[ONNX] [Relay] Update unique operator to match ONNX output (1D only) (#…

…8099)

* Fix topi test case and docs (tvm was returning inverse_indices and claiming it was indices)

* Passes on CPU, fix unique op test

* more changes

* mtrying to fix optional outputs in onnx importer

* TupleGetItem is being passed a stringgit add python/tvm/relay/frontend/onnx.py debugging print statements

* Unique is passing onnx unit tests

* fix indices

* change comment

* fix return of compute unique

* black

* fix lint

* Some stray .asnumpy()s got through my merge, fix)

* fix lint

* revert changed .numpys

* missed a few

* fix more .asnumpy

* fix black

* Fix op level 3 test

* remove prints

* Fix pytorch and tf importers

* black

* fix lint

* fix indentation

* fix topi test

python/tvm/relay/frontend/onnx.py

@@ -2955,6 +2955,39 @@ def _impl_v11(cls, inputs, attr, params):
         return out
 
 
+class Unique(OnnxOpConverter):
+    """Operator converter for unique"""
+
+    @classmethod
+    def _impl_v11(cls, inputs, attr, params):
+        if len(inputs) != 1:
+            raise ValueError("Unique expects 1 input")
+
+        data = inputs[0]
+        axis = attr.get("axis", None)
+        if axis is None:  # If axis is None, flatten the input before calling unique
+            data = _op.reshape(data, _op.const([-1]))
+        else:
+            data_shape = infer_shape(data)
+            if len(data_shape) != 1:
+                raise ValueError("TVM only supports 1D Unique operator.")
+        is_sorted = attr.get("sorted", 1)  # sorted is 0 or 1, 1 by default
+
+        # ONNX documentation lists return_counts as optional but there is no input to specify
+        # whether it is returned. Therefore we'll just always return it.
+        unique = _op.unique(data, is_sorted=(is_sorted == 1), return_counts=True)
+        num_unique = unique[3]
+
+        trim_unique_lambda = lambda input: _op.strided_slice(input, _op.const([0]), num_unique)
+
+        unique_vals = trim_unique_lambda(unique[0])
+        indices = trim_unique_lambda(unique[1])
+        inverse_indices = unique[2]
+        counts = trim_unique_lambda(unique[4])
+        # ONNX unique returns unique, indices, inverse_indices, (optional) counts
+        return _expr.TupleWrapper(_expr.Tuple([unique_vals, indices, inverse_indices, counts]), 4)
+
+
 # compatible operators that do NOT require any conversion.
 _identity_list = []
 
@@ -3118,6 +3151,7 @@ def _get_convert_map(opset):
         "NonZero": NonZero.get_converter(opset),
         "Range": Range.get_converter(opset),
         "CumSum": CumSum.get_converter(opset),
+        "Unique": Unique.get_converter(opset),
         # defs/control_flow
         "Loop": Loop.get_converter(opset),
         "If": If.get_converter(opset),
@@ -3306,6 +3340,12 @@ def from_onnx(self, graph, opset, get_output_expr=False):
                 outputs_num = 1
             else:
                 outputs_num = len(op)
+
+            if outputs_num == 1:
+                op = fold_constant(op)
+            else:
+                op = _expr.TupleWrapper(fold_constant(op.astuple()), len(op))
+
             if outputs_num > 1:
                 # ONNX supports optional outputs for some nodes.
                 # This block searches for missing outputs in the ONNX graph
@@ -3327,8 +3367,8 @@ def from_onnx(self, graph, opset, get_output_expr=False):
                     # Create the new op with valid outputs
                     if len(outputs) == 1:
                         op = outputs[0]
-                    else:
-                        op = _expr.TupleWrapper(outputs, len(outputs))
+                    elif len(outputs) != outputs_num:
+                        op = _expr.TupleWrapper(_expr.Tuple(outputs), len(outputs))
                     # Drop invalid outputs for the onnx node
                     outputs_num = len(outputs)
                     node_output = [output for output in node_output if output != ""]
@@ -3337,10 +3377,10 @@ def from_onnx(self, graph, opset, get_output_expr=False):
             ), "Number of output mismatch {} vs {} in {}.".format(
                 len(node_output), outputs_num, op_name
             )
+
             if outputs_num == 1:
-                self._nodes[node_output[0]] = fold_constant(op)
+                self._nodes[node_output[0]] = op
             else:
-                op = _expr.TupleWrapper(fold_constant(op.astuple()), len(op))
                 for k, i in zip(list(node_output), range(len(node_output))):
                     self._nodes[k] = op[i]
 

