CCO model requirements:

To be able to use a torch model as input for CCO it must satisfy a traceability requirement.

Traceability

The input model must be traceable, meaning that it does not contain any dynamic python operations like if statements. To check if a model is indeed traceable, use the function torch.fx.symbolic_trace(). For more information, see the PyTorch documentation.

Similarly, if you are using Hugging Face Transformers library, you can use transformers.utils.fx.symbolic_trace() to check the traceability of your model. For more information, see the Transformers documentation.

Example

In this example we will refactor a simple non-traceable model into a traceable model:

import torch
from torch import nn


class NonTraceableModel(nn.Module):
    def __init__(self):
        super(NonTraceableModel, self).__init__()
        self.linear = nn.Linear(10, 10)

    def forward(self, x):
        # use "if" statement which will NOT be traceable
        if x.sum() > 0:
            return self.linear(x)
        else:
            return torch.zeros(10)


# will crash and therefore cannot be used for CCO
non_traceable_model = NonTraceableModel()
torch.fx.symbolic_trace(non_traceable_model)


class TraceableModel(nn.Module):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.linear = nn.Linear(10, 10)

    def forward(self, x):
        # use torch operations instead of "if" statement
        condition = x.sum() > 0
        linear_output = self.linear(x)
        linear_output = linear_output * condition
        return linear_output


# will run successfully and therefore can be used for CCO
traceable_model = TraceableModel()
torch.fx.symbolic_trace(traceable_model)