Merge pull request #97 from andfoy/add_stack

Add stack

lib/extorch/native/tensor/ops/manipulation.ex

@@ -1553,7 +1553,66 @@ defmodule ExTorch.Native.Tensor.Ops.Manipulation do
         iex> b.size
         {1, 3, 4, 5}
     """
-    @spec squeeze(ExTorch.Tensor.t(), integer() | tuple() | [integer()] | nil) :: ExTorch.Tensor.t()
+    @spec squeeze(ExTorch.Tensor.t(), integer() | tuple() | [integer()] | nil) ::
+            ExTorch.Tensor.t()
     defbinding(squeeze(input, dim \\ nil))
+
+    @doc """
+    Concatenates a sequence of tensors along a new dimension.
+
+    All tensors need to be of the same size. This function is analogous to `ExTorch.cat/3`
+
+    ## Arguments
+    - `tensors` (`[ExTorch.Tensor] | tuple()`) - A sequence of tensors of the same type. Non-empty
+    tensors provided must have the same shape.
+
+    ## Optional arguments
+    - `dim` (`integer()`) - the dimension over which the tensors are concatenated. Default: 0
+    - `out` (`ExTorch.Tensor | nil`) - an optional pre-allocated tensor used to store the
+    concatenation output. Default: nil
+
+    ## Examples
+        iex> a = ExTorch.rand({3, 4})
+        #Tensor<
+        [[0.7419, 0.4063, 0.0514, 0.4281],
+         [0.7350, 0.1977, 0.5593, 0.1701],
+         [0.4135, 0.7213, 0.9591, 0.2798]]
+        [size: {3, 4}, dtype: :float, device: :cpu, requires_grad: false]>
+
+        # Concatenate tensors into a new dimension at the beginning
+        iex> ExTorch.stack([a, a, a])
+        #Tensor<
+        [[[0.7419, 0.4063, 0.0514, 0.4281],
+          [0.7350, 0.1977, 0.5593, 0.1701],
+          [0.4135, 0.7213, 0.9591, 0.2798]],
+
+         [[0.7419, 0.4063, 0.0514, 0.4281],
+          [0.7350, 0.1977, 0.5593, 0.1701],
+          [0.4135, 0.7213, 0.9591, 0.2798]],
+
+         [[0.7419, 0.4063, 0.0514, 0.4281],
+          [0.7350, 0.1977, 0.5593, 0.1701],
+          [0.4135, 0.7213, 0.9591, 0.2798]]]
+        [size: {3, 3, 4}, dtype: :float, device: :cpu, requires_grad: false]>
+
+        # Concatenate tensors into a new dimension at the first position
+        iex> ExTorch.stack([a, a, a], 1)
+        #Tensor<
+        [[[0.7419, 0.4063, 0.0514, 0.4281],
+          [0.7419, 0.4063, 0.0514, 0.4281],
+          [0.7419, 0.4063, 0.0514, 0.4281]],
+
+         [[0.7350, 0.1977, 0.5593, 0.1701],
+          [0.7350, 0.1977, 0.5593, 0.1701],
+          [0.7350, 0.1977, 0.5593, 0.1701]],
+
+         [[0.4135, 0.7213, 0.9591, 0.2798],
+          [0.4135, 0.7213, 0.9591, 0.2798],
+          [0.4135, 0.7213, 0.9591, 0.2798]]]
+        [size: {3, 3, 4}, dtype: :float, device: :cpu, requires_grad: false]>
+    """
+    @spec stack([ExTorch.Tensor.t()] | tuple(), integer(), ExTorch.Tensor.t() | nil) ::
+            ExTorch.Tensor.t()
+    defbinding(stack(input, dim \\ 0, out \\ nil))
   end
 end

native/extorch/include/manipulation.h

@@ -174,3 +174,5 @@ TensorList split(
 std::shared_ptr<CrossTensor> squeeze(
     const std::shared_ptr<CrossTensor> &input,
     rust::Vec<int64_t> dims);
+
+std::shared_ptr<CrossTensor> stack(TensorList seq, int64_t dim, TensorOut opt_out);

native/extorch/src/csrc/manipulation.cc

@@ -722,3 +722,16 @@ std::shared_ptr<CrossTensor> squeeze(
     }
     return std::make_shared<CrossTensor>(std::move(out_tensor));
 }
+
+std::shared_ptr<CrossTensor> stack(TensorList seq, int64_t dim, TensorOut opt_out) {
+    CrossTensor out_tensor;
+    std::vector<CrossTensor> tensor_seq = unpack_tensor_list(seq);
+
+    if(opt_out.used) {
+        out_tensor = *opt_out.tensor.get();
+        out_tensor = torch::stack_out(out_tensor, tensor_seq, dim);
+    } else {
+        out_tensor = torch::stack(tensor_seq, dim);
+    }
+    return std::make_shared<CrossTensor>(std::move(out_tensor));
+}

native/extorch/src/lib.rs

@@ -124,6 +124,7 @@ rustler::init!(
         scatter_reduce,
         split,
         squeeze,
+        stack,
 
         // Tensor comparing operations
         allclose,

native/extorch/src/native/tensor/ops.rs.in

@@ -227,3 +227,6 @@ fn split(
 
 /// Remove specified or all singleton dimensions from a input tensor.
 fn squeeze(input: &SharedPtr<CrossTensor>, dims: Vec<i64>) -> Result<SharedPtr<CrossTensor>>;
+
+/// Concatenate a sequence of tensors across a new dimension.
+fn stack(seq: TensorList, dim: i64, out: TensorOut) -> Result<SharedPtr<CrossTensor>>;

native/extorch/src/nifs/tensor_ops.rs

@@ -281,3 +281,5 @@ nif_impl!(
     input: TensorStruct<'a>,
     dims: Size
 );
+
+nif_impl!(stack, TensorStruct<'a>, seq: TensorList, dim: i64, out: TensorOut);

test/tensor/manipulation_test.exs

@@ -803,4 +803,41 @@ defmodule ExTorchTest.Tensor.ManipulationTest do
     out = ExTorch.squeeze(input, {0, 2})
     assert out.size == expected
   end
+
+  test "stack/1" do
+    input = ExTorch.rand({3, 4, 2})
+    exp_input = ExTorch.unsqueeze(input, 0)
+    expected = ExTorch.cat([exp_input, exp_input, exp_input])
+
+    out = ExTorch.stack([input, input, input])
+    assert ExTorch.allclose(out, expected)
+  end
+
+  test "stack/2" do
+    input = ExTorch.rand({3, 4, 2})
+    exp_input = ExTorch.unsqueeze(input, 2)
+    expected = ExTorch.cat([exp_input, exp_input, exp_input], 2)
+
+    out = ExTorch.stack([input, input, input], 2)
+    assert ExTorch.allclose(out, expected)
+  end
+
+  test "stack/2 with kwargs" do
+    input = ExTorch.rand({3, 4, 2})
+    exp_input = ExTorch.unsqueeze(input, 3)
+    expected = ExTorch.cat([exp_input, exp_input, exp_input], dim: 3)
+
+    out = ExTorch.stack([input, input, input], dim: 3)
+    assert ExTorch.allclose(out, expected)
+  end
+
+  test "stack/3" do
+    input = ExTorch.rand({3, 4, 2})
+    exp_input = ExTorch.unsqueeze(input, 1)
+    expected = ExTorch.cat([exp_input, exp_input, exp_input], 1)
+
+    out = ExTorch.empty_like(expected)
+    ExTorch.stack([input, input, input], 1, out)
+    assert ExTorch.allclose(out, expected)
+  end
 end