Gather elements Operator

docs/CHANGELOG.md

@@ -4,6 +4,11 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased] - 2023-11-20
+
+## Added
+- Gather Elements Operator.
+
 ## [Unreleased] - 2023-11-06
 
 ## Added

docs/SUMMARY.md

@@ -105,6 +105,7 @@
     * [tensor.reduce\_sum\_square](framework/operators/tensor/tensor.reduce\_sum\_square.md)
     * [tensor.reduce\_l2](framework/operators/tensor/tensor.reduce\_l2.md)
     * [tensor.reduce\_l1](framework/operators/tensor/tensor.reduce\_l1.md)
+    * [tensor.gather\_elements](framework/operators/tensor/tensor.gather\_elements.md)
     * [tensor.sequence\_length](framework/operators/tensor/tensor.sequence\_length.md)
     * [tensor.sequence\_at](framework/operators/tensor/tensor.sequence\_at.md)
     * [tensor.reduce\_min](framework/operators/tensor/tensor.reduce\_min.md)

docs/framework/compatibility.md

@@ -86,6 +86,7 @@ You can see below the list of current supported ONNX Operators:
 |   [ConstantOfShape](operators/tensor/tensor.constant_of_shape.md)   | :white\_check\_mark: |
 |          [ReduceL1](operators/tensor/tensor.reduce\_l1.md)          | :white\_check\_mark: |
 |         [ReduceL2](operators/tensor/tensor.reduce\_l2.md)           | :white\_check\_mark: |
+|    [GatherElements](operators/tensor/tensor.gather/_elements.md)   | :white\_check\_mark: |
 |         [SequenceLength](operators/tensor/tensor.sequence\_length.md)   | :white\_check\_mark: |
 |        [SequenceAt](operators/tensor/tensor.sequence\_at.md)        | :white\_check\_mark: |
 | [SequenceConstruct](operators/tensor/tensor.sequence\_construct.md) | :white\_check\_mark: |

docs/framework/operators/tensor/README.md

@@ -101,6 +101,9 @@ use orion::operators::tensor::TensorTrait;
 | [`tensor.scatter`](tensor.scatter.md) | Produces a copy of input data, and updates value to values specified by updates at specific index positions specified by indices. |
 | [`tensor.reduce_sum_square`](tensor.reduce\_sum\_square.md) | Computes the sum square of the input tensor's elements along the provided axes.  |
 | [`tensor.reduce_l2`](tensor.reduce\_l2.md) | Computes the L2 norm of the input tensor's elements along the provided axes. |
+| [`tensor.gather_elements`](tensor.gather\_elements.md) | GatherElements is an indexing operation that produces its output by indexing into the input data tensor at index positions determined by elements of the indices tensor. |
+| [`tensor.reduce_min`](tensor.reduce\_min.md) | Computes the min of the input tensor's elements along the provided axes. |
+| [`tensor.sequence_construct`](tensor.sequence\_construct.md) | Constructs a tensor sequence containing the input tensors. |
 | [`tensor.sequence_length`](tensor.sequence\_length.md) | Returns the length of the input sequence. |
 | [`tensor.shrink`](tensor.shrink.md) | Shrinks the input tensor element-wise to the output tensor with the same datatype and shape based on a defined formula. |
 | [`tensor.sequence_empty`](tensor.sequence\_empty.md) | Returns an empty tensor sequence. |

