TensorFlow Blog

Well, in reality tools like Tensorflow probability can help you model both aleatoric and epistemic uncertainty with probabilistic layers that have learnable priors and posteriors. The issue there is that for the average ML person might not have the required math skills to model the problem in these terms. For instance, if you look at https://blog.tensorflow.org/2019/03/regression-with-probabilistic-laye… until the case 4 it's easy to follow and digest, but if you look at the _Tabula rasa_ section I am pretty sure that such content isn't understandable by many. Where you get stuck because the ideas become too complex depends on your math skills.

This is just patently false. Most of the folks I know that have an M1 are students that were saving up to upgrade from a much older computer and got the M1 Air. I can assure you that they don't have a decent amount of money. Tensorflow has a branch that's optimized for metal with impressive performance. [1] It's fast enough to do transfer learning quickly on a large resnet, which is a common use-case for photo/video editing apps that have ML-powered workflows. It's best for everyone to do this locally: maintains privacy for the user and eliminates cloud costs for the developer. Also, not everyone has an imagenet sized dataset. A lot of applied ML uses small networks where prototyping is doable on a local machine. [1] https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performanc…

Apple can start leaning on their specialized ML cores and accelerators - Can you elaborate a bit more on this ? I was wondering how this would unfold and looks like things are moving in that direction https://blog.tensorflow.org/2020/04/tensorflow-lite-core-ml-delegate-f… If TF models can interoperate with CoreML - boy that'll literally be a home run for Apple cos eventually all ML frameworks will follow suit.

You're right about human, that would explain it, but still I find it surprising that such "common item" as a human is not there. Did you also try on items from the list ? If there is a match (and this is not frequent), to me it's still very low confidence (like noise or luck). It seems to be a repacking of https://blog.tensorflow.org/2020/03/higher-accuracy-on-vision-models-w… So an old release from 5 years ago (like very long time in AI-world), and AFAIK it has been superseded by YOLO-NAS and other models. MediaPipe feels really old tool, except for some specific subtasks like face tracking. And as a side-note, the OKR-system at Google is a very serious thing, there are lot of people internally gaming the system, and that could explain why it is a "new" launch, instead of a rather disappointing rebrand of the 2020-version. I'd rather recommend building on more modern tools, such as: https://huggingface.co/spaces/HuggingFaceTB/SmolVLM-256M-Instruct-WebG… (runs on iPhone with < 1GB of memory)

Prophet has gotten a lot of attention since being released in 2017, I think because the idea of a fully automatic solution is very appealing to people. One of the original developers, Sean Taylor, recently posted a nice retrospective on the project's successes and failures: https://medium.com/@seanjtaylor/a-personal-retrospective-on-prophet-f2… He quotes one of his earlier tweets: If I could build it again, I’d start with automating the evaluation of forecasts. It’s silly to build models if you’re not willing to commit to an evaluation procedure. I’d also probably remove most of the automation of the modeling. People should explicitly make these choices. Having worked on similar Bayesian time-series forecasting tools at Google, this matches my experience (though I've never used Prophet seriously, so please don't take this as any direct judgement of it as a software package). There is a lot of value in a framework that lets you easily experiment with different model structures (our version of this was the structural time series tools in TensorFlow Probability, see, e.g., https://blog.tensorflow.org/2019/03/structural-time-series-modeling-in… But if you're forecasting something you actually care about, it's usually worth the time to try to understand yourself what structure makes sense for your problem, and do a careful evaluation on held-out data with respect to whatever metric you're really trying to optimize. A fully automated search over model structures…

Great job. For someone who does not know Fortran, would you agree that the conclusion that can be drawn here is that PyTorch is good enough? > As you can see, fastGPT is slightly faster than PyTorch when doing as fair comparison as we can (both using OpenBLAS as a backend and both using caching, the default in PyTorch). You can also see that fastGPT loads the model very quickly and runs immediately, while both PyTorch and picoGPT take a long time to both load the model and to import all the Python libraries. One one core from your benchmarks I see that the Fortran implementation is about 4% faster than the PyTorch one. On 4 cores about 13% faster. Excluding the IO, the main advantage seems to come not from switching the language, but from using Apple's Accelerate framework. It appears to me that this framework is now available for Tensorflow [1] and PyTorch [2] too. Do you expect that once you port the code to GPU you will see a significant improvement over the GPU version of PyTorch ? [1] https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performanc… [2] https://towardsdatascience.com/installing-pytorch-on-apple-m1-chip-wit…

