Repository navigation
Simplify generic parameterization on some operations #176
Description
I'm working on the reified generation for the Kotlin API, and I'm noticing that lots of methods have unnecessary type parameters that makes the reified usage much less nice (since you have to specify all type parameters if you specify one). The best example is probably cast, which has the signiture:
<U extends TType, T extends TType> Cast<U> cast(Operand<T> x, Class<U> DstT, Cast.Options... options)T is completely unnecessary and could be replaced with ? without issue, but it prevents cast<TInx32>(x) usage from Kotlin. This shows up in a number of Ops, mostly with the unnecessary type parameters on the input. It essentially needs a "is this type param only bounded by TType and only used on inputs" check.
Originally posted by @rnett in #174 (comment)
Activity
karllessard commentedon Dec 28, 2020
I think it would be worth it to make this simplification yes, but the rule might be a bit more tricky. In fact, a generic parameter should be enforced if any input or output of the method should share the same parameter. In other words:
Even if a single parameter is not bound to
TTypebut to one of its subtype (likeTNumber), we can handle it with a wildcard, e.g. for a hypothetictoFloatoperation:So this logic must be added to the C++ op generator.
rnett commentedon Jan 14, 2021
@karllessard do you have any idea when you would get to this? Should we wait for it before merging in the Kotlin API?
karllessard commentedon Jan 14, 2021
Hey @rnett , I was not necessarily planning to do this work anytime soon, do you want to give it a try?
rnett commentedon Jan 15, 2021
I can take a look, at least. Do you have any recommendations on IDE setup? I had issues getting bazel to work with CLion when I tried it.
karllessard commentedon Jan 15, 2021
When I was doing more C/C++ code than Java for TensorFlow, I was using Eclipse which supports well both languages, even in the same project (at that time, I had to edit the project file manually to enable this hybridity, I don't know if that can be done differently now).
rnett commentedon Jan 22, 2021
@karllessard I've got this mostly working, I think, but Javadocs aren't being generated any more, even on
master. I'm running with./bazel-out/k8-opt/bin/external/org_tensorflow/tensorflow/libtensorflow_cc.so --output_dir=./src/gen/java --api_dirs=./src/bazel/api_def, any idea what would cause this? Usingbazel-tensorflow-core-api/external/org_tensorflow/tensorflow/core/api_defdoesn't work either, and makes everything acoreop.karllessard commentedon Jan 23, 2021
I'm not sure to understand, you are running thelibtensorflow_cc.solibrary? or that's probably a mistake, should be thegen_op_javabinary or something like that.(oh yes I recall now, you need to pass the path to the library as well as an argument, that's something relatively new.)
But you need to pass both dirs in the
api_dirsoption, separated by commas if I recall correctly. Also the order matters (sorry for not being more clear in my instructions but I'm not in front of my computer right now)rnett commentedon Jan 23, 2021
Ok, It looks like it should be
bazel-tensorflow-core-api/external/org_tensorflow/tensorflow/core/api_def/base_api,src/bazel/api_def. This doesn't generate a few ops (DummySeedGenerator,TensorScatterMax, andTensorScatterMin), should I be concerned? The other order doesn't generate them either and misses the group for some data ops.rnett commentedon Jan 23, 2021
Also, for anyone else trying to use the bazel project, I got it working w/ an old version of CLion (so that I could use the bazel plugin), as long as ran
./configurein a clonedtensorflowproject, and copied the resulting.tf_configure.bazelrctotensorflow-core-api.@karllessard do you think we should make a
CONTRIBUTORS.mdstyle file with build tips like this, the exact command to use for cc op generation, etc?karllessard commentedon Jan 24, 2021
Yes, we've mentioned that during the last meeting, we definitely need better documentation when it comes to build from sources and guide the users in the process, we have a lot of issues raised by them that would probably be avoided by just having this documentation, please feel free to start it!