docs/framework/operators/tensor/tensor.gather_elements.md

@@ -0,0 +1,47 @@
+# tensor.gather_elements
+
+```rust 
+   fn gather_elements(self: @Tensor<T>, indices: Tensor<T>, axis: Option<usize>) -> Tensor<T>;
+```
+
+GatherElements is an indexing operation that produces its output by indexing into the input data tensor at index positions determined by elements of the indices tensor.
+
+## Args
+
+* `self`(`@Tensor<T>`) - The input tensor.
+* `indices`(`Tensor<T>`) - Tensor of indices.
+* `axis`(`Option<usize>`) - Axis to gather_elements on. Default: axis=0.
+
+## Panics
+
+* Panics if index values are not within bounds [-s, s-1] along axis of size s.
+
+## Returns 
+
+A new `Tensor<T>` .
+
+## Example
+
+```rust
+use array::{ArrayTrait, SpanTrait};
+
+use orion::operators::tensor::{TensorTrait, Tensor, U32Tensor};
+
+fn gather_elements_example() -> Tensor<u32> {
+    let tensor = TensorTrait::<u32>::new(
+        shape: array![3, 3].span(), 
+        data: array![[ 1, 2, 3],[4, 5, 6], [7, 8, 9]].span(), 
+    );
+    let indices = TensorTrait::<u32>::new(
+        shape: array![1, 2, 0].span(), 
+        data: array![2, 0, 0].span(), 
+    );
+
+    return tensor.gather_elements(
+        indices: indices, 
+        axis: Option::None(()), 
+    );
+}
+>>> [[4. 8. 3.]
+     [7. 2. 3.]]
+```