Tensorflow is accelerated just fine on AMD thanks to ROCm - everything you're praising Apple for doing, AMD has done for their hardware. You're disparaging AMD for not doing something they have actually done. There was even a tensorflow blog article announcing upstreaming AMD support, in 2018[0] > At this point, it is basically a sealed deal - if you're even remotely dabbling in data science, you better be working on a Mac. This is hilarious. You might have missed that the benchmark was missing Nvidia (or even AMD) graphics cards; I can't think of a lower bar for comparing ML performance than against Intel GPUs - perhaps Intel CPUs? While Apple has brilliant engineering, the M1 cannot possibly outperform the obscene number of transistors Nvidia & AMD throw at the task, even in older, mid-range cards. Not to mention power dissipation. If you're dabbling, you're better off with Google's Colab[1] which has (free) hardware acceleration which is roughly on par with my 3-year-old RX580 for my Tensorflow projects. Colab will work on anything that can run a browser. 0. https://blog.tensorflow.org/2018/08/amd-rocm-gpu-support-for-tensorflo… 1. https://colab.research.google.com/

It uses tflite in the background which can GPU accelerate with OpenGL ES 3.1 or OpenCL[0]. So it should work on pretty much any SOC. And you really notice that the model is dumber on GPU, because OpenGL doesn't take accuracy that seriously. [0] https://blog.tensorflow.org/2020/08/faster-mobile-gpu-inference-with-o…

New version release 1 month ago: https://blog.tensorflow.org/2024/03/whats-new-in-tensorflow-216.html

Better PDF with architecture picture about where it's used: https://developer.qualcomm.com/download/hexagon/hexagon-dsp-architectu… More links: https://en.wikichip.org/wiki/qualcomm/microarchitectures/hexagon https://pages.cs.wisc.edu/~danav/pubs/qcom/hexagon_micro2014_v6.pdf https://blog.tensorflow.org/2019/12/accelerating-tensorflow-lite-on-qu…

Very cool. Can you comment on how this compares with tensorflow similarity? https://blog.tensorflow.org/2021/09/introducing-tensorflow-similarity.…

DocTR: https://github.com/mindee/doctr It also has TensorFlow.js version to run in-browser: https://blog.tensorflow.org/2022/06/ocr-in-browser-using-tensorflowjs.…

Yes indeed: https://blog.tensorflow.org/2022/02/tfrt-progress-update.html?m=1 The ML Pathways infra is more complicated, and about more efficient scheduling many ops (in a graph) across many accelerators (rather than a small model on a single local GPU)

Whether PyTorch supports it... I'm not sure, but Tensorflow is still an immensely popular framework, and it is supported: https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performanc… https://developer.apple.com/metal/tensorflow-plugin/

These are not insurmountable, probably only a matter of time. Tensorflow has already been ported and optimised. https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performanc…

There has been an M1 fork of Tensorflow almost since the chip launched last year. I believe Apple did the leg work. It’s a hoop to jump through, yes, and no ones training big image models or transformers with this, but I imagine students or someone sandboxing a problem offline would benefit from the increased performance over CPU only. https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performanc…

The M1 gpu was comparable to a 1080. https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performanc… I believe they are working on PyTorch support.

