aten::repeat_interleave is considered an unsupported Tensor and causing an error when trying to convert to TFLite about tinyneuralnetwork

[peterjc123]

It is actually a quite complicated operator to translate. It has the following usages.
For example, we have x = torch.tensor([[1,2],[3,4]]) and the function signature torch.repeat_interleave(input, repeats=None, dim=None)

1. torch.repeat_interleave(x, 2)
2. torch.repeat_interleave(x, torch.tensor([1,2,3,4]))
3. torch.repeat_interleave(x, torch.tensor([1,2]), dim=0)
4. torch.repeat_interleave(x.flatten())

They can be split into two very different categories.

1. if repeats is int and dim is None, then it can be translated as x.reshape(-1,1).repeat(1,repeats).view(-1)
2. For other cases, we need to transform the repeats to indices to make the operation like a indexing operation.

len_repeats = len(repeats) # For `repeats` as an int, it should be the size of the dim
raw_indices = torch.arange(len_repeats)
indices = torch.repeat_interleave(raw_indices, repeats) # This can be calculated by concatenating the tiled raw_indices. e.g. [0] * 1 + [1] * 2 + [2] * 3

[peterjc123]

Well, I think I will not implement the logic for dynamic repeats tensor because it may lead to dynamic-shaped tensors.