[FRONTEND][ONNX] Some bug fixes and Shape operator fixed for relay. (#2850)

* [FRONTEND][ONNX] Some bug fixes and Shape operator fixed for relay.

* 	* test cases

* 	* ci error

python/tvm/relay/frontend/common.py

@@ -321,6 +321,10 @@ class AttrCvt(object):
         else:
             assert callable(self._op_name), "op_name can either be string or callable"
             op_name = self._op_name(attrs)
+
+        # ignore 'tvm_custom' always
+        self._ignores.append('tvm_custom')
+
         # convert attributes
         new_attrs = {}
         for k in attrs.keys():
@@ -329,7 +333,8 @@ class AttrCvt(object):
             elif k in self._disables:
                 logging.warning("Attribute %s is disabled in relay.sym.%s", k, op_name)
             elif k in self._ignores:
-                logging.debug("Attribute %s is ignored in relay.sym.%s", k, op_name)
+                if k != 'tvm_custom':
+                    logging.warning("Attribute %s is ignored in relay.sym.%s", k, op_name)
             elif k in self._transforms:
                 new_name, defaults, transform = self._parse_default(self._transforms[k])
                 if defaults is None:
@@ -416,4 +421,6 @@ class Renamer(object):
         self._new_name = new_name
 
     def __call__(self, inputs, attrs, *args):
+        if 'tvm_custom' in attrs:
+            attrs.pop('tvm_custom')
         return get_relay_op(self._new_name)(*inputs, **attrs)

python/tvm/relay/frontend/onnx.py

@@ -106,7 +106,7 @@ class Pool(OnnxOpConverter):
                 'pads': ('padding', (0, 0), revert_caffe2_pad)
             },
             # very weird attributes here in onnx, force check
-            ignores=['dilations'],
+            ignores=['dilations', 'auto_pad'],
             # TODO(zhreshold): make sure ceil_mode in onnx, and layout?
             extras={'ceil_mode': False},
             custom_check=dimension_constraint())(inputs, attr, params)
@@ -160,6 +160,7 @@ class Conv(OnnxOpConverter):
                           'dilations': ('dilation', (0, 0)),
                           'pads': ('padding', (0, 0), revert_caffe2_pad),
                           'group': ('groups', 1)},
+                      ignores=['auto_pad'],
                       custom_check=dimension_constraint())(inputs[:2], attr, params)
         use_bias = len(inputs) == 3
         if use_bias:
@@ -332,7 +333,21 @@ class Reshape(OnnxOpConverter):
             shape = tuple(params[inputs[1].name_hint].asnumpy())
             out = _op.reshape(inputs[0], shape)
         else:
-            out = _op.reshape_like(inputs[0], inputs[1])
+            # Try to infer shape by precompute prune if possible.
+            # TODO: good to check inputs to be in params.
+            #       to be enhanced when relay support list_input_names API of NNVM
+            logging.warning("Infering Reshape argument by precompute")
+            func = _expr.Function(ir_pass.free_vars(inputs[1]), inputs[1])
+            with tvm.relay.build_config(opt_level=0):
+                graph, lib, params = tvm.relay.build(func, target="llvm", params=params)
+            ctx = tvm.context("llvm", 0)
+            from tvm.contrib import graph_runtime
+            m = graph_runtime.create(graph, lib, ctx)
+            m.set_input(**params)
+            m.run()
+            params_new = m.get_output(0)
+            inputs.pop(1)
+            out = _op.reshape(inputs[0], tuple(params_new.asnumpy().astype('int32').flatten()))
 
         return out
 
@@ -477,10 +492,7 @@ class Shape(OnnxOpConverter):
 
     @classmethod
     def _impl_v1(cls, inputs, attr, params):
-        # Result of this operator is prominently used by reshape operator.
-        # Just pass the input as it is so that reshape_like can be used there.
-        logging.warning("Shape: Differently implemented in relay as a bypass (dummy operator)")
-        return inputs[0]
+        return _op.shape_of(inputs[0])
 
 class Cast(OnnxOpConverter):
     """ Operator converter for Cast.
@@ -494,7 +506,7 @@ class Cast(OnnxOpConverter):
     def _impl_v5(cls, inputs, attr, params):
         try:
             from onnx.mapping import TENSOR_TYPE_TO_NP_TYPE
-            attr['to'] = TENSOR_TYPE_TO_NP_TYPE[attr['to']]
+            attr['to'] = str(TENSOR_TYPE_TO_NP_TYPE[attr['to']])
         except ImportError as e:
             raise ImportError(
                 "Unable to import onnx.mapping which is required {}".format(e))
@@ -674,6 +686,11 @@ class ReduceMean(Reduce):
     """
     name = 'mean'
 