python/tvm/relay/frontend/pytorch.py

@@ -2294,16 +2294,18 @@ def unique(self, inputs, input_types):
             logging.warning("TVM always assumes sorted=True for torch.unique")
             is_sorted = True
         if return_counts:
-            [unique, indices, num_uniq, counts] = _op.unique(
+            [unique, indices, inverse_indices, num_uniq, counts] = _op.unique(
                 data, is_sorted=is_sorted, return_counts=True
             )
             unique_sliced = _op.strided_slice(unique, begin=[0], end=num_uniq, slice_mode="size")
             counts_sliced = _op.strided_slice(counts, begin=[0], end=num_uniq, slice_mode="size")
-            return (unique_sliced, indices, counts_sliced)
+            return (unique_sliced, inverse_indices, counts_sliced)
         else:
-            [unique, indices, num_uniq] = _op.unique(data, is_sorted=is_sorted, return_counts=False)
+            [unique, indices, inverse_indices, num_uniq] = _op.unique(
+                data, is_sorted=is_sorted, return_counts=False
+            )
             unique_sliced = _op.strided_slice(unique, begin=[0], end=num_uniq, slice_mode="size")
-            return (unique_sliced, indices)
+            return (unique_sliced, inverse_indices)
 
     # Operator mappings
     def create_convert_map(self):

python/tvm/relay/frontend/tensorflow.py

@@ -2702,19 +2702,21 @@ def _impl(inputs, attr, params, mod):
         assert len(inputs) == 1
         data = inputs[0]
         if return_counts:
-            [unique, indices, num_uniq, counts] = _op.unique(
+            [unique, _, inverse_indices, num_uniq, counts] = _op.unique(
                 data, is_sorted=False, return_counts=True
             )
             unique_sliced = _op.strided_slice(unique, begin=[0], end=num_uniq, slice_mode="size")
             counts_sliced = _op.strided_slice(counts, begin=[0], end=num_uniq, slice_mode="size")
             return _expr.TupleWrapper(
-                _expr.Tuple([unique_sliced, indices, counts_sliced]),
+                _expr.Tuple([unique_sliced, inverse_indices, counts_sliced]),
                 3,
             )
-        [unique, indices, num_uniq] = _op.unique(data, is_sorted=False, return_counts=False)
+        [unique, _, inverse_indices, num_uniq] = _op.unique(
+            data, is_sorted=False, return_counts=False
+        )
         unique_sliced = _op.strided_slice(unique, begin=[0], end=num_uniq, slice_mode="size")
         return _expr.TupleWrapper(
-            _expr.Tuple([unique_sliced, indices]),
+            _expr.Tuple([unique_sliced, inverse_indices]),
             2,
         )
 

python/tvm/relay/op/_transform.py

@@ -1045,24 +1045,28 @@ def ensure_tensor(tensor):
 def _unique_shape(data_shape):
     unique_shape = output_tensor((1,), "int64")
     indices_shape = output_tensor((1,), "int64")
+    inverse_indices_shape = output_tensor((1,), "int64")
     num_unique_shape = output_tensor((1,), "int64")
     unique_shape[0] = data_shape[0]
     indices_shape[0] = data_shape[0]
+    inverse_indices_shape[0] = data_shape[0]
     num_unique_shape[0] = int64(1)
-    return (unique_shape, indices_shape, num_unique_shape)
+    return (unique_shape, indices_shape, inverse_indices_shape, num_unique_shape)
 
 
 @script
 def _unique_with_counts_shape(data_shape):
     unique_shape = output_tensor((1,), "int64")
     indices_shape = output_tensor((1,), "int64")
+    inverse_indices_shape = output_tensor((1,), "int64")
     num_unique_shape = output_tensor((1,), "int64")
     counts_shape = output_tensor((1,), "int64")
     unique_shape[0] = data_shape[0]
     indices_shape[0] = data_shape[0]
+    inverse_indices_shape[0] = data_shape[0]
     num_unique_shape[0] = int64(1)
     counts_shape[0] = data_shape[0]
-    return (unique_shape, indices_shape, num_unique_shape, counts_shape)
+    return (unique_shape, indices_shape, inverse_indices_shape, num_unique_shape, counts_shape)
 
 
 @_reg.register_shape_func("unique", False)

python/tvm/relay/op/transform.py

@@ -1658,20 +1658,24 @@ def unique(data, is_sorted=True, return_counts=False):
     data : relay.Expr
         A 1-D tensor of integers.
 
-    sorted : bool
+    is_sorted : bool
         Whether to sort the unique elements in ascending order before returning as output.
 
     return_counts : bool
         Whether to return the count of each unique element.
 
     Returns
     -------
-    output : relay.Expr
+    unique : relay.Expr
         A 1-D tensor containing the unique elements of the input data tensor.
 
     indices : relay.Expr
         A 1-D tensor containing the index of each data element in the output tensor.
 
