ONNX2MPSX.py

import onnx
from onnx import numpy_helper
import numpy as np
import argparse


def _fp32_to_fp16_info(tensor_infos):
    return [
        onnx.helper.make_tensor_value_info(
            ti.name,
            10,
            [dim.dim_value or dim.dim_param for dim in ti.type.tensor_type.shape.dim]
        ) if ti.type.tensor_type.elem_type == 1 else ti for ti in tensor_infos
    ]


def _fp32_to_fp16(tensors):
    return [
        numpy_helper.from_array(
            numpy_helper.to_array(t).astype(np.float16),
            name=t.name
        ) if t.data_type == 1 else t for t in tensors
    ]


def find_tensors_for_transposition(nodes):
    swizzle_candidates = {}
    for node in nodes:
        if node.op_type == 'Conv':
            for attr in node.attribute:
                if attr.name == 'group' and attr.i != 1:
                    swizzle_candidates[node.input[1]] = attr.i
    return swizzle_candidates


def transpose_depthwise_conv_weights(tensors, names):
    return [
        numpy_helper.from_array(
            np.transpose(numpy_helper.to_array(t), (1, 0, 2, 3)),
            name=t.name
        ) if t.name in names and t.dims[0] == names[t.name] else t for t in tensors
    ]


def convert_onnx_to_mpsx(model, halfs):
    swizzle_candidates = find_tensors_for_transposition(model.graph.node)
    swizzled_initializer = transpose_depthwise_conv_weights(
        model.graph.initializer, swizzle_candidates)
    new_model = onnx.helper.make_model(
        onnx.helper.make_graph(
            nodes=model.graph.node,
            name=model.graph.name,
            inputs=_fp32_to_fp16_info(
                model.graph.input) if halfs else model.graph.input,
            outputs=_fp32_to_fp16_info(
                model.graph.output) if halfs else model.graph.output,
            initializer=_fp32_to_fp16(
                swizzled_initializer) if halfs else swizzled_initializer
        ),
        producer_name='MPSX'
    )
    return new_model


def main():
    parser = argparse.ArgumentParser(
        description='Convert ONNX model to MPSX format')
    parser.add_argument('--half', required=False,
                        help='Use FP16 weights', action='store_true')
    parser.add_argument('--input', required=True, help='Path to ONNX model')
    parser.add_argument('--output', required=True, help='Path to MPSX model')
    args = parser.parse_args()

    onnx_model = onnx.load(args.input)

    if onnx_model.producer_name == 'MPSX':
        return

    onnx.helper.strip_doc_string(onnx_model)
    mpsx_model = convert_onnx_to_mpsx(onnx_model, args.half)
    onnx.save(mpsx_model, args.output)

    print('Done!')


if __name__ == "__main__":
    main()