+class ReduceProd(Reduce):
+    """ Operator converter for ArgMax.
+    """
+    name = 'prod'
+
 class ArgMax(OnnxOpConverter):
     """ Operator converter for ArgMax.
     """
@@ -826,6 +843,7 @@ def _get_convert_map(opset):
         'ReduceMin': ReduceMin.get_converter(opset),
         'ReduceSum': ReduceSum.get_converter(opset),
         'ReduceMean': ReduceMean.get_converter(opset),
+        'ReduceProd': ReduceProd.get_converter(opset),
         # 'ReduceProd'
         # 'ReduceLogSumExp'
         'ArgMax': ArgMax.get_converter(opset),
@@ -842,8 +860,7 @@ def _get_convert_map(opset):
         'Squeeze': AttrCvt('squeeze', {'axes': 'axis'}),
         'Unsqueeze': Unsqueeze.get_converter(opset),
         'Pad': Pad.get_converter(opset),
-        # TODO(zhreshold) Shape op is implemented as bypass op in relay
-        # 'Shape': Shape.get_converter(opset),
+        'Shape': Shape.get_converter(opset),
     }
 
 
@@ -883,6 +900,7 @@ class GraphProto(object):
         ----------
         graph : onnx protobuf object
             The loaded onnx graph
+
         opset : opset version
 
         Returns
@@ -911,12 +929,12 @@ class GraphProto(object):
                                               dtype=self._params[i_name].dtype)
             else:
                 self._num_input += 1
-                shape = self._shape[i_name] if i_name in self._shape else ()
+                tshape = self._shape[i_name] if i_name in self._shape else ()
                 if isinstance(self._dtype, dict):
                     dtype = self._dtype[i_name] if i_name in self._dtype else d_type
                 else:
                     dtype = d_type
-                self._nodes[i_name] = new_var(i_name, shape=shape, dtype=dtype)
+                self._nodes[i_name] = new_var(i_name, shape=tshape, dtype=dtype)
         # construct nodes, nodes are stored as directed acyclic graph
         for node in graph.node:
             op_name = node.op_type
@@ -936,6 +954,10 @@ class GraphProto(object):
                     self._nodes[i_name] = new_var(node.output[0], shape=(), dtype=dtype)
                     inputs.append(self._nodes[i_name])
 
+                i_name = self._parse_value_proto(node)
+                attr['tvm_custom'] = {}
+                attr['tvm_custom']['name'] = i_name
+
                 op = self._convert_operator(op_name, inputs, attr, opset)
                 node_output = self._fix_outputs(op_name, node.output)
                 if not isinstance(op, _expr.TupleWrapper):

tests/python/frontend/onnx/test_forward.py

@@ -113,35 +113,36 @@ def test_reshape():
 
     tvm.testing.assert_allclose(ref_shape, tvm_out.shape)
 
-def test_reshape_like():
+def test_shape():
     in_shape = (4, 3, 3, 4)
-    ref_shape = (3, 4, 4, 3)
+    ref_shape = (6, 2, 4, 3)
 
-    ref_array = np.random.uniform(size=ref_shape).astype('float32')
+    ref_array = np.array(ref_shape)
     ref_node = onnx.helper.make_node('Constant',
                                  inputs=[],
                                  outputs=['ref_in'],
                                  value=onnx.helper.make_tensor(name = 'const_tensor',
-                                                               data_type = onnx.TensorProto.FLOAT,
+                                                               data_type = onnx.TensorProto.INT32,
                                                                dims = ref_array.shape,
-                                                               vals = ref_array.flatten().astype(float)))
-    copy_node = helper.make_node("Identity", ["ref_in"], ["copy_in"])
-    reshape_node = helper.make_node("Reshape", ["in", "copy_in"], ["out"])
+                                                               vals = ref_array.flatten().astype(int)))
+    reshape_node = helper.make_node("Reshape", ["in", "ref_in"], ["out"])
+
+    shape_node = helper.make_node("Shape", ['out'], ['final_out'])
 
-    graph = helper.make_graph([ref_node, copy_node, reshape_node],
-                              "reshape_like_test",
+    graph = helper.make_graph([ref_node, reshape_node, shape_node],
+                              "shape_test",
                               inputs = [helper.make_tensor_value_info("in",
                                             TensorProto.FLOAT, list(in_shape))],
-                              outputs = [helper.make_tensor_value_info("out",
+                              outputs = [helper.make_tensor_value_info("final_out",
                                             TensorProto.FLOAT, list(ref_shape))])
 
-    model = helper.make_model(graph, producer_name='reshape_like_test')
+    model = helper.make_model(graph, producer_name='shape_test')
 
     for target, ctx in ctx_list():
-        x = np.random.uniform(size=in_shape).astype('float32')
-        tvm_out = get_tvm_output(model, x, target, ctx, ref_shape, 'float32')
+        x = np.random.uniform(size=in_shape).astype('int32')
+        tvm_out = get_tvm_output(model, x, target, ctx, ref_shape, 'int32')
 
-    tvm.testing.assert_allclose(ref_shape, tvm_out.shape)
+    tvm.testing.assert_allclose(ref_shape, tvm_out)
 
 def _test_power_iteration(x_shape, y_shape):
     if isinstance(y_shape, int):
@@ -995,7 +996,7 @@ def test_LogSoftmax():
 
 if __name__ == '__main__':
     test_reshape()
-    test_reshape_like()
+    test_shape()
     test_power()
     test_squeeze()
     test_unsqueeze()