[Frontend][Tensorflow]add batch_dim support for gatherV2 (#7951)

* add batch_dim support

* fix lint

* add check for num of arguments for topi.take

* fix gpu test cases

* add check for batch_dims in take_grad

Author: Xingyu Zhou

include/tvm/relay/attrs/transform.h

@@ -145,10 +145,14 @@ struct GatherAttrs : public tvm::AttrsNode<GatherAttrs> {
 };
 
 struct TakeAttrs : public tvm::AttrsNode<TakeAttrs> {
+  Integer batch_dims;
   Integer axis;
   std::string mode;
 
   TVM_DECLARE_ATTRS(TakeAttrs, "relay.attrs.TakeAttrs") {
+    TVM_ATTR_FIELD(batch_dims)
+        .set_default(0)
+        .describe("The batch_dims over which to select values.");
     TVM_ATTR_FIELD(axis)
         .set_default(NullValue<Integer>())
         .describe("The axis over which to select values.");

include/tvm/topi/transform.h

@@ -763,15 +763,17 @@ inline Array<Tensor> split_sections(const Tensor& x, int num_sections, int axis,
  *
  * \param a The source array.
  * \param indices The indices of the values to extract.
+ * \param batch_dims The number of batch dimensions.
  * \param mode The mode of the operation.
  * \param name The name of the operation.
  * \param mode The mode of to handle out of bound indices.
  * \param tag The tag to mark the operation.
  *
  * \return A Tensor whose op member is the take operation
  */
-inline Tensor take(const Tensor& a, const Tensor& indices, std::string mode = "clip",
-                   std::string name = "T_take", std::string tag = kInjective) {
+inline Tensor take(const Tensor& a, const Tensor& indices, int batch_dims,
+                   std::string mode = "clip", std::string name = "T_take",
+                   std::string tag = kInjective) {
   Array<PrimExpr> a_shape = a->shape;
   Array<PrimExpr> out_shape = indices->shape;
   PrimExpr a_size = 1;
@@ -846,6 +848,7 @@ inline Tensor sequence_mask(const Tensor& data, const Tensor& valid_length, doub
  *
  * \param a The source array.
  * \param indices The indices of the values to extract.
+ * \param batch_dims The number of batch dimensions. By default is 0.
  * \param axis The axis over which to select values. By default,
  * the flattened input array is used.
  * \param mode The mode for handling out of bound indices.
@@ -854,46 +857,99 @@ inline Tensor sequence_mask(const Tensor& data, const Tensor& valid_length, doub
  *
  * \return A Tensor whose op member is the take operation
  */
-inline Tensor take(const Tensor& a, const Tensor& indices, int axis, std::string mode = "clip",
-                   std::string name = "T_take", std::string tag = kInjective) {
+inline Tensor take(const Tensor& a, const Tensor& indices, int batch_dims, int axis,
+                   std::string mode = "clip", std::string name = "T_take",
+                   std::string tag = kInjective) {
   if (axis < 0) {
     axis += static_cast<int>(a->shape.size());
   }
   ICHECK_GE(axis, 0) << "axis out of bounds";
   ICHECK_LT(axis, a->shape.size()) << "axis out of bounds";
   auto axis_dim = a->shape[axis];
-
   int indices_len = static_cast<int>(indices->shape.size());
-  Array<PrimExpr> out_shape;
-  for (size_t i = 0; i < a->shape.size(); ++i) {
-    if (axis == static_cast<int>(i)) {
-      for (size_t j = 0; j < indices->shape.size(); ++j) {
-        out_shape.push_back(indices->shape[j]);
-      }
-    } else {
-      out_shape.push_back(a->shape[i]);
+
+  int batch_dims_ = batch_dims;
+  if (batch_dims_ != 0) {
+    ICHECK_GE(batch_dims_, -static_cast<int>(indices->shape.size())) << "batch_dims out of bounds";
+    ICHECK_LE(batch_dims_, indices->shape.size()) << "batch_dims out of bounds";
+
+    if (batch_dims_ < 0) {
+      batch_dims_ = indices->shape.size() + batch_dims_;
     }
+
+    ICHECK_LT(batch_dims_, a->shape.size()) << "batch_dims out of bounds";
+    ICHECK_LE(batch_dims_, axis) << "batch_dims must be less than or equal to axis";
+    for (int i = 0; i < batch_dims_; ++i) {
+      auto addr1 = a->shape[i];
+      auto addr2 = indices->shape[i];
+      auto v1 = static_cast<IntImm*>(&addr1)->get()->value;
