Alexander Lutsenko

Hi! Starting with `2.16-rc0`, Tensorflow resolves to Keras 3. The new Keras is not fully backward-compatible. Nobuco is tested with `2.15` and will continue to rely on the old Keras...

Transitioning to Keras 3 turned out more complicated than was advertised (see https://github.com/keras-team/keras/issues/19314). I'll get back to it as soon as there is more clarity regarding the new Keras' design.

Hi! To trace a model, we need to perform inference on it. Do you have the inference script?

Hi! Could you run the conversion with `save_trace_html=True` and show me the `trace.html` file?

Hi! Here's one failure case: ```python class DynamicShape(nn.Module): def __init__(self): super().__init__() def forward(self, x): _, h, w = x.shape # DO NOT cast pytorch tensors to numpy / python primitives...

> Or perhaps a way to insert "fake nodes" in order to step into the graph more easily with the python debugger. You can try something like this: ```python @nobuco.traceable...

> Could it be possible to mock the int function, or the __int__ method in order to identify when a traced shape tensor is going to return to a native...

> Seem like YOLO-World really wasn't written with exporting the model in mind. Yep, the flexibility of Pytorch spoils people. > I am wondering if there would be an easier...

> Except by inserting code like print_shape everywhere of course. I can't see any other way. It's certainly possible to insert a bunch of `print_shape`s automatically, but I'm not sure...

Take a look at this, seems to be working reasonably well https://github.com/AlexanderLutsenko/nobuco/blob/aa4745e6abb1124d90f7d3ace6d282f923f08a40/nobuco/node_converters/grid_sampling.py#L38