nodegen/node/gather_elements.py

@@ -0,0 +1,268 @@
+import numpy as np
+from nodegen.node import RunAll
+from ..helpers import make_test, to_fp, Tensor, Dtype, FixedImpl, Trait
+
+def gather_elements(data, indices, axis=0):  # type: ignore
+    data_swaped = np.swapaxes(data, 0, axis)
+    index_swaped = np.swapaxes(indices, 0, axis)
+    gathered = np.choose(index_swaped, data_swaped, mode="wrap")
+    y = np.swapaxes(gathered, 0, axis)
+    return y
+
+class Gather_elements(RunAll):
+
+    @staticmethod
+    def gather_elements_fp16x16():
+        def gather_elements_3D():
+            def default():
+                x1 = np.arange(0,27).reshape(3,3,3).astype(np.int64)
+                x2 = np.random.randint(low = 0,high=2, size=(3,3,3)).astype(np.uint32)
+                y = gather_elements(x1, x2, axis=0)
+
+                x1 = Tensor(Dtype.FP16x16, x1.shape, to_fp(x1.flatten(), FixedImpl.FP16x16))
+                x2 = Tensor(Dtype.U32, x2.shape, x2.flatten()) 
+                y = Tensor(Dtype.FP16x16, y.shape, to_fp(
+                y.flatten(), FixedImpl.FP16x16)) 
+
+                name = "gather_elements_fp16x16_3d_default"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(0))", 
+                    name= name)
+
+            def axis1():
+                x1 = np.arange(0,27).reshape(3,3,3).astype(np.int64)
+                x2 = np.random.randint(low = 0,high=3, size=(3,3,3)).astype(np.uint32)
+                y = gather_elements(x1, x2, axis=1)
+
+                x1 = Tensor(Dtype.FP16x16, x1.shape, to_fp(x1.flatten(), FixedImpl.FP16x16))
+                x2 = Tensor(Dtype.U32, x2.shape, x2.flatten()) 
+                y = Tensor(Dtype.FP16x16, y.shape, to_fp(
+                y.flatten(), FixedImpl.FP16x16)) 
+
+                name = "gather_elements_fp16x16_3d_axis1"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(1))", 
+                    name= name)
+
+            def axis2():
+                x1 = np.arange(0,27).reshape(3,3,3).astype(np.int64)
+                x2 = np.random.randint(low = 0,high=3, size=(3,3,3)).astype(np.uint32)
+                y = gather_elements(x1, x2, axis=2)
+
+                x1 = Tensor(Dtype.FP16x16, x1.shape, to_fp(x1.flatten(), FixedImpl.FP16x16))
+                x2 = Tensor(Dtype.U32, x2.shape, x2.flatten()) 
+                y = Tensor(Dtype.FP16x16, y.shape, to_fp(
+                y.flatten(), FixedImpl.FP16x16)) 
+
+                name = "gather_elements_fp16x16_3d_axis2"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(2))", 
+                    name= name)
+
+            default()
+            axis1()
+            axis2()
+        gather_elements_3D()
+
+
+    @staticmethod
+    def gather_elements_fp8x23():
+        def gather_elements_3D():
+            def default():
+                x1 = np.arange(0,27).reshape(3,3,3).astype(np.int64)
+                x2 = np.random.randint(low = 0,high=2, size=(3,3,3)).astype(np.int64)
+                y = gather_elements(x1, x2, axis=0)
+
+                x1 = Tensor(Dtype.FP8x23, x1.shape, to_fp(x1.flatten(), FixedImpl.FP8x23))
+                x2 = Tensor(Dtype.U32, x2.shape, x2.flatten()) 
+                y = Tensor(Dtype.FP8x23, y.shape, to_fp(y.flatten(), FixedImpl.FP8x23))
+
+                name = "gather_elements_fp8x23_3d_default"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(0))", 
+                    name= name)
+
+            def axis1():
+                x1 = np.arange(0,27).reshape(3,3,3).astype(np.int64)
+                x2 = np.random.randint(low = 0,high=3, size=(3,3,3)).astype(np.int64)
+                y = gather_elements(x1, x2, axis=1)
+
+                x1 = Tensor(Dtype.FP8x23, x1.shape, to_fp(x1.flatten(), FixedImpl.FP8x23))
+                x2 = Tensor(Dtype.U32, x2.shape, x2.flatten()) 
+                y = Tensor(Dtype.FP8x23, y.shape, to_fp(y.flatten(), FixedImpl.FP8x23))
+
+                name = "gather_elements_fp8x23_3d_axis1"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(1))", 
+                    name= name)
+
+            def axis2():
+                x1 = np.arange(0,27).reshape(3,3,3).astype(np.int64)
+                x2 = np.random.randint(low = 0,high=3, size=(3,3,3)).astype(np.int64)
+                y = gather_elements(x1, x2, axis=2)
+
+                x1 = Tensor(Dtype.FP8x23, x1.shape, to_fp(x1.flatten(), FixedImpl.FP8x23))
+                x2 = Tensor(Dtype.U32, x2.shape, x2.flatten()) 
+                y = Tensor(Dtype.FP8x23, y.shape, to_fp(y.flatten(), FixedImpl.FP8x23))
+
+                name = "gather_elements_fp8x23_3d_axis2"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(2))", 
+                    name= name)
+
+            default()
+            axis1()
+            axis2()
+        gather_elements_3D()
+
+
+    @staticmethod
+    def gather_elements_i8():
+        def gather_elements_3D():
+            def default():
+                x1 = np.arange(0,9).reshape(3,3).astype(np.int8)
+                x2 = np.random.randint(low = 0,high=2, size=(3,3)).astype(np.int8)
+                y = gather_elements(x1, x2, axis=0)
+
+                x1 =  Tensor(Dtype.I8, x1.shape, x1.flatten())
+                x2 = Tensor(Dtype.U32, x2.shape, x2.flatten()) 
+                y =  Tensor(Dtype.I8, y.shape, y.flatten()) 
+
