[Good First Issue] [TF FE]: Support Atan2 operation for TensorFlow models

[rkazants]

Context

OpenVINO component responsible for support of TensorFlow models is called as TensorFlow Frontend (TF FE). TF FE converts a model represented in TensorFlow opset to a model in OpenVINO opset.

In order to infer TensorFlow models with Atan2 operation by OpenVINO, TF FE needs to be extended with this operation support.

What needs to be done?

For Atan2 operation support, you need to add the corresponding loader into TF FE op directory and to register it into the dictionary of Loaders. One loader is responsible for conversion (or decomposition) of one type of TensorFlow operation.

Here is an example of loader implementation for TensorFlow Einsum operation:

OutputVector translate_einsum_op(const NodeContext& node) { 
     auto op_type = node.get_op_type(); 
     TENSORFLOW_OP_VALIDATION(node, op_type == "Einsum", "Internal error: incorrect usage of translate_einsum_op."); 
     auto equation = node.get_attribute<std::string>("equation"); 
  
     OutputVector inputs; 
     for (size_t input_ind = 0; input_ind < node.get_input_size(); ++input_ind) { 
         inputs.push_back(node.get_input(input_ind)); 
     } 
  
     auto einsum = make_shared<Einsum>(inputs, equation); 
     set_node_name(node.get_name(), einsum); 
     return {einsum}; 
 } 

In this example, translate_einsum_op converts TF Einsum into OV Einsum. NodeContext object passed into the loader packs all info about inputs and attributes of Einsum operation. The loader retrieves an attribute of the equation by using the NodeContext::get_attribute() method, prepares input vector, creates Einsum operation from OV opset and returns a vector of outputs.

Responsibility of a loader is to parse operation attributes, prepare inputs and express TF operation via OV operations sub-graph. Example for Einsum demonstrates the resulted sub-graph with one operation. In PR #19007 you can see operation decomposition into multiple node sub-graph.

Hint

The output range values for Atan2 is [-pi; +pi] different from Atan. You need to pay attention to signs of operands x and y. For more information, read the details in wiki.

Example Pull Requests

• [TF FE][TF Hub] Support DivNoNan and EnsureShape operations #19007

Resources

• What is OpenVINO?
• OpenVINO TensorFlow Frontend
• Contribution guide
• Blog post on contributing to OpenVINO
• User documentation

Contact points

@openvinotoolkit/openvino-tf-frontend-maintainers

Ticket

TBD

[rghvsh]

Hey! @rkazants I'd like to work on this project can i get started?

[rkazants]

Hey! @rkazants I'd like to work on this project can i get started?

Sure, very welcome to contribute. The task is yours.

Best regards,
Roman

[rkazants]

Hi @rghvsh, is there progress with this task? Please let me know if you need some help.

Thanks,
Roman

[rghvsh]

Hey! @rkazants I've started working on the issue and have made some progress.

Just had a small question. Wanted to make sure that in src/frontends/tensorflow/src/op_table.cpp Atan2 needs to be added under binary operations with opset 8.

Thank You,
-- @rghvsh

[rkazants]

tends/tensorflow/src/op_table.cpp Atan2 ne

Hi @rghvsh, please pay attention that it is Not sufficient to add:

"Atan2", CreatorFunction(translate_binary_op<opset8::Atan>)

That is because opset8::Atan and Atan2 have different inputs number. opset8::Atan is unary operation. You can see description of operations here: https://docs.openvino.ai/2023.1/openvino_docs_ops_opset12.html. Also, TF Atan2 value depends on signs of operands x and y. So you need to develop a decomposition of Atan2 using available operations from OpenVINO opset. Please use already developed loaders for other operations as examples.

Best regards,
Roman

[rghvsh]

Got it!

[rghvsh]

Hey! @rkazants Is this code correct for this issue wanted to check with and get your feed back before a PR. If there are any mistakes I'll fix them ASAP.

Thank You
-- @rghvsh

#include "common_op_table.hpp"
#include "src/core/reference/include/openvino/reference/atan2.hpp"

using namespace std;
using namespace ov::opset10;

namespace ov {
namespace frontend {
namespace tensorflow {
namespace op {
OutputVector translate_atan2_op(const NodeContext& node) {

default_op_checks(node, 2, {"atan2"});
auto y = node.get_input(0);
auto x = node.get_input(1);
int input_size = sizeof(x)/sizeof(x[0]);

// Creating variable of the same size to store value
auto pout = create_same_type_const_scalar<float32>(x, 0);

// Value of atan2.
auto atan2 = make_shared<atan2>(y, x, pout, input_size);

set_node_name(node.get_name(), atan2);
return {atan2};

}
} // namespace op
} // namespace tensorflow
} // namespace frontend
} // namespace ov

[rkazants]

Hey! @rkazants Is this code correct for this issue wanted to check with and get your feed back before a PR. If there are any mistakes I'll fix them ASAP.

