dims [-1] don't match configuration dims [-1]

Hello everyone, i server tensorflow saved model. My model configuration is:

  inputs['input_1'] tensor_info:
      dtype: DT_FLOAT
      shape: (-1, -1, -1, 3)
      name: serving_default_input_1:0
The given SavedModel SignatureDef contains the following output(s):
  outputs['yolo_nms_0'] tensor_info:
      dtype: DT_FLOAT
      shape: (-1, 100, 4)
      name: StatefulPartitionedCall:0
  outputs['yolo_nms_1'] tensor_info:
      dtype: DT_FLOAT
      shape: (-1, 100)
      name: StatefulPartitionedCall:1
  outputs['yolo_nms_2'] tensor_info:
      dtype: DT_FLOAT
      shape: (-1, 100)
      name: StatefulPartitionedCall:2
  outputs['yolo_nms_3'] tensor_info:
      dtype: DT_INT32
      shape: (-1)
      name: StatefulPartitionedCall:3

So my config file is:

input [
    {
        name: "input_1"
        data_type: TYPE_FP32
        dims: [ 320, 320, 3 ]
    }
]

output [
   {
      name: "yolo_nms_0"
      data_type: TYPE_FP32
      dims: [ 100, 4 ]
      label_filename: "labels.txt"
   },
   {
      name: "yolo_nms_1"
      data_type: TYPE_FP32
      dims: [ 100 ]
      label_filename: "labels.txt"
   },
   {
      name: "yolo_nms_2"
      data_type: TYPE_FP32
      dims: [ 100 ]
      label_filename: "labels.txt"
   },
   {
      name: "yolo_nms_3"
      data_type: TYPE_INT8
      dims: [-1 ]
   }

]

But when i server model using tensorrt inference server. i got error: E0602 10:52:59.510099 1 retrier.cc:37] Loading servable: {name: vehicle-detection ve rsion: 1} failed: Invalid argument: unable to load model ‘vehicle-detection’, output ‘yolo_nms_3’ dims [-1] don’t match configuration dims [-1]

How can i solve this problem?

[deadeyegoodwin]

The current 19.05 release uses TensorFlow 1.13.1 and it looks like the upcoming 19.06 release also uses Tensorflow 1.13.1. So you likely need to wait until 19.07 to get TensorFlow 1.14 support. As @GuanLuo mentioned an alternative is to use the master branch which now has a cmake-based build where you can build your own version of TensorFlow and use it in the TRTIS build. The online docs have not yet been updated with the new instructions but you can find them here: https://github.com/NVIDIA/tensorrt-inference-server/blob/master/docs/build.rst

If you decide to try the cmake build yourself be sure to let us know if you run into difficulties.

[GuanLuo]

Is “CombinedNonMaxSuppression” being added recently in TensorFlow? From here it seems like it is added in r1.14. So maybe the TRTIS you are using is linked with an older version of TensorFlow.