TensorFlow Probability: Learning with confidence (TF Dev Summit '19)
Вставка
- Опубліковано 14 чер 2024
- TensorFlow Probability (TFP) is a Python library built on TensorFlow that makes it easy to combine probabilistic models and deep learning on modern hardware (TPU, GPU). It's for data scientists, statisticians, and ML researchers/practitioners who want to encode domain knowledge to understand data and make predictions with uncertainty estimates. In this talk we focus on the "layers" module and demonstrate how TFP "distributions" fit naturally with Keras to enable estimating aleatoric and/or epistemic uncertainty.
See the revamped dev site → www.tensorflow.org/
Watch all TensorFlow Dev Summit '19 sessions → bit.ly/TFDS19Sessions
Event homepage → bit.ly/TFDS19
Subscribe to the TensorFlow UA-cam channel → bit.ly/TensorFlow1
Speaker:
Josh Dillon, Software Engineer
Music by Terra Monk → bit.ly/TerraMonkTFDS
Event Photo Album → bit.ly/TFSummit19
event: TensorFlow Dev Summit 2019; re_ty: Publish; product: TensorFlow - TensorFlow Probability; fullname: Josh Dillon; - Наука та технологія
As an applied mathematician, if I had to thank Google for a single thing, it would be this. Really going one step further in scientific inference.
I'm impressed with the presenter's delivery. Well done! And funny!
What would be really cool would be a series of instructional videos on using TF Probability, like the Coursera course for Tensorflow recently. The material seems be be more challenging than tensorflow for CNN.
The more I dig into tensorflow, the more I realize how incredible is this tool. Great job, I am sure investing time learning tensorflow is no waste of time
Hell yes TFP team, making my life easier! (and making my code readable, through Keras, for my team members) :D
Don't watch this if you get jealous of what other people know. I'm feeling like I've been remanded back to Machine Learning 101.
Hey it just means you have more upside!
This is amazing!
Outstanding
7:17 - interestingly, I think the reason why we end up with a line for the mean and variance is because there is only one layer in the network
(and hence, mean = theta0 + theta1 * x1 + theta2*x1 +... and hence this will be equation of a line.)
Amazing man
In the end, it looks like a Gaussian process with automatic feature extraction layers!
1. Before training the network it must return the prior distribution of the outputs for any input, am I wrong?
2. If we have (1) how to show that the training procedure is using Bayes Rule to update the priors, I mean replacing X by X|y?
And nine days later... AttributeError: module 'tensorflow_probability.python.layers' has no attribute 'VariationGaussianProcess'
pip install tfp-nightly
Is there a newer version of the book? Josh said: "check out this book which we rewrote using TF probability" .... where can this newer version be found? Thanks!
It's on github. github.com/CamDavidsonPilon/Probabilistic-Programming-and-Bayesian-Methods-for-Hackers/blob/master/Chapter1_Introduction/Ch1_Introduction_TFP.ipynb
7:30 - Looks like a Gaussian Process Deep Neural Network (GPDNN) ?
What is the probability I'll be using this? 100%
lol, good one.
this was my prior too!
I don’t think this estimate took into account the risk of complete human extinction within hours after posting it.
@TensorFlow one question , I want to use it in anaconda . First I install tensorflow-probability but than if I want to use It I get the error "module TensorFlow has no attribute contrib" . Before that I upgraded TensorFlow. So I am very confused
If you have updated to tensorflow 2.0, then you need to install tfp-nightly. See github.com/tensorflow/probability/issues/320
Do you have tf2.0? I believe this part of tf2.0
how to update all reference to tfp.distributions instead of tf.distributions?
The sample code doesn't seem to work in TF 2.0? Is there an update for TF 2.0?
great
wow
1) Where do you get the "dots" from?
2) Does all data have to be converted to these "dots" or can a CSV file work?
3) How would you create data and convert them into these "dots", is there a special program.
4) What is the best way for tensorflow to predict data? I have been watching a ton of videos and this is the first time I seen these "dot" system all the videos I have been watching have been using CSV files from MNIST Dataset. (either locally or linked to a website"
** I am from a journalistic background and I do appreciate the presentations but a lot of these tutorials are either : Outdated, code does not work on computer, but will work on theirs (teachers)
Not clear on how you get data, how to create your own data, where do you put your data(files) so that tensorflow can access it.?
I'm a little late to reply. If you are referring to the plotted example points as "dots", those are likely generated by adding random noise to a predetermined function, and could be stored in a CSV, binary, or in memory. You could generate the data using Excel or MATLAB/Octave, but most likely all this was done in Python or R using a Jupyter Notebook. You may need to watch some tutorials on neural networks to get a gist for how machine learning works and how TF is used, essentially TF is a framework most-often used for creating and training neural networks and other architectures that require backpropagation.
In addition to calling and running the Tensorflow API, most of the preprocessing (generating data, normalizing, sorting) and post-processing (saving, plotting, etc) will be done using a programming language such as Python or R.
Is TFP offering a similar functionality (or vice versa) as Uber's Pyro ?
Like what specifically? (Also note: a lot of pyro is actually based on TFP including distribution shape semantics, bijectors, use of coroutine for joint distributions, etc.)
"But wait, there's more..."
Where can I get the book?
Here, of course - www.amazon.com/Bayesian-Methods-Hackers-Probabilistic-Addison-Wesley-ebook-dp-B016060UHA/dp/B016060UHA/ref=mt_kindle
camdavidsonpilon.github.io/Probabilistic-Programming-and-Bayesian-Methods-for-Hackers/
github.com/CamDavidsonPilon/Probabilistic-Programming-and-Bayesian-Methods-for-Hackers
camdavidsonpilon.github.io/Probabilistic-Programming-and-Bayesian-Methods-for-Hackers/
www.amazon.com/Bayesian-Methods-Hackers-Probabilistic-Addison-Wesley-ebook/dp/B016060UHA/ref=sr_1_1?keywords=bayes+for+hacker&qid=1552242788&s=gateway&sr=8-1-spell
github.com/CamDavidsonPilon/Probabilistic-Programming-and-Bayesian-Methods-for-Hackers
tf.keras is so bloody awkward to type everytime.
from tensorflow import keras as tk
Or avoid it entirely like
`from tensorflow.python.keras.layers import Dense`
2:57 glitch in the tensor
Mixed models
Okay I feel stupid.
Don’t feel stupid. you don’t need to know everything, this is just machine learning but with LOTS of probability theory. You can do the same things with just neural networks.
rip ai for u if ur bad at math
Tbh all you need is numerical analysis and bayesian statistics. If you know these two plus a programming language then you are golden
@@harjitsingh7308 Lol, surely there's a lot more math (derivatives, optimization theory, etc) involved
Pytorch is losing ground. Lack of engineers is real.
@SMC Good point, Is Pyro offering a similar functionality as Tensforflow probability ? I've recently stumbled across Pyro (or even probablistic programming as a matter of fact)
Ok, what is 'Unknown known'? a.k.a You don't know what you know.
* Known unknown: data variance, Heteroskedastic, aleatoric uncertainty, you know what you don't know.
* Unknown unknown: Bayesian posterior p(w|x), distribution over the weights, epistemic uncertainty, you don't know what you don't know.