+                name = "gather_elements_i8_3d_default"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(0))", 
+                    name= name)
+
+            def axis1():
+                x1 = np.arange(0,9).reshape(3,3).astype(np.int8)
+                x2 = np.random.randint(low = 0,high=2, size=(3,3)).astype(np.int8)
+                y = gather_elements(x1, x2, axis=1)
+
+                x1 =  Tensor(Dtype.I8, x1.shape, x1.flatten())
+                x2 = Tensor(Dtype.U32, x2.shape, x2.flatten()) 
+                y =  Tensor(Dtype.I8, y.shape, y.flatten()) 
+
+                name = "gather_elements_i8_3d_axis1"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(1))", 
+                    name= name)
+
+            default()
+            axis1()
+        gather_elements_3D()
+
+
+    @staticmethod
+    def gather_elements_i32():
+        def gather_elements_3D():
+            def default():
+                x1 = np.arange(0,24).reshape(4,2,3).astype(np.int32)
+                x2 = np.random.randint(low = 0,high=2, size=(5,2,3)).astype(np.int32)
+                y = gather_elements(x1, x2, axis=0)
+
+                x1 =  Tensor(Dtype.I32, x1.shape, x1.flatten())
+                x2 = Tensor(Dtype.U32, x2.shape, x2.flatten()) 
+                y =  Tensor(Dtype.I32, y.shape, y.flatten()) 
+
+                name = "gather_elements_i32_3d_default"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(0))", 
+                    name= name)
+
+            def axis1():
+                x1 = np.arange(0,24).reshape(4,2,3).astype(np.int32)
+                x2 = np.random.randint(low = 0,high=2, size=(4,3,3)).astype(np.int32)
+                y = gather_elements(x1, x2, axis=1)
+
+                x1 =  Tensor(Dtype.I32, x1.shape, x1.flatten())
+                x2 = Tensor(Dtype.U32, x2.shape, x2.flatten()) 
+                y =  Tensor(Dtype.I32, y.shape, y.flatten()) 
+
+                name = "gather_elements_i32_3d_axis1"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(1))", 
+                    name= name)
+
+            def axis2():
+                x1 = np.arange(0,24).reshape(4,2,3).astype(np.int32)
+                x2 = np.random.randint(low = 0,high=2, size=(4,2,4)).astype(np.int32)
+                y = gather_elements(x1, x2, axis=2)
+
+                x1 =  Tensor(Dtype.I32, x1.shape, x1.flatten())
+                x2 = Tensor(Dtype.U32, x2.shape, x2.flatten()) 
+                y =  Tensor(Dtype.I32, y.shape, y.flatten()) 
+
+                name = "gather_elements_i32_3d_axis2"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(2))", 
+                    name= name)
+
+            default()
+            axis1()
+            axis2()
+        gather_elements_3D()
+
+    @staticmethod
+    def gather_elements_u32():
+        def gather_elements_3D():
+            def default():
+                x1 = np.arange(0,108).reshape(3,3,4,3).astype(np.int32)
+                x2 = np.random.randint(low = 0,high=3, size=(10,3,4,3)).astype(np.int32)
+                y = gather_elements(x1, x2, axis=0)
+
+                x1 =  Tensor(Dtype.U32, x1.shape, x1.flatten())
+                x2 =  Tensor(Dtype.U32, x2.shape, x2.flatten())
+                y =  Tensor(Dtype.U32, y.shape, y.flatten()) 
+
+                name = "gather_elements_u32_default"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(0))", 
+                    name= name)
+
+            def axis1():
+                x1 = np.arange(0,108).reshape(3,3,4,3).astype(np.int32)
+                x2 = np.random.randint(low = 0,high=3, size=(3,5,4,3)).astype(np.int32)
+                y = gather_elements(x1, x2, axis=1)
+
+                x1 =  Tensor(Dtype.U32, x1.shape, x1.flatten())
+                x2 =  Tensor(Dtype.U32, x2.shape, x2.flatten())
+                y =  Tensor(Dtype.U32, y.shape, y.flatten()) 
+
+                name = "gather_elements_u32_axis1"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(1))", 
+                    name= name)
+
+            def axis2():
+                x1 = np.arange(0,108).reshape(3,3,4,3).astype(np.int32)
+                x2 = np.random.randint(low = 0,high=3, size=(3,3,4,3)).astype(np.int32)
+                y = gather_elements(x1, x2, axis=2)
+
+                x1 =  Tensor(Dtype.U32, x1.shape, x1.flatten())
+                x2 =  Tensor(Dtype.U32, x2.shape, x2.flatten())
+                y =  Tensor(Dtype.U32, y.shape, y.flatten()) 
+
+                name = "gather_elements_u32_axis2"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(2))", 
+                    name= name)
+
+            def axis3():
+                x1 = np.arange(0,108).reshape(3,3,4,3).astype(np.int32)
+                x2 = np.random.randint(low = 0,high=3, size=(3,3,4,6)).astype(np.int32)
+                y = gather_elements(x1, x2, axis=3)
+
+                x1 =  Tensor(Dtype.U32, x1.shape, x1.flatten())
+                x2 =  Tensor(Dtype.U32, x2.shape, x2.flatten())
+                y =  Tensor(Dtype.U32, y.shape, y.flatten()) 
+
+                name = "gather_elements_u32_axis3"
+                make_test(
+                    inputs = [x1, x2], output = y, func_sig = "input_0.gather_elements(indices:input_1, axis:Option::Some(3))", 
+                    name= name)
+
+            default()
+            axis1()
+            axis2()
+            axis3()
+        gather_elements_3D()