Mxnet转Onnx 踩坑记录

news2024/9/29 19:22:29

0. 前言

使用将MXNET模型转换为ONNX的过程中有很多算子不兼容，在此对那些不兼容的算子替换。在此之前需要安装mxnet分支v1.x版本作为mx2onnx的工具，git地址如下：

mxnet/python/mxnet/onnx at v1.x · apache/mxnet · GitHub

同时还参考了如下的两个知乎链接：

https://zhuanlan.zhihu.com/p/166267806

https://zhuanlan.zhihu.com/p/165294876

1. UpSample

1.1 采用Resize实现

mxnet/contrib/onnx/mx2onnx/_op_translations.py
def create_helper_tensor_node(input_vals, output_name, kwargs):

    """create extra tensor node from numpy values"""

    data_type = onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[input_vals.dtype]



    tensor_node = onnx.helper.make_tensor_value_info(

        name=output_name,

        elem_type=data_type,

        shape=input_vals.shape

    )

    kwargs["initializer"].append(

        onnx.helper.make_tensor(

            name=output_name,

            data_type=data_type,

            dims=input_vals.shape,

            vals=input_vals.flatten().tolist(),

            raw=False,

        )

    )



    return tensor_node



@mx_op.register("UpSampling")

def convert_upsample(node, **kwargs):

    """Map MXNet's UpSampling operator attributes to onnx's Upsample operator

    and return the created node.

    """

   

    name, input_nodes, attrs = get_inputs(node, kwargs)

   

    sample_type = attrs.get('sample_type', 'nearest')

    sample_type = 'linear' if sample_type == 'bilinear' else sample_type

    scale = convert_string_to_list(attrs.get('scale'))

    scaleh = scalew = float(scale[0])

    if len(scale) > 1:

        scaleh = float(scale[0])

        scalew = float(scale[1])

    scale = np.array([1.0, 1.0, scaleh, scalew], dtype=np.float32)

    roi = np.array([], dtype=np.float32)

    node_roi=create_helper_tensor_node(roi, name+'roi', kwargs)

    node_sca=create_helper_tensor_node(scale, name+'scale', kwargs)

   

    node = onnx.helper.make_node(

        'Resize',

        inputs=[input_nodes[0], name+'roi', name+'scale'],

        outputs=[name],

        coordinate_transformation_mode='asymmetric',

        mode=sample_type,

        nearest_mode='floor',

        name=name

    )

    return [node_roi, node_sca, node]

1.2 采用ConvTranspose实现

@mx_op.register("UpSampling")

def convert_upsample(node, **kwargs):

    """Map MXNet's UpSampling operator attributes to onnx's Upsample operator

    and return the created node.

    """

    import math

    name, inputs, attrs = get_inputs(node, kwargs)

   

    #==NearestNeighbor ==

    channels=64  #此处需要手动修改！！！

    scale=int(attrs.get('scale'))

    pad=math.floor((scale - 1)/2.0)

    weight = np.ones((channels,1,scale,scale), dtype=np.float32)

    weight_node=create_helper_tensor_node(weight, name+'_weight', kwargs)

    pad_dims = [pad, pad]

    pad_dims = pad_dims + pad_dims

    #print(pad_dims)

    deconv_node = onnx.helper.make_node(

        "ConvTranspose",

        inputs=[inputs[0], name+'_weight'],

        outputs=[name],

        auto_pad="VALID",

        strides=[scale, scale],

        kernel_shape=[scale,scale],

        pads=pad_dims,

        group=channels,

        name=name)

   

return [deconv_node]

2. 使用Slice代替Crop.

mxnet/contrib/onnx/mx2onnx/_op_translations.py

def create_helper_shape_node(input_node, node_name):

    """create extra transpose node for dot operator"""

    trans_node = onnx.helper.make_node(

        'Shape',

        inputs=[input_node],

        outputs=[node_name],

        name=node_name

    )

   

    return trans_node



@mx_op.register("Crop")

def convert_crop(node, **kwargs):

    """Map MXNet's crop operator attributes to onnx's Crop operator

    and return the created node.

    """

   

    name, inputs, attrs = get_inputs(node, kwargs)

   

    start=np.array([0, 0, 0, 0], dtype=np.int) #index是int类型

    start_node=create_helper_tensor_node(start, name+'__starts', kwargs)

    shape_node = create_helper_shape_node(inputs[1], inputs[1]+'__shape')

   

    crop_node = onnx.helper.make_node(

        "Slice",

        inputs=[inputs[0], name+'__starts', inputs[1]+'__shape'], #data、start、end

        outputs=[name],

        name=name

    )

   

    logging.warning(

        "Using an experimental ONNX operator: Crop. " \

        "Its definition can change.")



    return [start_node, shape_node, crop_node]

3. ONNX softmax维度转换问题

onnx的实现的softmax在处理多维输入(NCHW)存在问题。

@mx_op.register("softmax")

def convert_softmax(node, **kwargs):

    """Map MXNet's softmax operator attributes to onnx's Softmax operator

    and return the created node.

    """

    name, input_nodes, attrs = get_inputs(node, kwargs)

    axis = int(attrs.get("axis", -1))

   

    c_softmax_node = []

    axis=-1

   

    transpose_node1 = onnx.helper.make_node(

        "Transpose",

        inputs=input_nodes,

        perm=(0,2,3,1), #NCHW--NHWC--(NHW,C)

        name=name+'_tr1',

        outputs=[name+'_tr1']

    )

   

    softmax_node = onnx.helper.make_node(

        "Softmax",

        inputs=[name+'_tr1'],

        axis=axis,

        name=name+'',

        outputs=[name+'']

    )

   

    transpose_node2 = onnx.helper.make_node(

        "Transpose",

        inputs=[name+''],

        perm=(0,3,1,2), #NHWC--NCHW

        name=name+'_tr2',

        outputs=[name+'_tr2']

    )

   

    c_softmax_node.append(transpose_node1)

    c_softmax_node.append(softmax_node)

    c_softmax_node.append(transpose_node2)

   

    return c_softmax_node