tensorrt savedmodel pb can not be imported into tensorboard and serving with tf serving
superhg2012 opened this issue · 0 comments
Hi, I am using tf-trt in my Tacotron model.
My question : the tf-trt optimized saved model can not import into tensorboard and deployed through TF-Serving
tensorflow : 1.13.0-rc0
tensorrt : 5.0.2.6
My steps :
1 export savedmodel using saved_model_builder:
saved_model.pb
variables
-- variables.data-00000-of-00001
-- variables.index
2 tensorrt optimization
workspace_size = 1 << 30
trt.create_inference_graph(
input_graph_def=None,
outputs=None,
max_batch_size=32,
max_workspace_size_bytes=workspace_size,
input_saved_model_dir=os.path.join(args.export_dir, args.version),
output_saved_model_dir="FP32_savedmodel_dir",
precision_mode="FP32")
the result is the savedmodel in "FP32_savedmodel_dir" directory
when deploying the FP32_trt_savedmodel files with TF-Serving docker, an error info :
"The TF function for the TRT segment could not be empty"
Also I tried to import the savedmodel pb to tensorboard using tensorflow tools "import_pb_to_tensorboard.py", the events was generated successfully, however when start tensorboard and load the events failed , it just hanged there.
could you help ?