Implement aten::index | feat(torchlib) (#862) #883

justinchuby · 2023-07-17T17:13:13Z

Stack from ghstack (oldest at bottom):

-> Implement aten::index | feat(torchlib) (#862) #883

This change implements the logic for aten::index and adds tests for different nd index combinations and permutations.

Understanding `aten::index`

For arg0 with shape [7, 3, 4, 5, 6]
The indexing operation arg0[0, :, 1:2, tensor([[4,5]])] will be translated to

+>  select: i64[3, 4, 5, 6] = torch.ops.aten.select.int(arg0, 0, 0);
+>  slice_1: i64[3, 4, 5, 6] = torch.ops.aten.slice.Tensor(select, 0, 0, 9223372036854775807);
+>  slice_2: i64[3, 1, 5, 6] = torch.ops.aten.slice.Tensor(slice_1, 1, 1, 2);
+>  index: i64[3, 1, 1, 2, 6] = torch.ops.aten.index.Tensor(slice_2, [None, None, arg1]);

Here,

indices = [None, None, arg1] is equivalent to indices = [None, None, arg1, None]
The operation arg0[0, :, 1:2, tensor([[4,5]])] is equivalent to arg0[0, :, 1:2, tensor([[4,5]]), :]
None in indices are like fillers for dimensions that cannot be removed in the process.

Gather op reference

Co-authored-by: BowenBao [email protected]

[ghstack-poisoned]

Remove the duplicate check_model message when displaying [ghstack-poisoned]

WIP: need more tests Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <[email protected]> [ghstack-poisoned]

WIP: need more tests Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> ghstack-source-id: a991236 Pull Request resolved: #883

Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> [ghstack-poisoned]

WIP: need more tests Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> ghstack-source-id: 8755417 Pull Request resolved: #883

Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> [ghstack-poisoned]

WIP: need more tests Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> ghstack-source-id: 078cf56 Pull Request resolved: #883 Signed-off-by: Justin Chu <[email protected]>

Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> [ghstack-poisoned]

WIP: need more tests Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> ghstack-source-id: e641533 Pull Request resolved: #883 Signed-off-by: Justin Chu <[email protected]>

--- **This change implements the logic for `aten::index` and adds tests for different nd index combinations and permutations.** ## Understanding `aten::index` For `arg0` with shape `[7, 3, 4, 5, 6]` The indexing operation `arg0[0, :, 1:2, tensor([[4,5]])]` will be translated to ``` +> select: i64[3, 4, 5, 6] = torch.ops.aten.select.int(arg0, 0, 0); +> slice_1: i64[3, 4, 5, 6] = torch.ops.aten.slice.Tensor(select, 0, 0, 9223372036854775807); +> slice_2: i64[3, 1, 5, 6] = torch.ops.aten.slice.Tensor(slice_1, 1, 1, 2); +> index: i64[3, 1, 1, 2, 6] = torch.ops.aten.index.Tensor(slice_2, [None, None, arg1]); ``` Here, - `indices = [None, None, arg1]` is equivalent to `indices = [None, None, arg1, None]` - The operation `arg0[0, :, 1:2, tensor([[4,5]])]` is equivalent to `arg0[0, :, 1:2, tensor([[4,5]]), :]` None in `indices` are like fillers for dimensions that cannot be removed in the process. ## Gather op reference - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> [ghstack-poisoned]

WIP: need more tests Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> ghstack-source-id: 194baee Pull Request resolved: #883 Signed-off-by: Justin Chu <[email protected]>

BowenBao

🎉 LGTM. Do you plan to add bool mask indices support in this or another PR?

BowenBao · 2023-07-17T18:00:38Z

onnxscript/tests/function_libs/torch_lib/ops_test_common.py

@@ -473,6 +471,8 @@ def _capture_graph_and_evaluate_torch_script_evaluator(function: Callable, args,
                        input.value = subarg
                        sequence_input.append(input)
                        ort_inputs[input_name] = subarg
+                    else:
+                        sequence_input.append(subarg)


Let's put comment here explaining why this is needed in case we forget?