Tensorflow (and by extension Keras) offload most of the actual work to C++ or C, so having those compile to WebAssembly would (I imagine) be a herculean effort. Instead, The TF team maintains TFJS, which can run on WebAssembly[0]. There are also tractjs[1], and onnyxjs[2], both of which allow you to run (most) ONNX models (which is an open standard for specifying ML models) using WebAssembly and WebGL(only onnyxjs supports WebGL). A bunch of frameworks (caffe, pytorch, TF) support exporting to/importing from ONNX. [0] https://blog.tensorflow.org/2020/03/introducing-webassembly-backend-fo… [1] https://github.com/bminixhofer/tractjs [2] https://github.com/microsoft/onnxjs

In order to benefit from optimizations in *this blog post* the model needs to be quantized to 8-bit integers. However, XNNPACK supports floating-point inference as well (including with FP16 weights), see https://blog.tensorflow.org/2020/07/accelerating-tensorflow-lite-xnnpa…

Conversation helper. Imagine you talk to a person. It can give you a list of ice breakers to open up the conversation based on the person's appearance and surroundings. It can alert you whether the person is not comfortable or very attracted to you. Some people like to miss these subtle cues. You can combine this AR technology with Deep Learning technology such as Pose Tracker. https://blog.tensorflow.org/2021/05/next-generation-pose-detection-wit…

I dont believe so...and looks like neither does either Apple or Google/Tensorflow - https://developer.apple.com/documentation/mlcompute - https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performanc… For a more personal take on your answer - do consider the rest of world. For example, Ryzen is very popular in India. Discrete GPU are unaffordable for that college student who wants to train a non-English NLP model on GPU.

> Swift was a weird choice for statically typed TensorFlow, being only popular on the platform, that does not have GPU/TPU support in TensorFlow, which is, basically, a requirement for any serious work. The fact, that they had to fork the compiler did not help either. This is available now (not sure if it's in mainline). https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performanc…

That depends on how tensorflow core ML performance compares to M1 neural engine core performance and the type and amount of memory. According to these notes, it performs roughly 4x as fast compared to an intel macbook which runs cpu only. 1.7GHz quad-core Intel Core i7-based 13-inch MacBook Pro system with Intel Iris Plus Graphics 645, 16GB of RAM, https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performanc… So I'd guess it's slower then an 1080 non-Ti

Well unless they also support acceleration on AMD GPUs, this is not so interesting. Training on x86_64 CPU cores or integrated Intel GPUs is really slow compared to training on modern NVIDIA GPUs with Tensor Cores (or AMD GPUs with ROCm, if you can get the ROCm stack running without crashing with obscure bugs). Also see the benchmarks in the marketing PR: https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performanc… The M1 blows away Intel CPUs with integrated GPUs (and modern NVIDIA GPUs will probably blow away the M1 results, otherwise they'd show the competition ;)).

TensorFlow announced their new Facial Landmarks model gained Iris features last month[1]. There's an online browser-only demo[2]. Note that this model doesn't say where the user is looking, it just extracts 2d locations of iris & surrounding face. Older software like WebGazer.js[3] provides more complete apis, reporting where on the screen someone is looking. The approach here seems to be server side. I wonder what tech they are using under the hood. [1] https://blog.tensorflow.org/2020/11/iris-landmark-tracking-in-browser-… [2] https://storage.googleapis.com/tfjs-models/demos/face-landmarks-detect… [3] https://webgazer.cs.brown.edu/

https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performanc… So way better than a regular Intel CPU, I've heard comparisons with the GTX 1050. Definitely not going to replace a top of the line GPU for ML.

In fairness, it's been possible to convert a TensorFlow model to a CoreML model for a while, and in April TensorFlow Lite added a CoreML delegate to run models on the Neural Engine. https://blog.tensorflow.org/2020/04/tensorflow-lite-core-ml-delegate-f… So don't think of it as Apple walked right in with this so much as Apple has been shipping the neural engine for years and now they're finally making it available on macOS.

Then try this one https://blog.tensorflow.org/2020/06/estimating-pitch-with-spice-and-te… There is a online demo here https://tfhub.dev/google/lite-model/spice/1

Google seems to think it is currently useful. They have a WASM back end to tensorflow.js to use when WebGL isn’t available and you only have CPUs to run on https://blog.tensorflow.org/2020/03/introducing-webassembly-backend-fo…

One more link from TensorFlow Blog: https://blog.tensorflow.org/2020/08/announcing-tensorflow-lite-micro-e…