Deep Learning with PyTorch for more Fun and Profit (Part II)

Speaker 1 [00:02]

um hello everyone um this is a talk for deep learning for fun and profit uh it's not so much about pytorch i use pytorch i love pytorch but i don't have too much pytorch in here actually a little bit by myself um i think you have seen my face already at this conference maybe more than enough um i am a partner and senior consultant for data science um at the Königsweg, we're a boutique consultancy in Mannheim. I'm also working in the Python community as a program chair for EuroPython, EuroSciPy. So I was quite active this year. So I'm happy this is the last conference I have to work on. So one thing I have to mention here, we launched PyData in Frankfurt. So if you're from the Frankfurt area and you want to come around, check it out. We are on Meetup. If you want to give a talk, just ping me And the next PyData meetup is at the 9th of November at the Tech Quarter in Frankfurt. This is an ongoing talk, so a little bit real quick catch-up on the story so far. Actually, the first iteration or the first part was made for PyData Berlin this summer, how I made my computer write the first story. And it continued at Europython with some updates. Both talks are online on YouTube, so I'm going to give you a quick walkthrough, but the extended versions are, yeah, you have to check them back online. So, quick update. I really have to rush a bit because there's a lot of ground to cover. The inspiration came from Gene Kogan. Gene Kogan is an artist, and he gave a keynote at PyData London in 2016. And some of you know, actually, my IT and data science, my second career, my first career was in the 90s in the music industry. we were doing techno and house labels and it was already about digitalization but that's a story for another time so of course i'm always drawn to art um in a way um and creativity um so i his talks also online so hey great um finally um the beginning of the year i had a project and i finally could get an e-gpu to build experiments on this very computer if you don't know what e-gpu is There's like Thunderbolt, USB-C. You can plug an external graphic card into a MacBook or another computer and basically run your experiments locally. This is what I wanted to do. Have something local here. And if I need to scale or do something, need more power, move it to the cloud. So this is just like an NVIDIA GTX 1080. It's not yet plug and play. But there's a great source. If you want to look into space, eGPUO, there's some great guys that made some great scripts, checking your computer, installing the right drivers, and all the things. Okay. So the first thing after the business project, I said, okay, now I have the equipment, now I can play and explore the space more, especially in the art place. So I started doing style transfer. Who knows style transfer? Okay. So style transfer is, for those of you who haven't heard about it, learn how basically a painter painted a picture, apply it to a photograph, and create something new. And so I said, well, okay, let's jump into this. And this is a comic. This is a comic I used to read as a kid. I went to the public library, and I liked these French comics. Valeria and Veronique, it's in German. I think it's a little different name in French, but I really love the style and stuff. yeah what's it is uh yeah maybe um okay and for example i love these comics and so this is like how a modern flash comics look like so okay it's okay but i don't like this stuff so i asked myself the question can i basically learn from this comic back from the 80s and transfer it on a picture like this and how will this look like this is my very first experiment so for those you don't know we learned from the the image and I squared the learning part and we learned how it's painted there and applied here and this was the result and I was just like oh my god yes Wow this worked pretty well and so I was really like excited took holiday photos pictures I had taken at an art museums like this yeah you see like the different styles this is the flash style This is like the villainy mix on and then the left is the original You can also like have like psych learn like psychedelic styles and apply them to Game of Thrones the tiger holiday pictures Okay, this was fun. Actually, you can even apply it to a whole page of a modern comic and this is all not yet Optimized yet. So this was just like very first baby steps One nice finding on the way was what about like old and blurry pictures like this very old flash comic from the 50s I found a really bad scan on the internet. Nice finding, and this is probably also a useful thing for business and other things. You can really, it adapted the style, but it became readable again. Basically, the resolution is there. Basically, you could almost print it. You can also apply styles to white noise and have a picture like this. So we learned this style, applied just to white noise and a picture like that. And I just sent it to some friends. I said, hey, I'm at Tate Modern in London. I just sent them the picture. How do you like this painting? I really fancy it. And they said, hey, well, yeah, it's a great painting. This year we were at Europython in Edinburgh, and this is a very famous picture. If you see a lot, if you see Edinburgh, this is Edinburgh. And I just tweeted it to announce my talk at Europython about the same talk here, like the freeze version. And I just thought, okay, I just wanted to, like, promote it. And then somebody said, okay, I took this picture. And I said, okay, it's like I thought it was an attendee knowing about the domain. And I said, oh, yeah, it's a reality transfer. And he just replied, no, it was a picture. I took it a few days ago. And I was wondering, when did you draw the painting? And I just replied, no, I didn't draw it. It was made by Artificial Intelligence. And then I said, oh, I see. Even more amazing. And of course, you can just go to a conference and do your own comic. Once you have trained a model, it's really fast to apply. And you