人脸识别的iresnet，在graph模式下开了fuse_add_to_output，无法通过数值check

[marigoold]

复现代码

import tempfile
import oneflow as flow
from oneflow_onnx.oneflow2onnx.util import convert_to_onnx_and_check
from flowvision.models.face_recognition import iresnet50

model = iresnet50().to("cuda")

class ModelGraph(flow.nn.Graph):
    def __init__(self, model):
        super().__init__()
        self.config.allow_fuse_add_to_output(True)
        self.backbone = model

    def build(self, x):
        x = x.to("cuda")
        out = self.backbone(x)
        return out

model.eval()
model_graph = ModelGraph(model)
model_graph._compile(flow.randn(1, 3, 112, 112))


with tempfile.TemporaryDirectory() as tmpdirname:
    flow.save(model.state_dict(), tmpdirname)
    convert_to_onnx_and_check(
        model_graph, flow_weight_dir=tmpdirname, onnx_model_path="./", print_outlier=True)

报错

File ~/miniconda/lib/python3.9/site-packages/oneflow_onnx/oneflow2onnx/util.py:102, in compare_result(a, b, rtol, atol, print_outlier)
    100         if np.abs(a[i] - b[i]) > atol + rtol * np.abs(b[i]):
    101             print("a[{}]={}, b[{}]={}".format(i, a[i], i, b[i]))
--> 102 assert np.allclose(a, b, rtol=rtol, atol=atol)

AssertionError:

但是 self.config.allow_fuse_add_to_output(True) 这一行注释掉就可以成功转换。

[marigoold]

随机初始化的模型，设置成eval时候，bn的running stat可能不稳定导致参数爆炸，可以用预训练的模型权重倒入再测试