Should we consider elif None? It's easier to catch what we are not expecting.

Good point. I wonder if there are things we don't expect that can sneak in? Nested lists?

No idea. Just in case.

SG. Created #892 for now.

justinchuby · 2023-07-17T18:06:19Z

🎉 LGTM. Do you plan to add bool mask indices support in this or another PR?

I plan to do it in another PR to keep this one manageable.

codecov · 2023-07-17T18:07:39Z

Codecov Report

Merging #883 (169debc) into main (bbdbf1e) will decrease coverage by 0.09%.
The diff coverage is n/a.

❗ Current head 169debc differs from pull request most recent head 9d941c5. Consider uploading reports for the commit 9d941c5 to get more accurate results

@@            Coverage Diff             @@
##             main     #883      +/-   ##
==========================================
- Coverage   76.59%   76.51%   -0.09%     
==========================================
  Files         112      112              
  Lines       13408    13408              
  Branches     1348     1348              
==========================================
- Hits        10270    10259      -11     
- Misses       2801     2812      +11     
  Partials      337      337

see 1 file with indirect coverage changes

onnxscript/tests/function_libs/torch_lib/ops_test_common.py

) Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #883 * #882 * __->__ #881 The torchscript ONNX graph generator creates numeric value names by default (`0`, `1`). These are not legal ONNX tensor names, since ONNX requires the names to be valid C variable names. This change updates the names by prepending a prefix `_val_` or `_const_` to make them valid ONNX names. It also improves readability by making the names less likely to be confused with shape values. I decided to use the `_` prefix to reduce the chance of name collision with FX names. After: ``` < ir_version: 8, opset_import: ["" : 18], producer_name: "pytorch", producer_version: "2.1.0" > torch_jit (float[5,5,5,5] input_0, int64[2] input_1_3) => (float[5,5,5,2] _val_10) { _val_2 = Transpose <perm = [0, 1, 2, 3]> (input_0) _val_3 = Max (input_1_3) _val_4 = Shape <start = 0> (_val_3) _val_5 = Expand (input_1_3, _val_4) _const_6 = Constant <value = int64 {-1}> () _val_7 = Unsqueeze (_val_5, _const_6) _val_8 = Concat <axis = -1> (_val_7) _val_9 = GatherND <batch_dims = 0> (_val_2, _val_8) _val_10 = Transpose <perm = [0, 1, 2, 3]> (_val_9) } ``` Before: ``` < ir_version: 8, opset_import: ["" : 18], producer_name: "pytorch", producer_version: "2.1.0" > torch_jit (float[5,5,5,5] input_0, int64[2] input_1_3) => (float[5,5,5,2] 10) { 2 = Transpose <perm = [0, 1, 2, 3]> (input_0) 3 = Max (input_1_3) 4 = Shape <start = 0> (3) 5 = Expand (input_1_3, 4) 6 = Constant <value = int64 {-1}> () 7 = Unsqueeze (5, 6) 8 = Concat <axis = -1> (7) 9 = GatherND <batch_dims = 0> (2, 8) 10 = Transpose <perm = [0, 1, 2, 3]> (9) } ```

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #883 * __->__ #882 * #881 Remove the duplicate check_model message when displaying

