Android TOCO无法使用--allow_custom_ops转换不支持的操作

我正在尝试转换在pascal voc数据集（）上训练的mobilenet_v2模型

我想以1024x1024分辨率推断，因此我使用以下命令导出了模型

 python export_model.py \
--checkpoint_path model/model.ckpt-30000 \
--export_path ./frozen_inference_graph.pb \
--model_variant="mobilenet_v2" \
--num_classes=21 \
--crop_size=1024 \
--crop_size=1024 \
--inference_scales=1.0

然后我使用toco命令将模型转换为tflite

toco --graph_def_file=/home/abdullah/models-master/research   /frozen_casted.pb --input_format=TENSORFLOW_GRAPHDEF                                                         

--output_format=TFLITE  

--output_file=deeplabv3_mnv2_pascal_trainval.tflite   
--inference_type=FLOAT --inference_input_type=QUANTIZED_UINT8
--input_arrays=ImageTensor 
--output_arrays=SemanticPredictions
--input_shapes=1,1024,1024,3 --default_ranges_min=0 
--default_ranges_max=255 --allow_custom_ops  
--std_dev_values=128 --mean_values=1

然后将模型转换为tflite，无错误：

    2019-07-16 00:28:18.188009: I tensorflow/contrib/lite/toco/import_tensorflow.cc:1171] Converting unsupported operation: Equal
2019-07-16 00:28:18.188052: I tensorflow/contrib/lite/toco/import_tensorflow.cc:1171] Converting unsupported operation: Equal
2019-07-16 00:28:18.188067: I tensorflow/contrib/lite/toco/import_tensorflow.cc:1171] Converting unsupported operation: LogicalOr
2019-07-16 00:28:18.188093: I tensorflow/contrib/lite/toco/import_tensorflow.cc:1171] Converting unsupported operation: Unpack
2019-07-16 00:28:18.188134: I tensorflow/contrib/lite/toco/import_tensorflow.cc:1171] Converting unsupported operation: LogicalAnd
2019-07-16 00:28:18.192256: I tensorflow/contrib/lite/toco/import_tensorflow.cc:1171] Converting unsupported operation: ResizeNearestNeighbor
2019-07-16 00:28:18.203603: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] Before Removing unused ops: 813 operators, 1237 arrays (0 quantized)
2019-07-16 00:28:18.232795: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] After Removing unused ops pass 1: 801 operators, 1215 arrays (0 quantized)
2019-07-16 00:28:18.269909: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] Before general graph transformations: 801 operators, 1215 arrays (0 quantized)
2019-07-16 00:28:18.293777: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] After general graph transformations pass 1: 132 operators, 325 arrays (0 quantized)
2019-07-16 00:28:18.296414: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] Before dequantization graph transformations: 132 operators, 325 arrays (0 quantized)
2019-07-16 00:28:18.298905: I tensorflow/contrib/lite/toco/allocate_transient_arrays.cc:311] Total transient array allocated size: 28311552 bytes, theoretical optimal value: 25165824 bytes.
2019-07-16 00:28:18.299732: W tensorflow/contrib/lite/toco/tflite/operator.cc:661] Ignoring unsupported attribute type with key '_output_shapes'
2019-07-16 00:28:18.299743: W tensorflow/contrib/lite/toco/tflite/operator.cc:661] Ignoring unsupported attribute type with key 'T'
2019-07-16 00:28:18.299946: W tensorflow/contrib/lite/toco/tflite/operator.cc:661] Ignoring unsupported attribute type with key 'T'
2019-07-16 00:28:18.299956: W tensorflow/contrib/lite/toco/tflite/operator.cc:661] Ignoring unsupported attribute type with key '_output_shapes'

但是，当我在tflite上运行模型时，它给出了以下错误

java.lang.IllegalArgumentException:内部错误：无法创建解释器：找不到版本为1的名称“ExpandDims”的自定义op 注册失败

对于转换，我使用了tensorflowtensorflow=='1.6.0-rc1'

build.gradle中的tflite实现是实现“org.tensorflow:tensorflow lite:+”

---

我不知道我在哪里失败了，请帮助我。

--允许自定义操作

被添加到tf1.10中，请参见，但tf1.10非常旧，无法将模型转换为tflite“nodedf在Op y:DstT中提到attr'Truncate”；attr=SrcT:type；attr=DstT:type>；nodedf:Cast=Cast[DstT\u FLOAT，SrcT=DT\u UINT8，Truncate=false，_output\u shapes=[？，？，3]]（挤压）。（检查GraphDef解释二进制文件是否与GraphDef生成二进制文件一起处于最新状态。）我的意思是它在TF 1.10和更高版本中可用。我知道您使用的是TF 1.6，它比TF 1.6旧得多。我建议使用TF nightly将模型转换为TF-lite。ExpandDims现在作为TensorFlow lite的适当内置操作受到支持。我将更新您的TensorFlow版本（用于转换模型）至少到1.14.0。tf1.10中添加了“允许自定义操作”，请参见，但tf1.10非常旧，无法将模型转换为tflite“NodeDef在Op y:DstT中提到attr‘Truncate’；attr=SrcT:type；attr=DstT:type>；NodeDef:Cast=Cast[DstT=DT\u FLOAT，SrcT=DT\u UINT8，Truncate=false，[u output\u shapes=[？，？，3]]（挤压）。（检查GraphDef解释二进制文件是否与GraphDef生成二进制文件一起处于最新状态。）我的意思是它在TF 1.10和更高版本中可用。我知道您使用的是TF 1.6，它比TF 1.6旧得多。我建议您使用TF nightly将模型转换为TF-lite。ExpandDims现在作为TensorFlow lite的适当内置操作受到支持。我会将您的TensorFlow版本（用于转换模型）更新为至少1.14.0。