unsupported tensor op : InceptionResnetV1/Conv2d_1a_3x3/BatchNorm/cond/AssignMovingAvg about armnn

[MatthewARM]

Hi @solderzzc the current tensorflow parser cannot guarantee the correct operation of control inputs (which is tensorflow's way of guaranteeing a particular order of execution).

You can try removing the code block at TfParser.cpp:489 to get past this parse error, but if the control inputs are guaranteeing execution of some variable assignment or something, you will probably just hit another parse error. The full list of supported TensorFlow ops is at: src/armnnTfParser/TensorFlowSupport.md

It seems to me that maybe it's not a frozen inference-only model? If you try following the steps at https://www.tensorflow.org/mobile/prepare_models it will remove all training-only nodes and you will have a much smaller, cleaner graph that is more likely to work with ArmNN.

All the best,
Matthew

[solderzzc]

Hi @MatthewARM

Thank for your reply. I removed the code block at TfParser.cpp to bypass parser error. The model has phase_train (ref: davidsandberg/facenet#357 (comment)) to control graph inference flow, so it might has issue during runtime.

I also called optimize_for_inference API of tensorflow to have training-only nodes removed(same behavior as mentioned in https://www.tensorflow.org/mobile/prepare_models), then hit another error log:

terminate called after throwing an instance of 'armnn::ParseException'
  what():  Currently only FLOAT is supported for tensorflow nodes (apart from Const)
Aborted

I think this is because that FLOAT variables were converted to CONST during Freeze Graph phase.

Reference:
https://github.com/davidsandberg/facenet/blob/master/src/freeze_graph.py#L88

graph_util.convert_variables_to_constants(
        sess, input_graph_def, output_node_names.split(","),
        variable_names_whitelist=whitelist_names)

Thanks
Simba

[MatthewARM]

Hi @solderzzc "Currently only FLOAT is supported for tensorflow nodes (apart from Const)" indicates that there is an integer operation in your network, do you know if that is the case? Currently we only support float operations.

If you can find out which operation, and what data type, it would help us prioritise the support work internally. Or we can find a workaround.

All the best,
Matthew

[solderzzc]

Hi, @MatthewARM
I just go through the network design of facenet, the non float ops should come from dropout layer of tensorflow.slim.

Network define with dropout

Slim dropout API

Best Regards
Simba

[MatthewARM]

Hi @solderzzc, dropout should definitely get removed by the 'strip unused nodes' transform. Maybe you could try the mobile inference preparation steps again?

[solderzzc]

Hi @MatthewARM, I loaded the graph into tensorboard, the 'Dropout' layer is not included in the freezed graph. The issue should be introduced by DT_BOOL phase_train.

Added more information in TfPhaser.cpp line 1951:

type: 10 op: Placeholder name: phase_train

After replace phase_train into bool constant:

terminate called after throwing an instance of 'armnn::ParseException'
  what():  Unknown DataType DT_BOOL for node
Aborted

After convert phase_train into FLOAT32(Just for bypass this exception):

terminate called after throwing an instance of 'armnn::ParseException'
  what():  Unsupported operation Switch in tensorflow::GraphDef
Aborted

Thanks
Simba

[MatthewARM]

Thanks for the detailed analysis, @solderzzc.

We can try to get a proper fix done in a future release, but I think you can work around the issue by removing all connections to 'phase_train' from your graph. In the ArmNN TensorFlow parser we will always assume that 'phase_train' is FALSE.

[solderzzc]

@MatthewARM I will try it and look forward for the next release.

Thanks