Add support for GatherV2 batch_dims attr (#1329)

* Add support for GatherV2 batch_dims attr

Signed-off-by: Tom Wildenhain <[email protected]>

* Fix tests

Signed-off-by: Tom Wildenhain <[email protected]>

Author: TomWildenhain-Microsoft

tests/test_backend.py

@@ -871,6 +871,26 @@ def func(x):
             return tf.identity(x_, name=_TFOUTPUT)
         self._run_test_case(func, [_OUTPUT], {_INPUT: x_val})
 
+    @check_tf_min_version("1.14")
+    @check_opset_min_version(12, "GatherND with batch_dims")
+    def test_gather_batch_dims_no_trans(self):
+        x_val = np.arange(2 * 2 * 3 * 5 * 4, dtype=np.float32).reshape((2, 2, 3, 5, 4))
+        idx_val = np.array([[[1, 0, 2, 0], [1, 1, 1, 0]], [[0, 0, 0, 0], [2, 1, 1, 0]]], dtype=np.int32)
+        def func(x, idx):
+            x_ = tf.gather(x, idx, batch_dims=2, axis=2)
+            return tf.identity(x_, name=_TFOUTPUT)
+        self._run_test_case(func, [_OUTPUT], {_INPUT: x_val, _INPUT1: idx_val})
+
+    @check_tf_min_version("1.14")
+    @check_opset_min_version(12, "GatherND with batch_dims")
+    def test_gather_batch_dims(self):
+        x_val = np.arange(2 * 2 * 3 * 5 * 4, dtype=np.float32).reshape((2, 2, 3, 5, 4))
+        idx_val = np.array([[[1, 0, 2, 0], [1, 1, 1, 0]], [[0, 0, 0, 0], [2, 1, 1, 0]]], dtype=np.int32)
+        def func(x, idx):
+            x_ = tf.gather(x, idx, batch_dims=2, axis=3)
+            return tf.identity(x_, name=_TFOUTPUT)
+        self._run_test_case(func, [_OUTPUT], {_INPUT: x_val, _INPUT1: idx_val})
+
     @check_opset_min_version(10, "Slice")
     def test_roll_axis_scalar(self):
         x_val = np.arange(4 * 3 * 5 * 2, dtype=np.float32).reshape((4, 3, 5, 2))

tf2onnx/onnx_opset/tensor.py

@@ -423,7 +423,10 @@ class GatherV2:
     @classmethod
     def version_1(cls, ctx, node, **kwargs):
         # for GatherV2 axis come as input
+        err_msg = "Opset 12 required for batch_dims attribute of GatherV2"
+        utils.make_sure(node.get_attr_value("batch_dims", 0) == 0, err_msg)
         node.type = "Gather"
+        utils.make_sure(node.inputs[2].is_const(), "Axis of GatherV2 node must be constant")
         axis = node.inputs[2].get_tensor_value()
         ctx.remove_input(node, node.input[2], 2)
         node.set_attr("axis", axis)
@@ -433,6 +436,42 @@ def version_11(cls, ctx, node, **kwargs):
         # no change
         cls.version_1(ctx, node, **kwargs)
 
+    @classmethod
+    def version_12(cls, ctx, node, **kwargs):
+        batch_dims = node.get_attr_value("batch_dims", 0)
+        if batch_dims == 0:
+            cls.version_1(ctx, node, **kwargs)
+            return
+        # If batch_dims is not zero, use GatherND to simulate Gather with batch dims.
+        data_inp, indices_inp, axis_inp = node.input
+        utils.make_sure(node.inputs[2].is_const(), "Axis of GatherV2 node must be constant")
+        axis = node.inputs[2].get_tensor_value()
+        ctx.remove_input(node, axis_inp, 2)
+        if ctx.get_dtype(indices_inp) != TensorProto.INT64:
+            indices_inp = ctx.make_node("Cast", [indices_inp], attr={'to': TensorProto.INT64}).output[0]
+        unperm = None
+        # GatherND doesn't take an axis so we have to transpose stuff around
+        if axis != batch_dims:
+            data_rank = ctx.get_rank(data_inp)
+            indices_rank = ctx.get_rank(indices_inp)
+            result_rank = data_rank + indices_rank - 1 - batch_dims
+            shift_amt = axis - batch_dims
+            err_msg = "Cannot convert GatherV2 with batch dims since inputs have unknown ranks."
+            utils.make_sure(data_rank is not None and indices_rank is not None, err_msg)
+            perm = list(range(data_rank))
+            perm = perm[:batch_dims] + perm[axis:axis+1] + perm[batch_dims:axis] + perm[axis+1:]
+            data_inp = ctx.make_node("Transpose", [data_inp], attr={'perm': perm}).output[0]
+            ctx.replace_input(node, node.input[0], data_inp, 0)
+            unperm = list(range(result_rank))
+            j = indices_rank+shift_amt
+            unperm = unperm[:batch_dims] + unperm[indices_rank:j] + unperm[batch_dims:indices_rank] + unperm[j:]
+        node.type = "GatherND"
+        unsqueeze_node = GraphBuilder(ctx).make_unsqueeze({'data': indices_inp, 'axes': [-1]})
+        ctx.replace_input(node, node.input[1], unsqueeze_node, 1)
+        if unperm is not None:
+            ctx.update_node_shape_dtype(node, override=True)
+            ctx.insert_new_node_on_output("Transpose", node.output[0], perm=unperm)
+
 
 def _make_gathernd_inner_loop(ctx, params, index, dtype):
     """create the inner loop for GatherNd."""