_output_tensor(reader, op);
+      dump_shape(os, (nullptr == out_tensor->shape_signature()) ? out_tensor->shape()
+                                                                : out_tensor->shape());
     }
     os << std::endl;
   }

compiler/circle-operator/src/Dump.h

@@ -29,6 +29,7 @@ struct DumpOption
   bool names = false;
   bool codes = false;
   bool all_graphs = false;
+  bool shapes = false;
 };
 
 class DumpOperators

compiler/one-cmds/one-import-onnx

@@ -24,6 +24,7 @@ import os
 import sys
 import tempfile
 import onnx
+from onnx.tools import update_model_dims
 
 # ONNX legalizer is an optional feature
 # It enables conversion of some operations, but in experimental phase for now
@@ -204,6 +205,26 @@ def _get_parser():
                         action='store_true',
                         help='Experimental disable BatchMatMul unfold')
 
+    # set static input shape
+    parser.add_argument(
+        '--input_shapes',
+        type=str,
+        help=
+        'Set static shape for input tensors in comma-separated list format, like \'[1,2,3]\'.'
+        'If the model has multiple inputs, tensor names should be provided as well.'
+        'In such a case, the argument should be in the following format: \'a[1,2,3],b[4,5,6],c[7,8]\'.'
+    )
+
+    # set static output shape
+    parser.add_argument(
+        '--output_shapes',
+        type=str,
+        help=
+        'Set static shape for output tensors in comma-separated list format, like \'[1,2,3]\'.'
+        'If the model has multiple outputs, tensor names should be provided as well.'
+        'In such a case, the argument should be in the following format: \'a[1,2,3],b[4,5,6],c[7,8]\'.'
+    )
+
     return parser
 
 
@@ -299,6 +320,50 @@ def _check_onnx2circle():
     return o2c_path
 
 
+def _extract_origin_tensor_shapes(tensors):
+    shapes_map = {}
+    for tensor in tensors:
+        shapes_map[tensor.name] = []
+        for dim_proto in tensor.type.tensor_type.shape.dim:
+            if dim_proto.HasField("dim_value"):
+                shapes_map[tensor.name].append(dim_proto.dim_value)
+            elif dim_proto.HasField("dim_param"):
+                shapes_map[tensor.name].append(dim_proto.dim_param)
+            else:
+                shapes_map[tensor.name].append(-1)  # dynamic
+    return shapes_map
+
+
+def _parse_shapes(shape_str, origin_tensor_shapes_map):
+    user_shapes_map = {}
+    for idx, single_shape in enumerate(shape_str.split(']')):
+        if single_shape:
+            tensor_name, single_shape = single_shape.split('[')
+            if not tensor_name:
+                # using tensor name model allowed only for single input models
+                if idx == 0 and len(origin_tensor_shapes_map) == 1:
+                    tensor_name = next(iter(origin_tensor_shapes_map.keys()))
+                else:
+                    raise ValueError(
+                        'You must provide tenors name for the model with multiple inputs/outputs'
+                    )
+            tensor_name = tensor_name.replace(',', '')
+            user_shapes_map[tensor_name] = []
+            for dim in single_shape.split(','):
+                user_shapes_map[tensor_name].append(int(dim))
+    for tensor_name in user_shapes_map.keys():
+        if tensor_name not in origin_tensor_shapes_map:
+            raise ValueError(
+                f'Tensor with name={tensor_name} do NOT match any input/output of the model'
+            )
+        if len(user_shapes_map[tensor_name]) != len(
+                origin_tensor_shapes_map[tensor_name]):
+            raise ValueError(
+                f'Rank of provided shape must be compatible with the origin tensor={tensor_name} from the model'
+            )
+    return user_shapes_map
+
+
 def _convert(args):
     _apply_verbosity(args.verbose)
 
@@ -323,6 +388,29 @@ def _convert(args):
 
         onnx_model = onnx.load(onnx_to_convert_path)
         _sanitize_io_names(onnx_model)
+        input_shape_provided = oneutils.is_valid_attr(args, 'input_shapes')
+        output_shape_provided = oneutils.is_valid_attr(args, 'output_shapes')
+        if input_shape_provided or output_shape_provided:  # any shape arg provided
+            input_shapes_map = _extract_origin_tensor_shapes(onnx_model.graph.input)
+            if input_shape_provided:
+                input_shapes_map = _parse_shapes(getattr(args, 'input_shapes'),
+                                                 input_shapes_map)
+                delattr(args, 'input_shapes'
+                        )  # avoid argument colision with tf2tflite conversion step
+            output_shapes_map = _extract_origin_tensor_shapes(onnx_model.graph.output)
+            if output_shape_provided:
+                output_shapes_map = _parse_shapes(getattr(args, 'output_shapes'),
+                                                  output_shapes_map)
+            onnx_model = update_model_dims.update_inputs_outputs_dims(
+                onnx_model, input_shapes_map, output_shapes_map)
+            try:
+                print(onnx.__file__)
+                onnx_model = onnx.shape_inference.infer_shapes(onnx_model,
+                                                               strict_mode=True)
+            except Exception as e:
+                raise RuntimeError(
+                    f'Exception with message="{e}" raised from external onnx library during shape inference'
+                )
         # TODO set model_updated to True only when sanitize updates onnx_model
         # NOTE with onnx2circle, we may not need sanitize(this needs verification)
         model_updated = True