this same SavedModelBundle load library works for tensorflow-text(2.8.1) resource using mac .dylib file.
when we run our test in Jenkins which uses Docker image, it tries to load its .so counterpart file (counterpart to same .dylib file for wordpiece_tokenizer and some other tensorflow-text ops) and we get these errors for these ops resource files.
java.lang.UnsatisfiedLinkError: java.lang.UnsatisfiedLinkError: /home/jenkins/workspace/lization_Search_build-all_PR-313/spell-corrector/target/scala-2.13/classes/bertNeuSpell/tokenizer/_wordpiece_tokenizer.so: undefined symbol: _ZN10tensorflow8OpKernel11TraceStringEPNS_15OpKernelContextEb
Same error for other .so files loaded to run Bert Tokenizer Graph like this:
java.lang.UnsatisfiedLinkError: /home/jenkins/workspace/lization_Search_build-all_PR-313/spell-corrector/target/scala-2.13/classes/bertNeuSpell/tokenizer/_normalize_ops.so: undefined symbol: _ZN10tensorflow8OpKernel11TraceStringEPNS_15OpKernelContextEb
and so on for other _regex_split.so file as well.
Does this mean docker image OS is not compatible with this tensorflow text (2.8.1) ops file. ?
We are only building our scala/java app using java tensorflow core platform 0.4.0 which works fine when these resource ops files are loading on mac using .dylib extensions and docker image doesn’t do any pip install for tensorflow or tensorflow-text in the docker file- we just directly load the .so files before loading the tokenizer Graph.
Another thing, since i couldn’t get .so file when i do pip install tensorflow-text locally on mac, we have databricks cluster, i had done copied files from the driver node, i am guessing that’s reason it won’t work. should i just run docker image , and do pip install in it and grab files out of docker image and checkin into our github repo ?