Training models faster with TensorFlow Hub (TensorFlow Meets)

Key Takeaways

[Music] hi everybody we're here at the tents flow developer summit at the tensorflow cafe and I'm chatting with Andrew Gasper ovitch from tensor flow hub and you announced tensorflow hub today can you tell us all about it yeah so tensorflow hub is the new way to share what we're calling modules which is meant to be a reusable component so it's a little bit less than a model it's just the reusable portion okay but it is the graph the weights and potentially any assets that come with it so it's all packaged up into a component that makes it easy to share in just one line of code and one line of code yeah yeah that was a requirement from the beginning actually no more than one line of code where do you go from here off a lot of code yeah yeah well I guess the natural zero that's next year now one of the things you mentioned in your talk was about these modules being composable reusable and retrain of all could yes that's true that composable for us yeah so composable just means that you can do things like add your own classification so if you're talking about an image mod module it doesn't include the classification from the model so that is something that you can compose with what you're building in the case of text classification that may be something like an embedding module okay and then in other cases we also want to have just general-purpose modules that you can compose around almost like functions that you can call together and make building blocks for an entire model I say so so if I have a model that I don't know for example of doing OCR in an image and but it can also tag things within that image what saying is I could break out the OCR part and I could break out the tagging part and compose them into something new yeah absolutely absolutely and that's gonna be something really interesting to see what develops in the community over time just the ways of putting things together and sort of modules that people end up ryeol yeah so it's we're providing building all the communities providing building blocks yes we have we have a number of modules to start with and they're very general-purpose things doing you know image classification doing text classification we have those embedding modules but I think that over time what the community can contribute is what really will be interesting sounds great so we've mentioned that there composable and then reusable reusable just means that you can basically take something that already exists and apply it to your own particular problem makes perfect sense yeah and then I'm really interested in retraining below I mean we speak a lot about the attentive love for poets where you can retrain the final layer to make it flowers instead of general damages and is it to do a similar thing Gaudi yes yeah definitely so there is you know the classic sort of transfer learning case there where you would take an image classification module just up to the feature vectors and then retrain your own classification on top to do things but you know you can also go deeper than that you can start retraining text embedding modules and the thing about TF hub is that you can actually do fine-tuning and Retraining inside the module itself so if you have enough data you can actually turn on retraining inside the module because it's just a graph with the weights so you know you can really get better results because it's not just something that's static it's something that you can really include in your own model sounds good now in your talk you showed this for like classifying rabbits yes yeah yeah and you had that one picture of a rabbit which I think is the best picture of any in the entire day so we have to cut to that and show it's a beautiful yeah so tell us what did you do in that demo in the demo we basically just it's the same ideas the tensorflow for poets we start with a general purpose image classification model just up to the feature vectors and in the particular demo example we were classifying rabbits maybe including the Easter Bunny and you know we added our own classification on top and we fed in all of our own training examples and the end result is that you get something that is special to your task but includes all of the benefits of the general-purpose model nice it's kind of fun yeah how was accurate was at detecting rabbits it's very accurate we'll have to see how it works with the Easter Bunny but you didn't put the Easter Bunny in your test next year so so if I'm a developer and I want to build like I'm maybe I'm an expert in rabbit detection and you know and I want to build like a custom model and I want to contribute this into tensorflow hub how would I go about doing that well that's something that we're really excited to work on over the next few months right now a module is accessible via any URL and so you can host a module and you can use it in that one line of code but we want to create a platform where people can go and find modules for all sorts of different topics and have that one central place that you always know that you can get really high quality stuff in a variety of different areas and is there like some kind of curation that's something it remains to be you know seeing specifically what we're gonna do but we want to be able to open it up to it as wide a group of people as possible sounds good so where can I find this again it's tensorflow org slash hub tensorflow org slash hub I keep saying tensorflow hub / org we should remain tensorflow org slash hub yeah you know go check it out so thank you so much Andrew thank you it's been really fun and I love that slide okay so thanks everybody for watching this episode if you got any questions for me or if you have any questions for Andrew please leave them in the comments below and all of the links that we spoke about today will put in the text description so thank you so much and don't forget to hit that subscribe button [Music] you