+      auto v2 = static_cast<IntImm*>(&addr2)->get()->value;
+      ICHECK_EQ(v1, v2) << "a.shape[" << i << "] should be equal to indices.shape[" << i << "]";
+    }
+  }
+
+  // The result shape is a.shape[:axis] + indices.shape[batch_dims:] +
+  // a.shape[axis + 1:].
+
+  Array<PrimExpr> out_shape;
+  for (int i = 0; i < batch_dims_; ++i) {
+    out_shape.push_back(a->shape[i]);
+  }
+  for (int i = batch_dims_; i < axis; ++i) {
+    out_shape.push_back(a->shape[i]);
+  }
+  for (size_t i = static_cast<size_t>(batch_dims_); i < indices->shape.size(); ++i) {
+    out_shape.push_back(indices->shape[i]);
+  }
+  for (size_t i = axis + 1; i < a->shape.size(); ++i) {
+    out_shape.push_back(a->shape[i]);
   }
+
   if (mode == "clip") {
-    return compute(
-        out_shape,
-        [&](const Array<Var>& out_index) {
-          Array<PrimExpr> indices_position;
-          for (size_t j = axis; j < static_cast<size_t>(axis + indices_len); ++j) {
-            indices_position.push_back(out_index[j]);
-          }
-          Array<PrimExpr> real_indices;
-          for (size_t j = 0; j < static_cast<size_t>(axis); ++j) {
-            real_indices.push_back(out_index[j]);
-          }
-          auto idx = tvm::min(tvm::max(0, indices(indices_position)), axis_dim - 1);
-          real_indices.push_back(idx);
-          for (size_t j = axis + indices_len; j < out_index.size(); ++j) {
-            real_indices.push_back(out_index[j]);
-          }
-          return a(real_indices);
-        },
-        name, tag);
+    if (batch_dims_ == 0) {
+      return compute(
+          out_shape,
+          [&](const Array<Var>& out_index) {
+            Array<PrimExpr> indices_position;
+            for (size_t j = axis; j < static_cast<size_t>(axis + indices_len); ++j) {
+              indices_position.push_back(out_index[j]);
+            }
+            Array<PrimExpr> real_indices;
+            for (size_t j = 0; j < static_cast<size_t>(axis); ++j) {
+              real_indices.push_back(out_index[j]);
+            }
+            auto idx = tvm::min(tvm::max(0, indices(indices_position)), axis_dim - 1);
+            real_indices.push_back(idx);
+            for (size_t j = axis + indices_len; j < out_index.size(); ++j) {
+              real_indices.push_back(out_index[j]);
+            }
+            return a(real_indices);
+          },
+          name, tag);
+    } else {
+      return compute(
+          out_shape,
+          [&](const Array<Var>& out_index) {
+            Array<PrimExpr> indices_position;
+            for (size_t j = 0; j < static_cast<size_t>(batch_dims_); ++j) {
+              indices_position.push_back(out_index[j]);
+            }
+            for (size_t j = axis; j < static_cast<size_t>(axis + indices_len - batch_dims_); ++j) {
+              indices_position.push_back(out_index[j]);
+            }
+            Array<PrimExpr> real_indices;
+            for (size_t j = 0; j < static_cast<size_t>(axis); ++j) {
+              real_indices.push_back(out_index[j]);
+            }
+            auto idx = tvm::min(tvm::max(0, indices(indices_position)), axis_dim - 1);
+            real_indices.push_back(idx);
+            for (size_t j = axis + indices_len - batch_dims_; j < out_index.size(); ++j) {
+              real_indices.push_back(out_index[j]);
+            }
+            return a(real_indices);
+          },
+          name, tag);
+    }
   } else if (mode == "fast") {
     LOG(WARNING) << "Fast mode segfaults when there are out-of-bounds indices. "
                     "Make sure input indices are in bound";

python/tvm/relay/frontend/mxnet.py

@@ -911,7 +911,7 @@ def _mx_take(inputs, attrs):
     if mode == "raise":
         raise tvm.error.OpAttributeUnimplemented("take with raise mode is not supported yet")
     axis = attrs.get_int("axis", 0)
-    return _op.take(inputs[0], inputs[1].astype("int32"), axis, mode)
+    return _op.take(inputs[0], inputs[1].astype("int32"), axis=axis, mode=mode)
 
 
 def _mx_gather_nd(inputs, attrs):

