[ONNX] Extend ONNX frontend with SoftmaxCrossEntropyLoss function #20547 #22650
Conversation
i Updates onxx_import.in.cpp
There was a problem hiding this comment.
Hello @tanishka321, thank you for your contribution.
Decomposition subgraph can contain only extisting OpenVINO operators,
it's not possible to create default_opset::SoftmaxCrossEntropyLoss as such class doesn't exist (so such code can't be compiled). SoftmaxCrossEntropyLoss operation is not available as OpenVINO primitive operator (list of ov ops), that's why the ONNX Function expand mechanism is used.
@gkrivor As ONNX SoftmaxCrossEntropyLoss is training related operator, I would suggest to verify it's applicability for custom decomposition in OpenVINO ONNX FE.
| const auto coerced_logits = ov::op::util::flatten(logits, static_cast<int>(axis)); | ||
| const auto coerced_labels = ov::op::util::flatten(labels, static_cast<int>(axis)); | ||
|
|
||
| const auto result = std::make_shared<default_opset::SoftmaxCrossEntropyLoss>(coerced_logits, coerced_labels, 1); |
There was a problem hiding this comment.
It's not possible to create default_opset::SoftmaxCrossEntropyLoss as such class doesn't exist. SoftmaxCrossEntropyLoss operation is not available as OpenVINO primitive operator (list of ov ops), this is why the ONNX Function expand mechanism is used.
Decomposition should follow or optimize the current subgraph produced by ONNX Frontend.
| namespace set_1 { | ||
| OutputVector softmax_cross_entropy_loss(const Node& node) { | ||
| const auto logits = node.get_ng_inputs().at(0); | ||
| const auto labels = node.get_ng_inputs().at(1); | ||
|
|
||
| const auto axis = node.get_attribute_value<int64_t>("axis", 1); | ||
|
|
||
| std::shared_ptr<ngraph::Node> result; | ||
| result = onnx_softmax_cross_entropy_loss(logits, labels, axis); | ||
|
|
||
| return {result}; | ||
| } | ||
| } // namespace set_1 |
There was a problem hiding this comment.
This code is repeated 3 times for each namespace (set_1, set_11, set_13). It was a good direction to create common helper function onnx_softmax_cross_entropy_loss and reuse it, but more code can be unified. In fact there is no difference between the versions, just bfloat16 support has been added to the ONNX spec, so the import code can be shared between versions.
https://onnx.ai/onnx/operators/text_diff_SoftmaxCrossEntropyLoss_12_13.html
| const auto logits = node.get_ng_inputs().at(0); | ||
| const auto labels = node.get_ng_inputs().at(1); |
There was a problem hiding this comment.
It's better to call node.get_ng_inputs() once and assign it to a variable, then take required input by index.
| auto test_case = ov::test::TestCase(function, s_device); | ||
| test_case.add_input<int64_t>(Shape{2}, {1, 2}); | ||
| test_case.add_input<int64_t>(Shape{2}, {0, 2}); | ||
| test_case.add_expected_output<bool>(Shape{2}, {false, true}); |
There was a problem hiding this comment.
Why the type of the expected output values is bool?
According to the ONNX SoftmaxCrossEntropyLoss specification, it should be floating-point precision.
| #include "openvino/core/deprecated.hpp" | ||
| OPENVINO_SUPPRESS_DEPRECATED_START | ||
|
|
||
| #include "ngraph/node.hpp" | ||
| #include "onnx_import/core/node.hpp" | ||
|
|
||
| namespace ngraph { | ||
| namespace onnx_import { | ||
| namespace op { | ||
| namespace set_1 { | ||
|
|
||
| OutputVector softmax_cross_entropy_loss(const Node& node); |
There was a problem hiding this comment.
Please sync with the latest master. ONNX FE has been slightly refactored. The changes (namespaces, headers) need to be aligned with the rest of the frontends/onnx files.
Any of the onnx frontend op file can be taken as an example to follow, like:
/src/frontends/onnx/frontend/src/op/leaky_relu.hpp
/src/frontends/onnx/frontend/src/op/leaky_relu.cpp
|
This PR will be closed in a week because of 2 weeks of no activity. |
|
This PR was closed because it has been stalled for 2 week with no activity. |
Details:
Extend ONNX frontend with SoftmaxCrossEntropyLoss function
Tickets:
Closes #20547