…)" --- **This change implements the logic for `aten::index` and adds tests for different nd index combinations and permutations.** ## Understanding `aten::index` For `arg0` with shape `[7, 3, 4, 5, 6]` The indexing operation `arg0[0, :, 1:2, tensor([[4,5]])]` will be translated to ``` +> select: i64[3, 4, 5, 6] = torch.ops.aten.select.int(arg0, 0, 0); +> slice_1: i64[3, 4, 5, 6] = torch.ops.aten.slice.Tensor(select, 0, 0, 9223372036854775807); +> slice_2: i64[3, 1, 5, 6] = torch.ops.aten.slice.Tensor(slice_1, 1, 1, 2); +> index: i64[3, 1, 1, 2, 6] = torch.ops.aten.index.Tensor(slice_2, [None, None, arg1]); ``` Here, - `indices = [None, None, arg1]` is equivalent to `indices = [None, None, arg1, None]` - The operation `arg0[0, :, 1:2, tensor([[4,5]])]` is equivalent to `arg0[0, :, 1:2, tensor([[4,5]]), :]` None in `indices` are like fillers for dimensions that cannot be removed in the process. ## Gather op reference - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> [ghstack-poisoned]

--- **This change implements the logic for `aten::index` and adds tests for different nd index combinations and permutations.** ## Understanding `aten::index` For `arg0` with shape `[7, 3, 4, 5, 6]` The indexing operation `arg0[0, :, 1:2, tensor([[4,5]])]` will be translated to ``` +> select: i64[3, 4, 5, 6] = torch.ops.aten.select.int(arg0, 0, 0); +> slice_1: i64[3, 4, 5, 6] = torch.ops.aten.slice.Tensor(select, 0, 0, 9223372036854775807); +> slice_2: i64[3, 1, 5, 6] = torch.ops.aten.slice.Tensor(slice_1, 1, 1, 2); +> index: i64[3, 1, 1, 2, 6] = torch.ops.aten.index.Tensor(slice_2, [None, None, arg1]); ``` Here, - `indices = [None, None, arg1]` is equivalent to `indices = [None, None, arg1, None]` - The operation `arg0[0, :, 1:2, tensor([[4,5]])]` is equivalent to `arg0[0, :, 1:2, tensor([[4,5]]), :]` None in `indices` are like fillers for dimensions that cannot be removed in the process. ## Gather op reference - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> [ghstack-poisoned]

WIP: need more tests Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> ghstack-source-id: 4b050c3 Pull Request resolved: #883 Signed-off-by: Justin Chu <[email protected]>

…)" --- **This change implements the logic for `aten::index` and adds tests for different nd index combinations and permutations.** ## Understanding `aten::index` For `arg0` with shape `[7, 3, 4, 5, 6]` The indexing operation `arg0[0, :, 1:2, tensor([[4,5]])]` will be translated to ``` +> select: i64[3, 4, 5, 6] = torch.ops.aten.select.int(arg0, 0, 0); +> slice_1: i64[3, 4, 5, 6] = torch.ops.aten.slice.Tensor(select, 0, 0, 9223372036854775807); +> slice_2: i64[3, 1, 5, 6] = torch.ops.aten.slice.Tensor(slice_1, 1, 1, 2); +> index: i64[3, 1, 1, 2, 6] = torch.ops.aten.index.Tensor(slice_2, [None, None, arg1]); ``` Here, - `indices = [None, None, arg1]` is equivalent to `indices = [None, None, arg1, None]` - The operation `arg0[0, :, 1:2, tensor([[4,5]])]` is equivalent to `arg0[0, :, 1:2, tensor([[4,5]]), :]` None in `indices` are like fillers for dimensions that cannot be removed in the process. ## Gather op reference - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> [ghstack-poisoned]

--- **This change implements the logic for `aten::index` and adds tests for different nd index combinations and permutations.** ## Understanding `aten::index` For `arg0` with shape `[7, 3, 4, 5, 6]` The indexing operation `arg0[0, :, 1:2, tensor([[4,5]])]` will be translated to ``` +> select: i64[3, 4, 5, 6] = torch.ops.aten.select.int(arg0, 0, 0); +> slice_1: i64[3, 4, 5, 6] = torch.ops.aten.slice.Tensor(select, 0, 0, 9223372036854775807); +> slice_2: i64[3, 1, 5, 6] = torch.ops.aten.slice.Tensor(slice_1, 1, 1, 2); +> index: i64[3, 1, 1, 2, 6] = torch.ops.aten.index.Tensor(slice_2, [None, None, arg1]); ``` Here, - `indices = [None, None, arg1]` is equivalent to `indices = [None, None, arg1, None]` - The operation `arg0[0, :, 1:2, tensor([[4,5]])]` is equivalent to `arg0[0, :, 1:2, tensor([[4,5]]), :]` None in `indices` are like fillers for dimensions that cannot be removed in the process. ## Gather op reference - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> [ghstack-poisoned]