python/tvm/relay/frontend/tensorflow.py

@@ -2002,14 +2002,19 @@ def _impl(inputs, attr, params, mod):
             axis = _get_num_param(params, inputs.pop(2))
         else:
             axis = 0
+        batch_dims = 0
         if int(attr.get("batch_dims", 0)) != 0:
-            raise tvm.error.OpAttributeUnImplemented("Attribute batch_dims is not supported")
+            batch_dims = int(attr.get("batch_dims", 0))
         new_input = inputs[0:2]
-        return AttrCvt(
+        op_ = AttrCvt(
             op_name="take",
-            extras={"axis": tvm.tir.const(axis, "int32")},
-            ignores=["Tindices", "Tparams", "validate_indices", "Taxis", "_class", "batch_dims"],
+            extras={
+                "axis": tvm.tir.const(axis, "int32"),
+                "batch_dims": tvm.tir.const(batch_dims, "int32"),
+            },
+            ignores=["Tindices", "Tparams", "validate_indices", "Taxis", "_class"],
         )(new_input, attr)
+        return op_
 
     return _impl
 

python/tvm/relay/op/_tensor_grad.py

@@ -711,6 +711,7 @@ def make_scalar_tensor(v):
     # TODO(@altanh): we currently assume indices are in range
     data, indices = orig.args
     axis = orig.attrs.axis
+    batch_dims = orig.attrs.batch_dims
     zero, one = map(make_scalar_tensor, [0, 1])
     data_grad = zeros_like(data)
     try:
@@ -726,6 +727,12 @@ def make_scalar_tensor(v):
         data_shape = (data_shape,)
     else:
         axis = int(axis)
+    if batch_dims is None:
+        batch_dims = 0
+    else:
+        batch_dims = int(batch_dims)
+    if batch_dims != 0:
+        raise OpError("take_grad only supports batch_dims equales to 0")
     strides = [1] * len(data_shape)
 
     if len(indices.checked_type.shape) == 0:

python/tvm/relay/op/_transform.py

@@ -390,7 +390,7 @@ def _take_no_axis_shape_func(indices_shape, out_ndim):
 
 
 @script
-def _take_with_axis_shape_func(data_shape, indices_shape, axis, out_ndim):
+def _take_with_axis_shape_func(data_shape, indices_shape, axis, batch_dims, out_ndim):
     out = output_tensor((out_ndim,), "int64")
     for i in const_range(axis):
         out[i] = data_shape[i]
@@ -399,10 +399,10 @@ def _take_with_axis_shape_func(data_shape, indices_shape, axis, out_ndim):
         for i in const_range(axis + 1, len(data_shape)):
             out[i - 1] = data_shape[i]
     else:
-        for i in const_range(len(indices_shape)):
-            out[axis + i] = indices_shape[i]
+        for i in const_range(len(indices_shape) - batch_dims):
+            out[axis + i] = indices_shape[i + batch_dims]
         for i in const_range(axis + 1, len(data_shape)):
-            out[len(indices_shape) + i - 1] = data_shape[i]
+            out[len(indices_shape) + i - 1 - batch_dims] = data_shape[i]
     return out
 
 
@@ -414,11 +414,16 @@ def take_shape_func(attrs, inputs, out_ndims):
     if attrs.axis is None:
         return [_take_no_axis_shape_func(inputs[1], out_ndims[0])]
     axis = get_const_int(attrs.axis)
+    batch_dims = get_const_int(attrs.batch_dims)
     data_ndim = int(inputs[0].shape[0])
+    if inputs[1].shape:
+        indicies_ndim = int(inputs[1].shape[0])
     if axis < 0:
         axis += data_ndim
     assert 0 <= axis < data_ndim
-    return [_take_with_axis_shape_func(*inputs, convert(axis), out_ndims[0])]
+    if batch_dims < 0:
+        batch_dims += indicies_ndim
+    return [_take_with_axis_shape_func(*inputs, convert(axis), convert(batch_dims), out_ndims[0])]
 
 
 @_reg.register_legalize("take")

python/tvm/relay/op/transform.py

@@ -388,7 +388,7 @@ def reshape_like(data, shape_like, lhs_begin=0, lhs_end=None, rhs_begin=0, rhs_e
     return _make.reshape_like(data, shape_like, lhs_begin, lhs_end, rhs_begin, rhs_end)
 
 
-def take(data, indices, axis=None, mode="clip"):
+def take(data, indices, axis=None, batch_dims=0, mode="clip"):
     """Take elements from an array along an axis.
 
     Parameters
@@ -403,6 +403,9 @@ def take(data, indices, axis=None, mode="clip"):
         The axis over which to select values. By default,
         the flattened input array is used.
 
