是否有一些ONNX规范里面不存在的算子，或者有一些非常规网络里面的算子，比如 Conv 的 Kernel 大小超过128

能否输入更为详细的错误信息，好做相应调整

import os

import onnx

m = onnx.load('/Users/rqg/test/fft512_128_MTFAANetV2_mini_v0.1/aec.onnx')

init_names = {}

for i in m.graph.initializer:
    init_names[i.name] = i.dims

value_info = {}

for i in m.graph.value_info:
    value_info[i.name] = [d.dim_value for d in i.type.tensor_type.shape.dim]

for i in m.graph.input:
    value_info[i.name] = [d.dim_value for d in i.type.tensor_type.shape.dim]

value_info.update(init_names)


def wait_for_input():
    input('Press Enter to continue...')


def get_node_inputs(n):
    nis = []
    nia = []
    for ni in n.input:
        if len(ni) > 0:
            if ni in init_names:
                nia.append(ni)
            else:
                nis.append(ni)
    return nis, nia


def cmp_node_attributes(n1, n2):
    if len(n1.attribute) != len(n2.attribute):
        return False

    na1 = {}
    na2 = {}

    for i in n1.attribute:
        na1[i.name] = (i.type, i.f, i.i, i.s, i.t, i.g, i.floats, i.ints, i.strings, i.tensors, i.graphs)
    for i in n2.attribute:
        na2[i.name] = (i.type, i.f, i.i, i.s, i.t, i.g, i.floats, i.ints, i.strings, i.tensors, i.graphs)

    for n in na1:
        if n not in na2:
            return False
        if na1[n] != na2[n]:
            return False

    return True


def same_op(n1, n2):
    if n1.op_type != n2.op_type:
        return False

    nis1, nia1 = get_node_inputs(n1)
    nis2, nia2 = get_node_inputs(n2)

    if len(nis1) != len(nis2) or len(nia1) != len(nia2):
        return False

    for i in range(len(nis1)):
        if value_info[nis1[i]] != value_info[nis2[i]]:
            return False

    for i in range(len(nia1)):
        if value_info[nia1[i]] != value_info[nia2[i]]:
            return False

    return cmp_node_attributes(n1, n2)


verify_node = []


def already_in_verify_node(n):
    for vn in verify_node:
        if same_op(n, vn):
            return True
    return False


for n in m.graph.node:
    if not already_in_verify_node(n):
        verify_node.append(n)

print(len(verify_node))

# tt_names = [
#     "Conv_631",
#     "BatchNormalization_1318",
#     "Relu_1319",
#     "Conv_1337",
# ]

for n in verify_node:
    # print(n.name)
    #
    # if n.name not in tt_names:
    #     continue
    node_inputs = []
    for ni in n.input:
        if ni not in init_names and len(ni) > 0:
            node_inputs.append(ni)

    onnx.utils.extract_model(
        '/Users/rqg/test/fft512_128_MTFAANetV2_mini_v0.1/aec.onnx',
        '/Users/rqg/test/fft512_128_MTFAANetV2_mini_v0.1/aec_tt.onnx',
        input_names=node_inputs,
        output_names=n.output,
    )
    os.system('mc cp /Users/rqg/test/fft512_128_MTFAANetV2_mini_v0.1/aec_tt.onnx my/xzgj/test/aec_tt.onnx')
    wait_for_input()

如何验证哪个OP 不支持的PYTHON脚本，自己写个 小程序，点下从http 上下载加载验证。

我是手动一个一个验证的 150多种op case

有些 conv相关的 case 会崩溃