WIP: need more tests Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> ghstack-source-id: b22c3f8 Pull Request resolved: #883 Signed-off-by: Justin Chu <[email protected]>

…)" --- **This change implements the logic for `aten::index` and adds tests for different nd index combinations and permutations.** ## Understanding `aten::index` For `arg0` with shape `[7, 3, 4, 5, 6]` The indexing operation `arg0[0, :, 1:2, tensor([[4,5]])]` will be translated to ``` +> select: i64[3, 4, 5, 6] = torch.ops.aten.select.int(arg0, 0, 0); +> slice_1: i64[3, 4, 5, 6] = torch.ops.aten.slice.Tensor(select, 0, 0, 9223372036854775807); +> slice_2: i64[3, 1, 5, 6] = torch.ops.aten.slice.Tensor(slice_1, 1, 1, 2); +> index: i64[3, 1, 1, 2, 6] = torch.ops.aten.index.Tensor(slice_2, [None, None, arg1]); ``` Here, - `indices = [None, None, arg1]` is equivalent to `indices = [None, None, arg1, None]` - The operation `arg0[0, :, 1:2, tensor([[4,5]])]` is equivalent to `arg0[0, :, 1:2, tensor([[4,5]]), :]` None in `indices` are like fillers for dimensions that cannot be removed in the process. ## Gather op reference - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> [ghstack-poisoned]

--- **This change implements the logic for `aten::index` and adds tests for different nd index combinations and permutations.** ## Understanding `aten::index` For `arg0` with shape `[7, 3, 4, 5, 6]` The indexing operation `arg0[0, :, 1:2, tensor([[4,5]])]` will be translated to ``` +> select: i64[3, 4, 5, 6] = torch.ops.aten.select.int(arg0, 0, 0); +> slice_1: i64[3, 4, 5, 6] = torch.ops.aten.slice.Tensor(select, 0, 0, 9223372036854775807); +> slice_2: i64[3, 1, 5, 6] = torch.ops.aten.slice.Tensor(slice_1, 1, 1, 2); +> index: i64[3, 1, 1, 2, 6] = torch.ops.aten.index.Tensor(slice_2, [None, None, arg1]); ``` Here, - `indices = [None, None, arg1]` is equivalent to `indices = [None, None, arg1, None]` - The operation `arg0[0, :, 1:2, tensor([[4,5]])]` is equivalent to `arg0[0, :, 1:2, tensor([[4,5]]), :]` None in `indices` are like fillers for dimensions that cannot be removed in the process. ## Gather op reference - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> [ghstack-poisoned]

WIP: need more tests Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> ghstack-source-id: 1c8f0e2 Pull Request resolved: #883 Signed-off-by: Justin Chu <[email protected]>

thiagocrepaldi

LGTM

thiagocrepaldi · 2023-07-17T18:30:51Z

onnxscript/tests/function_libs/torch_lib/extra_opinfo.py

+    index_1d = common_methods_invocations.index_variable(2, s, device=device)
+    index_2d = common_methods_invocations.index_variable((s + 1, 2), s, device=device)
+    index_3d = common_methods_invocations.index_variable((s + 2, s + 1, 2), s, device=device)
+    test_args = [


would itertools.{product,permutation,combination} help in ensuring all combinations are covered and make the code shorter?

