Imprecise conversion for custom log_softmax converter about nobuco

[AlexanderLutsenko]

Hi! First, here's a proper converter:

@converter(F.log_softmax, torch.log_softmax, torch.Tensor.log_softmax, channel_ordering_strategy=ChannelOrderingStrategy.MINIMUM_TRANSPOSITIONS)
def converter_log_softmax(input: Tensor, dim, *, dtype: Optional[_dtype]=None):
    num_dims = input.dim()

    def func(input, dim, *, dtype=None):
        if get_channel_order(input) == ChannelOrder.TENSORFLOW:
            dim = dim_pytorch2keras(dim, num_dims)
        return tf.nn.log_softmax(input, axis=dim)
    return func

Turns out, the converter is already there, and I registered it for torch.log_softmax, torch.Tensor.log_softmax, but forgot about F.log_softmax.

Your implementation does not work correctly because due to ChannelOrderingStrategy.MINIMUM_TRANSPOSITIONS it can receive input tensor in channel-last layout, and in this case, dim has to be adapted accordingly.

The dim < 0 check is redundant, as it's done inside dim_pytorch2keras. Output type cast can also be omitted, because if needed, its' performed automatically inside the wrapper layer.

[coxep]

Awesome, thanks! Got lightglue built, and I think it is working... want any of these chatgpt-sourced converters?
coxep@5afcdf0

[AlexanderLutsenko]

Awesome, thanks! Got lightglue built, and I think it is working... want any of these chatgpt-sourced converters? coxep@5afcdf0

Yep, I'll take it.

A word of warning: there's a good chance the model was not converted properly. See, LightGlue is quite dynamic, the number of iterations depends on the difficulty of the task, and Nobuco cannot automatically capture control flows. It might still work as-is, but you'd be missing out on performance. Speaking of performance. Tensorflow lags behind Pytorch considerably when it comes to transformers. Pytorch already integrated fast attention operations (notably, F.scaled_dot_product_attention), and there's no such thing in TFLite.

[coxep]

Thanks for the heads-up. I'm guessing that the conversion was suboptimal. The h5 file is ~80mb, but I'm at least getting the same correspondence as the pytorch model.

I did have to make some changes (disabled pruning, fixed the number of iterations / disabled early stopping, etc)