Adam Pocock comments

Results 375 comments of


                                            Adam Pocock

Instructions to enable Core ML with Apple Silicon

Fixing some of the free dimensions does move more ops to the CoreML execution provider, but not all of them with ORT v1.16.3, so on my regular M1 it's still...

Instructions to enable Core ML with Apple Silicon

I retested it with the new CoreML support in ORT 1.18.0 and unfortunately that still doesn't support many of the nodes used in this. Still needs more work inside ORT.

[java] Adding ability to load a model from a memory mapped byte buffer

> @Craigacp this might need a merge with main due to the pipeline definitions being updated. I can't run the 'Python format' CI with the current set of changes and...

Problem with loading model via SavedModelBundle.load()

What version of TF-Java are you using, what's the deployment environment (OS, Java version etc), and what platform does it work locally on? Is this using TF-DF?

Result of tf.image.resizeBilinear in Java is different from tf.image.resize In python

Are you specifying `halfPixelCenters = true` when you run that op in Java? It's set to true by Python's `resize`.

Result of tf.image.resizeBilinear in Java is different from tf.image.resize In python

Are you running it on the same hardware (i.e. CPU or GPU) for both languages? Also can you check the behaviour of tf python 2.4.1?

Getting errors on copyright comment when trying to build PR

I think it would be best if we turned off the license checking in spotless. It's going to complain about all my PRs with Oracle copyright on.

TF Java 0.3.1 shows a performance degradation on GPU compared to v 0.2.0 when loading Hugging Face models

Do you know if it's during tensor creation/access, model inference (and is that model trained in Python or Java), training or some other place? Also is this across multiple different...

Framework: Move Ops parameter to call method where possible

One other wrinkle is that the initialisations are only populated in the `Graph`, but they aren't persisted in the `GraphDef`. So we do tricks like [this](https://github.com/oracle/tribuo/blob/main/Interop/Tensorflow/src/main/java/org/tribuo/interop/tensorflow/TensorFlowTrainer.java#L477) in Tribuo to initialise...

BlockLSTM output values differ from Python

There's a determinism problem in the TF C API with how it builds the gradient operations too. I've reported it upstream but no one has looked at it yet.