Thank You -- @rghvsh

#include "common_op_table.hpp" #include "src/core/reference/include/openvino/reference/atan2.hpp"

using namespace std; using namespace ov::opset10;

namespace ov { namespace frontend { namespace tensorflow { namespace op { OutputVector translate_atan2_op(const NodeContext& node) {

default_op_checks(node, 2, {"atan2"});
auto y = node.get_input(0);
auto x = node.get_input(1);
int input_size = sizeof(x)/sizeof(x[0]);

// Creating variable of the same size to store value
auto pout = create_same_type_const_scalar<float32>(x, 0);

// Value of atan2.
auto atan2 = make_shared<atan2>(y, x, pout, input_size);

set_node_name(node.get_name(), atan2);
return {atan2};

} } // namespace op } // namespace tensorflow } // namespace frontend } // namespace ov

Hi @rghvsh,

Not quite correct:).

1. Please start from building OpenVINO because the code is not buildable. There is no atan2 operation in OV, also there is no sizeof method for ov::Outputov::Node. The instructions for build you can find here: https://github.com/openvinotoolkit/openvino/blob/master/docs/dev/build.md.
2. Build OV sub-graph that will compute Atan2 by the formula from wiki:
   

Examples to build sub-graph you can find in other loaders. Example: https://github.com/openvinotoolkit/openvino/blob/master/src/frontends/tensorflow_common/src/op/xdivy.cpp.

Best regards,
Roman

[rghvsh]

I see I'll start doing this

[rghvsh]

Hey! @rkazants Update:

1. I've built OpenVINO
2. Started building the subgraph

had a question about the input which will be a NodeClass type object what will be stored in the class can you point me towards the documentation tried finding but got none

Thank You

[rkazants]

Hey! @rkazants Update:

1. I've built OpenVINO
2. Started building the subgraph

had a question about the input which will be a NodeClass type object what will be stored in the class can you point me towards the documentation tried finding but got none

Thank You

Hi @rghvsh,

Great, keep this way. I think you mean ov::Node class that is a parent class for different operation classes from the opset. Here is a developer documentation for classes you need: https://docs.openvino.ai/2023.1/groupov_model_cpp_api.html

This documentation might be helpful but see already implemented loaders as examples too:)

Best regards,
Roman

[rkazants]

Hi @rghvsh.

Do you have progress working on this task. Please let me know if you need help.

Best regards,
Roman

[rghvsh]

Hey @rkazants I've made progress and am working in it.

Please let me know if you need help.

Sure!

Thank You
Raghav

[rghvsh]

default_op_checks(node, 2, {"atan2"});
auto y = node.get_input(0);
auto x = node.get_input(1);

Hey @rkazants this stament unpacks the the node object and gets x and y which are of Tensor datatype how to iterate over these(x and y).

[rkazants]

default_op_checks(node, 2, {"atan2"}); auto y = node.get_input(0); auto x = node.get_input(1);

Hey @rkazants this stament unpacks the the node object and gets x and y which are of Tensor datatype how to iterate over these(x and y).

Hi @rghvsh,

Not sure how you would like to iterate. You just need to build a graph on down of x and y to express atan2. No need to iterate because you have only two inputs.
Just give you an idea:

auto div_y_x = make_shared<Divide>(y, x);
auto atan = make_shared<Atan>(div_y_x);
auto atan_plus_pi = ...
auto atan_minus_pi = ...
auto zero_const = create_same_type_const_scalar<int32_t>(x, 0);
auto is_x_positive = make_shared<Greater>(x, const_zero);
...

Use Select operation from our opset to represent:

Best regards,
Roman

[rghvsh]

Hey @rkazants UPDATE :
I've understood how this works and have written the python test just had some question

In the second case we have two conditions how will those be implemented do we have to use somethging like this

 auto result = make_shared<Select>(is_x_negitive  && is_y_positive, atan_plus_pi ,  atan); 

and

is this right way to make a pi constant

auto pi_const = create_same_type_const_scalar<float>(x, pi);

Sorry for asking too many questions its my first time contributing to a large codebase.

Thank You
Raghav

[rkazants]

Hi @rghvsh,

Sorry for the delay. Here is an pseudo-code how to represent these conditions using our opset:

// handle the first condition
auto arctan_y_x = ...
auto result = arctan_y_x;

// handle the second condition
auto const_pi = create_same_type_const_scalar<double>(x, std::atan(1.0)*4);
auto x_negative = make_shared<Less>(x, const_zero);
auto y_non_negative = make_shared<GreaterEqual>(y, const_zero);
auto cond1 = make_shared<LogicalAnd>(x_negative, y_non_negative);
auto arctan_y_x_plus_pi = make_shared<Add>(arctan_y_x, const_pi)
auto result = make_shared<Select>(cond1, arctan_y_x_plus_pi, result);

// handle the third consition
auto y_negative = make_shared<Less>(y, const_zero);
auto cond2 = make_shared<LogicalAnd>(x_negative, y_negative);
auto arctan_y_x_minus_pi = make_shared<Add>(arctan_y_x, const_pi)
auto result = make_shared<Select>(cond2, arctan_y_x_minus_pi, result);

//...

For undefined behaviour, we can do nothing. Hope this helps.

Best regards,
Roman

[rghvsh]

Hey! @rkazants I just opened a PR #21076 if there are any changes that need to be implemented please let me know.

Thank You
Raghav