+    inverse_indices : relay.Expr
+        A 1-D tensor. For each entry in data, it contains the index of that data element in the
+        unique array.
+
     num_unique : relay.Expr
         A 1-D tensor with size=1 containing the number of unique elements in the input data tensor.
 
@@ -1698,5 +1702,5 @@ def unique(data, is_sorted=True, return_counts=False):
         num_unique     =  [5]
     """
     if return_counts:
-        return TupleWrapper(_make.unique(data, is_sorted, return_counts), 4)
-    return TupleWrapper(_make.unique(data, is_sorted, return_counts), 3)
+        return TupleWrapper(_make.unique(data, is_sorted, return_counts), 5)
+    return TupleWrapper(_make.unique(data, is_sorted, return_counts), 4)

python/tvm/topi/cuda/unique.py

@@ -119,7 +119,7 @@ def _calc_num_unique(inc_scan):
 
 
 def _calc_unique_ir(
-    data, argsorted_indices, inc_scan, index_converter, unique_elements, indices, counts
+    data, argsorted_indices, inc_scan, index_converter, unique_elements, inverse_indices, counts
 ):
     """Low level IR to calculate unique elements, inverse indices, and counts (optional) of
     unique elements of 1-D array.
@@ -143,7 +143,7 @@ def _calc_unique_ir(
     unique_elements : Buffer
         A buffer that stores the unique elements.
 
-    indices : Buffer
+    inverse_indices : Buffer
         A buffer that stores the the index of each input data element in the unique element array.
 
     counts (optional) : Buffer
@@ -154,7 +154,7 @@ def _calc_unique_ir(
     argsorted_indices_ptr = ib.buffer_ptr(argsorted_indices)
     inc_scan_ptr = ib.buffer_ptr(inc_scan)
     unique_elements_ptr = ib.buffer_ptr(unique_elements)
-    indices_ptr = ib.buffer_ptr(indices)
+    inverse_indices_ptr = ib.buffer_ptr(inverse_indices)
 
     index_converter_ptr = None
     if isinstance(index_converter, tir.Buffer):
@@ -163,7 +163,7 @@ def _calc_unique_ir(
     if isinstance(counts, tir.Buffer):
         counts_ptr = ib.buffer_ptr(counts)
         # use indices_ptr as a tmp buffer to store tids with inc_scan[tid] != inc_scan[tid-1]
-        unique_seq_indices_ptr = ib.buffer_ptr(indices)
+        unique_seq_indices_ptr = ib.buffer_ptr(inverse_indices)
 
     batch_size = data.shape[0]
     max_threads = _get_max_threads(batch_size)
@@ -218,7 +218,7 @@ def _calc_unique_ir(
                 if not index_converter_ptr
                 else index_converter_ptr[inc_scan_ptr[tid]]
             )
-            indices_ptr[data_idx] = unique_idx
+            inverse_indices_ptr[data_idx] = unique_idx
             with ib.if_scope(tid == 0):
                 unique_elements_ptr[unique_idx] = data_ptr[data_idx]
             with ib.else_scope():
@@ -293,11 +293,20 @@ def unique(data, is_sorted=True, return_counts=False):
 
     Returns
     -------
-    output : tvm.te.Tensor
-        A 1-D tensor containing the unique elements of the input data tensor.
+    unique : tvm.te.Tensor
+        A 1-D tensor containing the unique elements of the input data tensor. The same size as
+        the input data. If there are less unique elements than input data, the end of the tensor
+        is padded with zeros.
 
     indices : tvm.te.Tensor
-        A 1-D tensor containing the index of each data element in the output tensor.
+        A 1-D tensor. The same size as output. For each entry in output, it contains
+        the index of its first occurence in the input data. The end of the tensor is padded
+        with the length of the input data.
+
+    inverse_indices : tvm.te.Tensor
+        A 1-D tensor. For each entry in data, it contains the index of that data element in the
+        unique array. (Note that inverse_indices is very similar to indices if output is not
+        sorted)
 
     num_unique : tvm.te.Tensor
         A 1-D tensor with size=1 containing the number of unique elements in the input data tensor.
@@ -309,20 +318,23 @@ def unique(data, is_sorted=True, return_counts=False):
     --------
     .. code-block:: python
         [output, indices, num_unique] = unique([4, 5, 1, 2, 3, 3, 4, 5], False, False)
-        output         =  [4, 5, 1, 2, 3, ?, ?, ?]
-        indices        =  [0, 1, 2, 3, 4, 4, 0, 1]
-        num_unique     =  [5]
+        output          =  [4, 5, 1, 2, 3, ?, ?, ?]
+        indices         =  [0, 1, 2, 3, 4, ?, ?, ?]
+        inverse_indices =  [0, 1, 2, 3, 4, 4, 0, 1]
+        num_unique      =  [5]
 