Good point. Let me try that

So I tested with itertools but realized some combinations are invalid so we cannot enumerate them all using itertools. For the sake of clarity I propose that we keep the current explicit tests. I also added more test cases and comments

thiagocrepaldi · 2023-07-17T18:38:02Z

onnxscript/tests/function_libs/torch_lib/ops_test_common.py

@@ -276,10 +277,7 @@ def convert_tensor_to_numpy(input: Any) -> Any:


 def convert_kwargs_for_onnx(kwargs: dict[str, Any]) -> dict[str, Any]:
-    """Converts kwargs to be compatible with ONNX Runtime.
-
-    ONNX Runtime doesn't support torch.bool, so we convert them to torch.uint8.


for my knowledge, does ort support torch.bool now? or was this docstring outdated already?

I actually don't know if ORT supports bool (it should?), but I think this message was a mistake by copilot because we don't actually have this conversion logic as code. If we see issues with ORT I will make adjustments.

thiagocrepaldi · 2023-07-17T18:39:26Z

onnxscript/tests/function_libs/torch_lib/extra_opinfo.py

+    index_1d = common_methods_invocations.index_variable(2, s, device=device)
+    index_2d = common_methods_invocations.index_variable((s + 1, 2), s, device=device)
+    index_3d = common_methods_invocations.index_variable((s + 2, s + 1, 2), s, device=device)
+    test_args = [


could itertools.product (or friends) with length 1 to 4 could shorten this listen and ensure no combination is left out?

As above. Turns out some combinations are invalid to torch and it may be better to specify explicitly.

…)" --- **This change implements the logic for `aten::index` and adds tests for different nd index combinations and permutations.** ## Understanding `aten::index` For `arg0` with shape `[7, 3, 4, 5, 6]` The indexing operation `arg0[0, :, 1:2, tensor([[4,5]])]` will be translated to ``` +> select: i64[3, 4, 5, 6] = torch.ops.aten.select.int(arg0, 0, 0); +> slice_1: i64[3, 4, 5, 6] = torch.ops.aten.slice.Tensor(select, 0, 0, 9223372036854775807); +> slice_2: i64[3, 1, 5, 6] = torch.ops.aten.slice.Tensor(slice_1, 1, 1, 2); +> index: i64[3, 1, 1, 2, 6] = torch.ops.aten.index.Tensor(slice_2, [None, None, arg1]); ``` Here, - `indices = [None, None, arg1]` is equivalent to `indices = [None, None, arg1, None]` - The operation `arg0[0, :, 1:2, tensor([[4,5]])]` is equivalent to `arg0[0, :, 1:2, tensor([[4,5]]), :]` None in `indices` are like fillers for dimensions that cannot be removed in the process. ## Gather op reference - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> [ghstack-poisoned]

--- **This change implements the logic for `aten::index` and adds tests for different nd index combinations and permutations.** ## Understanding `aten::index` For `arg0` with shape `[7, 3, 4, 5, 6]` The indexing operation `arg0[0, :, 1:2, tensor([[4,5]])]` will be translated to ``` +> select: i64[3, 4, 5, 6] = torch.ops.aten.select.int(arg0, 0, 0); +> slice_1: i64[3, 4, 5, 6] = torch.ops.aten.slice.Tensor(select, 0, 0, 9223372036854775807); +> slice_2: i64[3, 1, 5, 6] = torch.ops.aten.slice.Tensor(slice_1, 1, 1, 2); +> index: i64[3, 1, 1, 2, 6] = torch.ops.aten.index.Tensor(slice_2, [None, None, arg1]); ``` Here, - `indices = [None, None, arg1]` is equivalent to `indices = [None, None, arg1, None]` - The operation `arg0[0, :, 1:2, tensor([[4,5]])]` is equivalent to `arg0[0, :, 1:2, tensor([[4,5]]), :]` None in `indices` are like fillers for dimensions that cannot be removed in the process. ## Gather op reference - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> [ghstack-poisoned]

