1****系统环境
硬件环境(Ascend/GPU/CPU): Ascend/GPU/CPU
MindSpore版本: mindspore=2.2.13
执行模式（PyNative/ Graph）:不限
Python版本: Python=3.9
操作系统平台: 不限
2****报错信息
2.1****问题描述
使用Mindspore构建只包含单个Dropout算子的模型，格式分别为ONNX和MINDIR，无法正确使用converter_lite工具转换为.ms格式。
import numpy as np
import mindspore
from mindspore import Parameter, ops, Tensor, nn
class Dropout(nn.Cell):
def __init__(self):
super(Dropout, self).__init__()
self.Dropout = ops.Dropout(keep_prob=0.8)
def construct(self, x):
return self.Dropout(x)
dropout = Dropout()
dummy_input = Tensor(np.random.rand(2000, 1600), mindspore.float32)
mindspore.export(dropout, dummy_input, file_name="demo10", file_format="MINDIR")复制
执行下面命令
./converter_lite --fmk=MINDIR --trainModel=true --modelFile=demo10.mindir --outputFile=demo10复制
有以下报错（trainModel参数是否存在都会导致同样的报错）：
2.1****报错信息
[ERROR] LITE(32309,ffff9357b010,converter_lite):2024-08-13-15:04:46.393.679 [mindspore/lite/tools/common/graph_util.cc:133] GetShapeVectorAndIdxFromCNode] Abstract is not abstract tensor. Default/Dropout-op6
[ERROR] LITE(32309,ffff9357b010,converter_lite):2024-08-13-15:04:46.393.727 [mindspore/lite/tools/common/graph_util.cc:236] TraceOutput] Get node shape failed.
[ERROR] LITE(32309,ffff9357b010,converter_lite):2024-08-13-15:04:46.393.745 [mindspore/lite/tools/common/graph_util.cc:661] GetFuncGraphOutputsInfo] Trace output failed.
[ERROR] CORE(32309,ffff9357b010,converter_lite):2024-08-13-15:04:46.393.784 [mindspore/core/utils/log_adapter.cc:433] operator^] Runtime error for null exception handler.
[ERROR] LITE(32309,ffff9357b010,converter_lite):2024-08-13-15:04:46.394.005 [mindspore/lite/tools/converter/converter.cc:1324] RunConverter] Exception occurred: Failure info [Get outputs info of funcgraph failed].
----------------------------------------------------
- Framework Unexpected Exception Raised:
----------------------------------------------------
This exception is caused by framework's unexpected error. Please create an issue at https://gitee.com/mindspore/mindspore/issues to get help.
----------------------------------------------------
- C++ Call Stack: (For framework developers)
----------------------------------------------------
mindspore/lite/tools/common/func_graph_utils.cc:83 GetFuncGraphOutputNames
[ERROR] LITE(32309,ffff9357b010,converter_lite):2024-08-13-15:04:46.394.038 [mindspore/lite/tools/converter/converter.cc:1328] RunConverter] Convert model failed
[ERROR] LITE(32309,ffff9357b010,converter_lite):2024-08-13-15:04:46.394.061 [mindspore/lite/tools/converter/cxx_api/converter.cc:374] Convert] Convert model failed, ret=Common error code.
ERROR [mindspore/lite/tools/converter/converter_lite/main.cc:104] main] Convert failed. Ret: Common error code.
Convert failed. Ret: Common error code.复制
3****根因分析
看报错信息是convert工具在处理output的时候遇到了不支持的类型
GetShapeVectorAndIdxFromCNode] Abstract is not abstract tensor
然后查看API
应该是mask的输出比较特殊导致convert处理失败。
4****解决方案
1.算子使用ops.dropout, 因为该算子只有一个输出
import numpy as np
import mindspore
from mindspore import Parameter, ops, Tensor, nn
class Dropout(nn.Cell):
def __init__(self):
super(Dropout, self).__init__()
def construct(self, x):
return ops.dropout(x, p=0.2)
dropout = Dropout()
dummy_input = Tensor(np.random.rand(2000, 1600), mindspore.float32)
mindspore.export(dropout, dummy_input, file_name="demo10", file_format="MINDIR")
2.不将Dropout的输出mask作为网络输出
import numpy as np
import mindspore
from mindspore import Parameter, ops, Tensor, nn
mindspore.set_context(device_target="CPU")
class Dropout(nn.Cell):
def __init__(self):
super(Dropout, self).__init__()
self.Dropout = ops.Dropout(keep_prob=0.8)
def construct(self, x):
x, _ = self.Dropout(x)
return x
dropout = Dropout()
dummy_input = Tensor(np.random.rand(2000, 1600), mindspore.float32)
mindspore.export(dropout, dummy_input, file_name="demo11", file_format="MINDIR")