         [output, indices, num_unique, counts] = unique([4, 5, 1, 2, 3, 3, 4, 5], False, True)
-        output         =  [4, 5, 1, 2, 3, ?, ?, ?]
-        indices        =  [0, 1, 2, 3, 4, 4, 0, 1]
-        num_unique     =  [5]
-        counts         =  [2, 2, 1, 1, 2, ?, ?, ?]
+        output          =  [4, 5, 1, 2, 3, ?, ?, ?]
+        indices         =  [0, 1, 2, 3, 4, ?, ?, ?]
+        inverse_indices =  [0, 1, 2, 3, 4, 4, 0, 1]
+        num_unique      =  [5]
+        counts          =  [2, 2, 1, 1, 2, ?, ?, ?]
 
         [output, indices, num_unique] = unique([4, 5, 1, 2, 3, 3, 4, 5], True)
-        output         =  [1, 2, 3, 4, 5, ?, ?, ?]
-        indices        =  [3, 4, 0, 1, 2, 2, 3, 4]
-        num_unique     =  [5]
+        output          =  [1, 2, 3, 4, 5, ?, ?, ?]
+        indices         =  [2, 3, 4, 0, 1, ?, ?, ?]
+        inverse_indices =  [3, 4, 0, 1, 2, 2, 3, 4]
+        num_unique      =  [5]
     """
     sorted_data = sort(data)
     argsorted_indices = argsort(data, dtype="int32")
@@ -355,29 +367,29 @@ def unique(data, is_sorted=True, return_counts=False):
         out_buffers = [unique_elements_buf, inverse_indices_buf]
         out_dtypes = [data.dtype, "int32"]
     # prepare inputs and fcompute
+    # calculate first occurence
+    first_occurence_buf = tir.decl_buffer(
+        data.shape, "int32", "first_occurence_buf", data_alignment=8
+    )
+    first_occurence = te.extern(
+        [data.shape],
+        [argsorted_indices, inc_scan],
+        lambda ins, outs: _calc_first_occurence_ir(ins[0], ins[1], outs[0]),
+        dtype=["int32"],
+        in_buffers=[argsorted_indices_buf, inc_scan_buf],
+        out_buffers=[first_occurence_buf],
+        name="_calc_first_occurence",
+        tag="_calc_first_occurence_gpu",
+    )
     if is_sorted:
         in_data = [data, argsorted_indices, inc_scan]
         in_buffers = [data_buf, argsorted_indices_buf, inc_scan_buf]
         if return_counts:
             fcompute = lambda ins, outs: _calc_unique_ir(*ins, None, *outs)
         else:
             fcompute = lambda ins, outs: _calc_unique_ir(*ins, None, *outs, None)
+        indices = first_occurence
     else:
-        # calculate the index converter if the unique elements should not be sorted
-        # calculate first occurence
-        first_occurence_buf = tir.decl_buffer(
-            data.shape, "int32", "first_occurence_buf", data_alignment=8
-        )
-        first_occurence = te.extern(
-            [data.shape],
-            [argsorted_indices, inc_scan],
-            lambda ins, outs: _calc_first_occurence_ir(ins[0], ins[1], outs[0]),
-            dtype=["int32"],
-            in_buffers=[argsorted_indices_buf, inc_scan_buf],
-            out_buffers=[first_occurence_buf],
-            name="_calc_first_occurence",
-            tag="_calc_first_occurence_gpu",
-        )
         # calculate index converter by sorting unique elements by their first occurence
         argsorted_first_occurence = argsort(first_occurence, dtype="int32")
         index_converter = argsort(argsorted_first_occurence, dtype="int32")
@@ -390,6 +402,7 @@ def unique(data, is_sorted=True, return_counts=False):
             fcompute = lambda ins, outs: _calc_unique_ir(*ins, *outs)
         else:
             fcompute = lambda ins, outs: _calc_unique_ir(*ins, *outs, None)
+        indices = sort(first_occurence)
     outs = te.extern(
         out_data_shape,
         in_data,
@@ -401,5 +414,5 @@ def unique(data, is_sorted=True, return_counts=False):
         tag="_calc_unique_gpu",
     )
     if return_counts:
-        return [outs[0], outs[1], num_unique_elements, outs[2]]
-    return [*outs, num_unique_elements]
+        return [outs[0], indices, outs[1], num_unique_elements, outs[2]]
+    return [outs[0], indices, outs[1], num_unique_elements]