WIP: need more tests Gather op: - https://github.com/openxla/xla/blob/main/docs/operation_semantics.md?rgh-link-date=2023-07-13T01%3A09%3A16Z#gather - https://www.pathpartnertech.com/gather-scatter-operation-in-deep-learning-framework/ --------- Co-authored-by: BowenBao <bowbaomicrosoft.com> ghstack-source-id: 4e36b48 Pull Request resolved: #883 Signed-off-by: Justin Chu <[email protected]>

justinchuby added 3 commits July 17, 2023 17:09

Rename intermediate values to _val and _const | feat(torchlib)

5bde197

[ghstack-poisoned]

Simplify model error message in test | test(torchlib)

3ff6ee2

Remove the duplicate check_model message when displaying [ghstack-poisoned]

This was referenced Jul 17, 2023

Rename intermediate values to _val and _const | feat(torchlib) #881

Merged

Simplify model error message in test | test(torchlib) #882

Merged

justinchuby mentioned this pull request Jul 17, 2023

Implement aten::index | feat(torchlib) #862

Merged

justinchuby requested a review from BowenBao July 17, 2023 17:13

justinchuby marked this pull request as draft July 17, 2023 17:14

justinchuby marked this pull request as ready for review July 17, 2023 17:47

justinchuby added module: torchlib Related to the torch/aten function lib in development change base before merge Remember to change the merge base to main when the PR is ready to merge labels Jul 17, 2023

justinchuby requested review from xiaowuhu and fatcat-z July 17, 2023 18:00

BowenBao approved these changes Jul 17, 2023

View reviewed changes

justinchuby commented Jul 17, 2023

View reviewed changes

onnxscript/tests/function_libs/torch_lib/ops_test_common.py Show resolved Hide resolved

justinchuby changed the base branch from gh/justinchuby/33/base to main July 17, 2023 18:11

justinchuby added 2 commits July 17, 2023 18:16

justinchuby added 2 commits July 17, 2023 18:19

thiagocrepaldi approved these changes Jul 17, 2023

View reviewed changes

justinchuby mentioned this pull request Jul 17, 2023

ATen ops list #258

Open

justinchuby added 2 commits July 17, 2023 21:40

justinchuby merged commit 2e0aee7 into main Jul 17, 2023

justinchuby deleted the gh/justinchuby/33/head branch July 17, 2023 21:52

This was referenced Jul 17, 2023

[torchlib] Explicitly check for None in sequence inputs #892

Open

[OpInfo] index.Tensor pytorch/pytorch#105483

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implement aten::index | feat(torchlib) (#862) #883

Implement aten::index | feat(torchlib) (#862) #883

justinchuby commented Jul 17, 2023 •

edited

Loading

BowenBao left a comment

BowenBao Jul 17, 2023

justinchuby Jul 17, 2023

titaiwangms Jul 17, 2023

justinchuby Jul 17, 2023

titaiwangms Jul 17, 2023

justinchuby Jul 17, 2023

justinchuby commented Jul 17, 2023

codecov bot commented Jul 17, 2023 •

edited

Loading

thiagocrepaldi left a comment

thiagocrepaldi Jul 17, 2023

justinchuby Jul 17, 2023

justinchuby Jul 17, 2023

thiagocrepaldi Jul 17, 2023

justinchuby Jul 17, 2023 •

edited

Loading

thiagocrepaldi Jul 17, 2023

justinchuby Jul 17, 2023

Implement aten::index | feat(torchlib) (#862) #883

Implement aten::index | feat(torchlib) (#862) #883

Conversation

justinchuby commented Jul 17, 2023 • edited Loading

Understanding aten::index

Gather op reference

BowenBao left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

justinchuby commented Jul 17, 2023

codecov bot commented Jul 17, 2023 • edited Loading

Codecov Report

thiagocrepaldi left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

justinchuby Jul 17, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

justinchuby commented Jul 17, 2023 •

edited

Loading

Understanding `aten::index`

codecov bot commented Jul 17, 2023 •

edited

Loading

justinchuby Jul 17, 2023 •

edited

Loading