can even have a whole comic once upon a time in October in Karlsruhe. and enter, like you see Peter and Sebastian here, and you see conference organizing is fun, the pretzels, you see stickers, all the stuff, and quite unbelievable, and you can have like, yeah, you can like document the whole thing. Here, I didn't really build the story, didn't have time for that but this is just like this morning so i take him a picture from a tweet and then we just have veskin's keynote here as in the comic style okay um the technology you know like local cpu chip ability to the cloud reproduce and another big issue was reproducibility and documentation of the experiments when i just like played around i didn't document a lot like random seed parameters i was just toying around and i think for the part is fine but then i had um i i i saw okay we need documentation i want to reproduce my experience because if i cannot explain why something works or not that's really bad so but you have to look at the other talks um for that so this is all now um solved so um you can imagine i got a little over excited with like the style transfer because everything was just like amazing it felt like this um and so i thought okay i very often don't have the possibility to present a real use case from a customer um so how can i deal with that um it's uh so what can i do to give you something of value something which is close to real life and to explain stuff and to say basically turn around say okay here's a challenge and now we're going to throw deep learning at it and see how far we can get so it's not just okay we have this tiny thing when we can solve it probably with deep learning so let's say okay let's take a full product and try to fully synthesize it and of course i was also drawn back to the days when i was a child so i thought what about um yeah who knows them who knows what it is oh yes who doesn't know this is like the three it's the three investigators in original is a book series from the u.s um it's like teenage detective and mystery crime stories um it goes back to 1986 it's not really popular in the u.s I think it's only popular in Germany. But in Germany, it's very, very popular. And so there are like 200 taped radiodramas. Hörspiel. Who knows Hörspiel? Okay. Everybody. I think this is all the Germans in the room. Okay. There's like 200 taped radiodramas. I'm going to explain them in a second. The radiodramas have sold more than 45 million copies over time. This series is going now for 40 years plus. So I thought, okay, eventually the people to speak the recordings will eventually die, so maybe it's time now to get into deep learning so we continue the next 40 years. So, okay, a little bit more radio dramas. Hörspiel, what is that? Taped radio dramas goes back pre-television. No, pre-internet, pre-television. We're now like in the 1920s when people are having radios and what is a tape radio drama. It's basically a movie without pictures. So people were just listening to the radios and you had actors telling you a story. And very famous, and if you don't know it, I can really recommend looking into it, one of the pioneers was Orson Welles. You see like 23-year-old Orson Welles in the 1920s or early 30s, I'm not sure, he made a radio drama and not only a radio drama he made the radio drama like a documentary and about aliens invading the united states and people were panicking because they thought it was real because the whole radio media thing was pretty new so really good um it's great so this is like a news article here so how does the three question look like so we have like the three question marks in the whispering mummy and it sounds like this yeah okay here you hear how it sounds like it's more like a teaser you see it's like This is a script. Luckily, I found scripts on the internet. They were handmade by people, so this was really good. And this is basically how the artwork looks like. And who's an expert would already say, hey, Alex, there's only half of them are fake already. But more on that later. So, okay, which are the ingredients to be synthesized? Actually, the story, the plot, what's happening, the dialogues, it's very dynamic human speech. people are interacting we have to have the artwork a cover just as i showed you um and we need spoken word and something to synthesize it um so um and now we're going to look into the different areas not all i covered yet uh because there's quite a lot of ground to cover i realized so which resources are required for each individual step um choosing the right ai or Neuronet for that. Data acquisition and processing. What are the costs and time? Because I'm a business consultant, so I have to advise my customers where to spend their money. So how much money and effort have to go into each of the single steps? And which real problems can we potentially solve and find on the way? Because experimentation in AI is something that is really an important part as well. Okay, so let's go to the journalization. So we had transcripts. We're still, like, in the upload slide. Extended versions are online on YouTube. The technology stuff, we covered it. The process is very important. ASL, always take the same process for each and every stack. Data acquisition, data cleansing, researcher paper, and a working solution. Verify it's actually working. It's not somebody just writing the next blog post. And adapt the solution for the specific use case, and then go to maximizing quality. And we're going to, so always like the very essence of this is always only exchange one variable at a time. So you can see if something's not working, you can investigate what is not working. And I think this is one of the most best, one of probably the best advice, never change more than one variable and stuff like that because it's hard to figure out what's next. So fabricate text. Who knows Andrew Capaciti's blog post, The Unreasonable Effectiveness of Recurrent Neural Networks? Okay. 20%? Okay. So, Andrew Capassi is a really great guy. He wrote this blog post a while ago. It's a really