Toby Roseman

You will need to manually edit the spec (i.e. the protobuf) for your model. Here is what I would try. First, get the `axis` value for your `SoftmaxND` layer. Then...

You could potentially use the NeuralNetworkBuilder. I would probably just edit the protobuf directly.

I can reproduce this issue. This is a neural network backend only issue. If you use the new mlprogram backend, then the outputs match. To use the new backend, add...

Don't worry about those two warnings. Using your most recent code, things look good to me. `(y - cy).sum()` being `31.7326` may seem like a lot, but given that `y`...

With the original code, it does work with `convert_to="mlprogram"`. However you're right this new combination does not work with `convert_to="mlprogram"`. It is curious that multiplying by `49` causes things to...

This is a problem with one of our optimization passes. Passing `pass_pipeline=ct.PassPipeline.EMPTY` to `ct.convert` cause the predictions to match.

@Gianluigi121 - This is a lot of code. Please give us a simpler example to reproduce your issue. Also please include all the code we need in order to run...

`model_traced(torch.Tensor([[49406, 4160]]).long())` works so an input shape of `(1,2)` should be valid. The following works: ```python # Convert traced model to CoreML text_input_shape = ct.Shape(shape=(1, ct.RangeDim(lower_bound=2, upper_bound=77, default=77))) model_coreml =...

The `out` variable from your code contains three tensors all of which are empty. [The PyTorch documentation](https://pytorch.org/vision/0.12/generated/torchvision.models.detection.fasterrcnn_mobilenet_v3_large_fpn.html) for that model uses a different input shape. Are you sure your PyTorch...

Right, as I said, all of the output tensors are empty. Also the fact that the first tensor is of shape `(0, 4)` seems wrong. Do you have a `example_image`...