[Good First Issue][NNCF][TorchFX]: Test model transformer

[daniil-lyakhov]

Greetings🐱! As a part of https://github.com/openvinotoolkit/nncf/issues/2766 TorchFX PTQ backend support, we are gladly presenting to you following issue

Context

The task is to cover FXModelTransformer by simple unit tests as it done for other backends: https://github.com/openvinotoolkit/nncf/blob/develop/tests/onnx/test_model_transformer.py

What needs to be done?

Unit tests in file tests/torch/fx/test_model_transformer.py :

• test_leaf_module_insertion_transformation
• test_bias_update_transformation
• test_qdq_insertion_transformation

Example Pull Requests

No response

Resources

• Contribution guide - start here!
• Intel DevHub Discord channel - engage in discussions, ask questions and talk to OpenVINO developers
• How to link your Pull Request to an issue

Contact points

@daniil-lyakhov

Ticket

141640

[awayzjj]

.take

[github-actions[bot]]

Thank you for looking into this issue! Please let us know if you have any questions or require any help.

[parthrastogicoder]

.take

[github-actions[bot]]

Thanks for being interested in this issue. It looks like this ticket is already assigned to a contributor. Please communicate with the assigned contributor to confirm the status of the issue.

[awayzjj]

@daniil-lyakhov Hi, Could you please push your branch before merging it into the develop branch? This will allow me to develop based on your branch.

Additionally, I noticed the file tests/torch/fx/test_sanity.py. Should I implement the unit tests in tests/torch/fx/test_model_transformer.py instead of tests/torch_fx/test_model_transformer.py?

Thank you!

[daniil-lyakhov]

Hi @awayzjj,

Thank you for your contribution! PR is on review right now and should be merged soon, I'll keep you updated. Yes, please use tests/torch/fx directory, I forgot to update the issue.

Thanks!

[daniil-lyakhov]

Hi @awayzjj, the base PR #2764 was merged, please rebase your cahnges

[awayzjj]

@daniil-lyakhov Hi, I attempted to implement the test_leaf_module_insertion_transformation as follows, but encountered two problems:

import torch
import torch.nn.functional as F
from torch import nn
from torch._export import capture_pre_autograd_graph

from nncf.common.graph.transformations.commands import TargetType
from nncf.common.graph.transformations.layout import TransformationLayout

def test_leaf_module_insertion_transformation():

    class InsertionPointTestModel(nn.Module):
        def __init__(self):
            super().__init__()
            self.conv1 = nn.Conv2d(1, 1, 1, 1)
            self.linear_wts = nn.Parameter(torch.FloatTensor(size=(100, 100)))
            self.conv2 = nn.Conv2d(1, 1, 1, 1)
            self.relu = nn.ReLU()

        def forward(self, input_):
            x = self.conv1(input_)
            x = x.flatten()
            x = nn.functional.linear(x, self.linear_wts)
            x = x.reshape((1, 1, 10, 10))
            x = self.conv2(x)
            x = self.relu(x)
            return x

    model = InsertionPointTestModel() 
    with torch.no_grad():
        ex_input = torch.ones([1, 1, 10, 10])
        model.eval()
        exported_model = capture_pre_autograd_graph(model, args=(ex_input,))
    print(exported_model.print_readable())

    from nncf.experimental.torch.fx.model_transformer import FXModelTransformer
    from nncf.torch.graph.transformations.commands import PTTargetPoint
    from nncf.experimental.torch.fx.transformations import leaf_module_insertion_transformation_builder
    from nncf.experimental.torch.fx.commands import FXApplyTransformationCommand
    model_transformer = FXModelTransformer(exported_model)

    conv1_node_name = "InsertionPointTestModel/NNCFConv2d[conv1]/conv2d_0"
    target_point = PTTargetPoint(
        target_type=TargetType.OPERATION_WITH_WEIGHTS, target_node_name=conv1_node_name, input_port_id=1
    )
    transformation = leaf_module_insertion_transformation_builder(
        exported_model, [target_point]
    )
    command = FXApplyTransformationCommand(
        transformation
    )
    transformation_layout = TransformationLayout()
    transformation_layout.register(command)
    model_transformer.transform(transformation_layout)

1. The tests fail with the following exception:

   

2. I have to place the following imports after exported_model = capture_pre_autograd_graph(model, args=(ex_input,))

   from nncf.experimental.torch.fx.model_transformer import FXModelTransformer
   from nncf.torch.graph.transformations.commands import PTTargetPoint
   from nncf.experimental.torch.fx.transformations import leaf_module_insertion_transformation_builder
   from nncf.experimental.torch.fx.commands import FXApplyTransformationCommand

   or it raises an error:

   

Could you give me some suggestions? Thank you very much!

[rk119]

@awayzjj you defined the variable conv1_node_name incorrectly. In this case, for example if you define it with conv1_node_name = "conv2d" instead as this is a valid node name, it should work. The node names in the Graph are different for Torch FX backend.

[awayzjj]

@rk119 Thank you so much! Your suggestions does work!