readable and very nice blog post, and he explained how we can use RNNs for text generation. So, he learned Shakespeare sonnets, and he trained a model for Shakespeare sonnets, and then created new Shakespeare sonnets. This was very impressive. This is very often quoted. He was in his PhD, and now he's, I think, the chief data scientist of Tesla or something like that. So, really great. I really can recommend reading it. The link's down there. So, the process here. I found fan-made transcripts, which was great. So, I have the text. I pre-processed them. I also identified the characters via some fan wiki. So to identify who's actually speaking, to match names to text. And the total text corpus is like 1.5 million words in the German language. And 56,000 are unique. This is an RNN. This is how it works. I don't want to go deeper here. And I just like, the question was, is this text corpus big enough to produce something? I mean, it sounds like a lot. Oh, yeah, it's like 200. Radiodramas, is this enough to produce something in the German language? and let's see and this is like the some stuff the neural network actually produced in german language and below is a google translate which is actually quite accurate so you see okay yeah i probably can sell this as german to many people around the world but not to germans i really like the schlokaliana actually that's a really nice work but you see okay it sounds But it's not really something useful so you can of course like also use it and send other stuff to your friends Okay, okay, I settle for character base. So actually we're learning on what's the network predicts. What's the next character? The cooking and characters we use this because we need it's a lower space and matrices because they're maybe only like 30, 36 characters to cover. We can also try it word-based, which of course is a lot more to compute because we have 56,000 unique words and it came up with something like that, which is also really nice and funny. So, well, okay. Wasn't quite as expected after my style friend, but we had fun. Okay, so I thought, okay, is RNN the only solution i came up on a lot of stuff during my research uh for this talk and so i i also okay what about chatbots so and i read about there's somebody who took star trek and star trek next generation transcripts and build a picar wolf and all these chatbots and i said okay let's see how that stuff works because reading the blog post was really okay yeah awesome stuff So I tweeted to Captain Picard. Actually, I took one line from the real transcript to make it easier for the neural net. So I threw something at the neural net, which the neural net actually should know. But actually, it was just like noise and, you know, fights and beans. And it was not really fun. It took me like a week to get rid of answers from Wolf and data and stuff. So, OK. So text was not nice. So, of course, deep learning now is the new kid got on my blog for three years now. So it feels new. But, of course, I just looked at my bookshelves. All this stuff is you can probably see for the covers. The problems are not new. The problems and deep learning and neural nets, they go back to the 60s, 50s even, with the same problems. It just woke up again. So, okay, next challenge, text, okay. Let's put the text to sleep, sleep it over a little bit. Maybe I can come up with something better. The next step, artwork. Artwork, we use a CNN. So we basically compare two pictures and basically the neural net takes smaller parts of the picture and the neural net can actually learn how a certain style. for example like the from the original paper um tubing at night with van gogh style so i tried that i we probably know already from the my introduction this works well many styles but can i also like have it learn on something on on the style of the three investigators because they have a very specific style um it's not just van gogh style it's like over the years styles have evolved and you can now take a quick guess 50 are fake which ones three two one these are actually fake they're fabricated i just choose a picture and throw in some art styles i'll have learned on the three investigators there um so this is fun actually the zelzheimer wecker was very obvious because this is like super classic so okay really nice so we can fabricate some artwork i guess um so there's also some stuff people on the internet will probably not tell you or mostly they only talk about stuff that works well uh for example like i had um a data set of pictures like 30 pictures to see how the network was learning so i always like throw people a nice set of different pictures to see what's going on for example this picture hardly ever worked on any style transfer which was quite so it's not always working um you can reason about why this is not working um well i have to check that on video um of course it's really nice to impress you with stuff like that but we also have to see there's a lot of stuff like that and even worse while training and doing the stuff. So it's not just like throw something, deep learning and magic. No, no magic. It's just linear algebra. To sum all my findings up, Ned Radcliffe is a very bright guy. He talked about quantum computing at EuroPython. He put it this way. Okay, yeah, well, now the updates. So I've looked more into the speech space. How can we fabricate speech? So basically some sent text to a model and have it basically talk to you in a human-like voice. This is the network. It's Techo Turn True. So basically the ideas we learn from male spectrograms. Male spectrograms is basically how we say something. So it's like a very rich spectrogram. And then we try to learn it, predict it, and align. This is like how the two are aligned. And if you see a straight line here on 45 degrees, it's really well aligned. So let's look into that. So training the network requires audio snippets, text, and the text. So I found a really nice data set of 24 hours recording by this one female speaker of 24 hours of non-fictional English data. Prepared, cut, really