[converter] aten::{index_select, scalarImplicit}

[peterjc123]

Fixes #181

[LUOBO123LUOBO123]

Another way to achieve。

class ATenIndexSelectOperator(ATenSliceSchema):
def parse(self, node, attrs, args, graph_converter):
super().parse(node, attrs, args, graph_converter)

    self.run(node)
    input_tensor = self.find_or_create_input(0, graph_converter)
    dim, index = self.input_tensors[1:]

    start_index = int(index[0])

    starts = np.zeros(input_tensor.tensor.ndim, dtype='int32')
    starts[dim] = start_index
    start_tensor = self.create_attr_tensor(starts)

    sizes = np.zeros(input_tensor.tensor.ndim, dtype='int32') - 1
    sizes[dim] = len(index)
    size_tensor = self.create_attr_tensor(sizes)

    inputs = [input_tensor, start_tensor, size_tensor]
    outputs = self.to_tfl_tensors(self.output_names, self.output_tensors)
    graph_converter.add_operator(tfl.SliceOperator(inputs, outputs)

[peterjc123]

Another way to achieve。

class ATenIndexSelectOperator(ATenSliceSchema): def parse(self, node, attrs, args, graph_converter): super().parse(node, attrs, args, graph_converter)

    self.run(node)
    input_tensor = self.find_or_create_input(0, graph_converter)
    dim, index = self.input_tensors[1:]

    start_index = int(index[0])

    starts = np.zeros(input_tensor.tensor.ndim, dtype='int32')
    starts[dim] = start_index
    start_tensor = self.create_attr_tensor(starts)

    sizes = np.zeros(input_tensor.tensor.ndim, dtype='int32') - 1
    sizes[dim] = len(index)
    size_tensor = self.create_attr_tensor(sizes)

    inputs = [input_tensor, start_tensor, size_tensor]
    outputs = self.to_tfl_tensors(self.output_names, self.output_tensors)
    graph_converter.add_operator(tfl.SliceOperator(inputs, outputs)

Thanks for sharing. This inspires me to think of some graph optimization work that can be done. I opened a new issue for that. #184