+    batch_dims : int
+        The number of batch dimensions. By default is 0.
+
     mode : str, optional
         Specifies how out-of-bound indices will behave [clip, wrap, fast].
         clip: clip to the range (default).
@@ -414,7 +417,7 @@ def take(data, indices, axis=None, mode="clip"):
     ret : relay.Expr
         The computed result.
     """
-    return _make.take(data, indices, axis, mode)
+    return _make.take(data, indices, batch_dims, axis, mode)
 
 
 def full(fill_value, shape=(), dtype=""):

python/tvm/topi/transform.py

@@ -396,7 +396,7 @@ def split(ary, indices_or_sections, axis=0):
     return cpp.split(ary, indices_or_sections, axis)
 
 
-def take(a, indices, axis=None, mode="clip"):
+def take(a, indices, axis=None, batch_dims=0, mode="clip"):
     """Take elements from an array along an axis.
 
     Parameters
@@ -411,6 +411,9 @@ def take(a, indices, axis=None, mode="clip"):
         The axis over which to select values. By default,
         the flattened input array is used.
 
+    batch_dims : int
+        The number of batch dimensions. By default is 0.
+
     mode : str, optional
         Specifies how out-of-bound indices will behave.
         clip - clip to the range (default)
@@ -422,8 +425,8 @@ def take(a, indices, axis=None, mode="clip"):
     ret : tvm.te.Tensor
     """
     if axis is None:
-        return cpp.take(a, indices, mode)
-    return cpp.take(a, indices, int(axis), mode)
+        return cpp.take(a, indices, int(batch_dims), mode)
+    return cpp.take(a, indices, int(batch_dims), int(axis), mode)
 
 
 @tvm.target.generic_func

src/relay/op/make_op.h

@@ -107,7 +107,7 @@ Expr MakeArange(Expr start, Expr stop, Expr step, DataType dtype);
 
 Expr MakeShapeOf(Expr data, DataType dtype);
 
-Expr MakeTake(Expr data, Expr indices, Integer axis, String mode);
+Expr MakeTake(Expr data, Expr indices, Integer batch_dims, Integer axis, String mode);
 
 }  // namespace relay
 }  // namespace tvm

src/relay/op/tensor/transform.cc

@@ -1204,15 +1204,24 @@ bool TakeRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
   const auto ndim_data = static_cast<int>(data->shape.size());
   const auto ndim_indices = static_cast<int>(indices->shape.size());
   int axis = static_cast<int>(param->axis->value);
+  int batch_dims = static_cast<int>(param->batch_dims->value);
   if (axis < 0) axis += ndim_data;
+  if (batch_dims < 0) axis += ndim_indices;
   ICHECK_LE(axis, ndim_data) << "axis should be with in data shape"
                              << ", but got = " << axis;
+  ICHECK_LE(batch_dims, ndim_indices) << "batch_dims should be with in indices shape"
+                                      << ", but got = " << batch_dims;
+  ICHECK_LE(batch_dims, axis) << "batch_dims should be less than or equal to axis"
+                              << ", but got = " << batch_dims;
 
-  oshape.reserve(ndim_data - 1 + ndim_indices);
-  for (int i = 0; i < axis; ++i) {
+  oshape.reserve(ndim_data - 1 + ndim_indices - batch_dims);
+  for (int i = 0; i < batch_dims; ++i) {
+    oshape.emplace_back(data->shape[i]);
+  }
+  for (int i = batch_dims; i < axis; ++i) {
     oshape.emplace_back(data->shape[i]);
   }
-  for (int i = 0; i < ndim_indices; ++i) {
+  for (int i = batch_dims; i < ndim_indices; ++i) {
     oshape.emplace_back(indices->shape[i]);
   }
   for (int i = axis + 1; i < ndim_data; ++i) {
@@ -1228,14 +1237,16 @@ Array<te::Tensor> TakeCompute(const Attrs& attrs, const Array<te::Tensor>& input
   const auto* param = attrs.as<TakeAttrs>();
   ICHECK(param != nullptr);
   if (!param->axis.defined()) {
-    return Array<te::Tensor>{topi::take(inputs[0], inputs[1], param->mode)};
+    return Array<te::Tensor>{topi::take(inputs[0], inputs[1], param->batch_dims, param->mode)};
   } else {
-    return Array<te::Tensor>{topi::take(inputs[0], inputs[1], param->axis, param->mode)};
+    return Array<te::Tensor>{
+        topi::take(inputs[0], inputs[1], param->batch_dims, param->axis, param->mode)};
   }
 }
 
-Expr MakeTake(Expr data, Expr indices, Integer axis, String mode) {
+Expr MakeTake(Expr data, Expr indices, Integer batch_dims, Integer axis, String mode) {
   auto attrs = make_object<TakeAttrs>();
+  attrs->batch_dims = std::move(batch_dims);
   attrs->axis = std::move(axis);
   attrs->mode = std::move(mode);
   static const Op& op = Op::Get("take");