TensorRT Non-Unique Argument Names for Saved Model

[RoboEvangelist]

I habe Ubuntu 18, tensorRT 7, tensorflow 2.1.0

Here is the error I get: ValueError: Got non-flat/non-unique argument names for SavedModel signature 'serving_base64': more than one argument to '__inference_signature_wrapper_144191' was named 'image_files:0'. Signatures have one Tensor per named input, so to have predictable names Python functions used to generate these signatures should avoid *args and Tensors in nested structures unless unique names are specified for each. Use tf.TensorSpec(..., name=...) to provide a name for a Tensor input.

Here is the conversion code I use:

import tensorflow as tf from tensorflow.python.compiler.tensorrt import trt_convert as trt params = trt.DEFAULT_TRT_CONVERSION_PARAMS._replace( precision_mode='FP16') converter = tf.experimental.tensorrt.Converter( input_saved_model_dir="/tmp/saved_model", conversion_params=params) converter.convert() converter.save("/tmp/saved_model/saved")

How exactly can we convert using tensorrt?

Thanks,

[RoboEvangelist]

I wonder if trimming the network and eliminating layers preceding "image_arrays" can help with this?

[tcxia]

@RoboEvangelist have you solved this problem?

[RoboEvangelist]

@tcxia no, I have not. I still don't know how to fix that problem I try to optimize the graph with tensorrt

[mingxingtan]

Hi @RoboEvangelist @tcxia , I removed the other signature. You may want to try it again?

[mingxingtan]

I am going to close this issue since signature is unique now. But feel free to open another issue if TensorRT has other issues. Thanks!

[RoboEvangelist]

I am going to close this issue since signature is unique now. But feel free to open another issue if TensorRT has other issues. Thanks!

Awesome! Thank you very much for this amazing work that you guys are doing!

[RoboEvangelist]

Yes, the fix worked. Thanks.