如何找到TensorRT不支持的操作

时间:2018-07-20 13:14:38

标签: tensorflow tensorrt

当我将我的张量流模型(保存为.pb文件)转换为uff文件时,错误日志如下:

Using output node final/lanenet_loss/instance_seg
Using output node final/lanenet_loss/binary_seg
Converting to UFF graph
Warning: No conversion function registered for layer: Slice yet.
Converting as custom op Slice final/lanenet_loss/Slice
name: "final/lanenet_loss/Slice"
op: "Slice"
input: "final/lanenet_loss/Shape_1"
input: "final/lanenet_loss/Slice/begin"
input: "final/lanenet_loss/Slice/size"
attr {
  key: "Index"
  value {
    type: DT_INT32
  }
}
attr {
  key: "T"
  value {
    type: DT_INT32
  }
}

Traceback (most recent call last):
  File "tfpb_to_uff.py", line 16, in <module>
    uff_model = uff.from_tensorflow(graphdef=output_graph_def, output_filename=output_path, output_nodes=["final/lanenet_loss/instance_seg", "final/lanenet_loss/binary_seg"], text=True)
  File "/home/dream/.local/lib/python3.5/site-packages/uff/converters/tensorflow/conversion_helpers.py", line 75, in from_tensorflow
    name="main")
  File "/home/dream/.local/lib/python3.5/site-packages/uff/converters/tensorflow/converter.py", line 64, in convert_tf2uff_graph
    uff_graph, input_replacements)
  File "/home/dream/.local/lib/python3.5/site-packages/uff/converters/tensorflow/converter.py", line 51, in convert_tf2uff_node
    op, name, tf_node, inputs, uff_graph, tf_nodes=tf_nodes)
  File "/home/dream/.local/lib/python3.5/site-packages/uff/converters/tensorflow/converter.py", line 28, in convert_layer
    fields = cls.parse_tf_attrs(tf_node.attr)
  File "/home/dream/.local/lib/python3.5/site-packages/uff/converters/tensorflow/converter.py", line 177, in parse_tf_attrs
    for key, val in attrs.items()}
  File "/home/dream/.local/lib/python3.5/site-packages/uff/converters/tensorflow/converter.py", line 177, in <dictcomp>
    for key, val in attrs.items()}
  File "/home/dream/.local/lib/python3.5/site-packages/uff/converters/tensorflow/converter.py", line 172, in parse_tf_attr_value
    return cls.convert_tf2uff_field(code, val)
  File "/home/dream/.local/lib/python3.5/site-packages/uff/converters/tensorflow/converter.py", line 146, in convert_tf2uff_field
    return TensorFlowToUFFConverter.convert_tf2numpy_dtype(val)
  File "/home/dream/.local/lib/python3.5/site-packages/uff/converters/tensorflow/converter.py", line 74, in convert_tf2numpy_dtype
    return np.dtype(dt[dtype])
TypeError: list indices must be integers or slices, not AttrValue

这意味着TensorRT当前不支持'Slice'层。 因此,我计划在代码中修改此层。 但是,即使我通过函数sess.graph.get_operation_by_name获取有关“ Slice”的信息,我也无法在代码中找到“ Slice”层:

graph list name: "final/lanenet_loss/Slice"
op: "Slice"
input: "final/lanenet_loss/Shape_1"
input: "final/lanenet_loss/Slice/begin"
input: "final/lanenet_loss/Slice/size"
attr {
  key: "Index"
  value {
    type: DT_INT32
  }
}
attr {
  key: "T"
  value {
    type: DT_INT32
  }
}

如何在代码行中找到“ Slice”层,以便可以通过TensorRT自定义层对其进行修改?

1 个答案:

答案 0 :(得分:1)

由于您是从Tensorflow进行解析的,因此最好查看TensorRT 支持的层。从TensorRT 4开始,支持以下几层:

  • 占位符
  • const
  • 添加,Sub,Mul,Div,最小值和最大值
  • BiasAdd
  • 负值,Abs,Sqrt,Rsqrt,Pow,Exp和Log
  • FusedBatchNorm
  • ReLU,TanH,乙状结肠
  • SoftMax
  • 平均值
  • ConcatV2
  • 重塑
  • 转置
  • Conv2D
  • DepthwiseConv2dNative
  • ConvTranspose2D
  • MaxPool
  • AvgPool
  • 如果紧随以下TensorFlow层之一,则支持
  • Pad: Conv2D,DepthwiseConv2dNative,MaxPool和AvgPool

从我在您的日志中看到的,您正在尝试部署LaneNet,是this paper的LaneNet吗?

如果是这种情况,那么它似乎是H-Net的一种变体,尚未阅读过,但根据该论文,其架构如下:

LaneNet Architecture

所以我看到了Convs,Relus,Maxpool和Linear,它们都受支持,不知道该BN,如果不在受支持的网络列表中,请检查一下以了解其引用的是哪一层您将不得不从头开始实施它。 祝你好运!