feat: add sum onnx import

[JachymPutta]

Checklist

• Confirmed that run-checks all script has been executed.
• Made sure the book is up to date with changes in this PR.

Related Issues/PRs

Helping with #1714

Changes

Added the ONNX models for sum and sum_int, the node seems to by covered by add

Testing

Manual

cargo test -p burn-import --color=always -- --color=always
cargo test -p onnx-tests --color=always

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 86.28%. Comparing base (99e1ba4) to head (c146776).
Report is 5 commits behind head on main.

❗ Current head c146776 differs from pull request most recent head fadccda

Please upload reports for the commit fadccda to get more accurate results.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1846      +/-   ##
==========================================
+ Coverage   86.24%   86.28%   +0.03%     
==========================================
  Files         767      769       +2     
  Lines       88575    88833     +258     
==========================================
+ Hits        76394    76648     +254     
- Misses      12181    12185       +4     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[antimora]

Yeah I think Add and Sum are similar but not the same. The python script generates Add node types. Probably it's difficult or not possible to achieve with PyTorch itself. So you might have to resort to onnx python utility. Here is the code that can generate one:

import onnx
import onnx.helper
import onnx.checker
import numpy as np

# Create input tensors
input1 = onnx.helper.make_tensor_value_info('input1', onnx.TensorProto.FLOAT, [3])
input2 = onnx.helper.make_tensor_value_info('input2', onnx.TensorProto.FLOAT, [3])
input3 = onnx.helper.make_tensor_value_info('input3', onnx.TensorProto.FLOAT, [3])

# Create output tensor
output = onnx.helper.make_tensor_value_info('output', onnx.TensorProto.FLOAT, [3])

# Create the Sum node
sum_node = onnx.helper.make_node(
    'Sum',
    inputs=['input1', 'input2', 'input3'],
    outputs=['output']
)

# Create the graph (GraphProto)
graph_def = onnx.helper.make_graph(
    nodes=[sum_node],
    name='SumGraph',
    inputs=[input1, input2, input3],
    outputs=[output]
)

# Create the model (ModelProto)
model_def = onnx.helper.make_model(graph_def, producer_name='sum-model')
onnx.checker.check_model(model_def)

# Save the ONNX model
onnx.save(model_def, 'sum_model.onnx')

print("ONNX model 'sum_model.onnx' generated successfully.")

However, for the underlying operator you'll have to use Burn's add operator multiple times. This is very similar to concat operation, see: https://github.com/tracel-ai/burn/blob/main/crates/burn-import/src/burn/node/concat.rs#L36-L38

Basically, a generated code might look like input1 + input2 + input3

[JachymPutta]

Ahh, I see, let me fix that!
Thank you!

[antimora]

There is one little rename required but overall looks great!

Thank you for your contribution! You are on a roll!

Approving in advance.

[antimora]

test_codegen_concat => test_codegen_sum

[JachymPutta]

Oh, I missed that! That's what I get for the mindless copying:)

[JachymPutta]

Should be good!