nice data set. And I had it learned. And now the network will always read this first sentence to you. A little bit louder, please. Okay, you see, like, after 10 hours, it's, like, starting to blow. You see something in rhythm happening, so it's okay. It's not very useful after 14 hours. Okay, after five days... So, actually, I was quite impressed. She didn't get to go out of the sea like any other. This is something I did not expect. Having a nice data set, throwing Tucker-Tron TrueNetwork, and producing these results within a few days. This was, I think, quite amazing, because you have to consider, I found the data open. It's an open-source data set online. I found a nice and really well-coded GitHub repository for an NVIDIA for the network. and um yeah actually it was a it's a this is low budget project i mean this is maybe like three days four days of work uh for me and basically um 50 bucks for computation or something like that so this is like super cheap so this is quite amazing so of course the next question is let's do this in german because um um if you vote german at this network it doesn't sound that good Okay, so how can I get snippets in German? Okay, that's not too good because there's no open data sets in German. Only a few, and I couldn't really use them. So, okay, I have to fabricate them. How can I do this? So, task building a German corpus to learn with. Audio snippets, text from the audio snippets. Options are audio books or newspapers where people read you the newspapers. I said, okay, that's a requirement, but I have a requirement for humans to prepare the data set. So I asked my business partner, Zio, Zio, can I have like two working students for maybe two months to produce the snippets for me? Zio said, hi, Alex, I really like your talks, but no. So I had to come up with something smart. And this is the other thing. Neuronets don't solve all the problems. You have to think. So I thought, I started thinking. I said, okay, I can use, I have the newspaper, so I can, I scraped the text from the website of the newspaper. I have the audio version. Somebody read that. I used a cloud service to transcribe it. So this transcription has built a sense like a JSON and was a timestamp when actually the word was spoken. So I can use some heuristics and the word time index to produce audio snippets and the text from the snippet. This worked quite well, actually. So I produced a set. But there were also some assumptions because you listen to all the stuff. And even like you saw, like, okay, like Wes just like said in the keynote, Google is throwing billions at TensorFlow. And if a company throws millions at something, there's an agenda. And the same is with Google and cloud services and what they want to make you probably want to make you believe. The transcription services are really good. But how many transcription services are there available in German? Yes. Amazon, raise your hands. Amazon, German, oh, you're all so well-educated in this space. Okay, that's shortening a bit. I thought, okay, Google, I was doing it with Google. The transcription was okay, basically. It catched what was said, but it was not the word spoken. So it was okay, quality is okay, but I had to use a lot of heuristics to make the matches. And I thought, okay, maybe I can find another service even cheaper. So I checked Azure, IBM, AWS, offline Python labs, and Apple. And there's only Google for German. There's no service on Microsoft or IBM or AWS. And Siri is only if you use an iOS device. Okay, so that's not so simple, but I solved it. I had a newspaper article read by like 25 speakers. Okay, I think we all got that. And now let's first listen to the original. So, this is an original text taken from the newspaper. Deutschland darf kein Talent vergeuden, so heißt es immer wieder. Wo aber bleiben in unseren Schulen die Angebote für die besonders Talentierten? And now we're listening to the Neuronet train of the 14 hours. Deutschland darf kein Talent vergeuden, so heißt es immer wieder. Deutschland darf kein Talent vergeuden, so heißt es immer wieder. You see, but I think it starts blabbering. You see there's a rhythm. It sounds human, although, like, also scary. My partner was, like, scared. He was listening next to us. Oh, stop doing this, allies. I'm really scared. But after 10 days. Yeah, so it's picking up. You can even recognize the speaker, which was, I think, also, like, quite amazing. The data set is still not super perfectly cut, so it's working. Also, another side note, after 14 days, training longer is not always for the better. That's another thing you have to consider. But I think that was quite cool, all the challenges, lessons learned here. The assumptions were neural networks are hard to set up. Neural networks are hard to train. Neural networks are hard to understand, but neural networks are smart. This is when you look at stuff online. Neural networks are smart. Data is like the new oil. We have heard this for quite some time now. And there's genius people like Ray Kurzweil claiming singularity is coming in 2045. And it's not like a black hole. It's like AI is smarter than humans are. And let's check into that a bit. um first assumptions neural networks we have seen it after did i tell you about any trouble setting up the neural net my most the biggest troubles were actually that python code published on github from smart people actually and big companies so there's a found stuff from published by facebook um just recently something really hot everybody wanted to look at but i said hey, it's Python 2.7, guys. Python 2.7 is going to die next year. So why start anything new in Python 2.7, guys? You see a lot of Jupyter notebooks. You see a lot of closures, people using variables within functions, and the function is somewhere defined at the top of the notebook, so five pages later. So that's not really readable. It's also not transportable. I mean, okay, it works running the Jupyter notebook. it might be nice to prove something but it's not good good coding practice also like using input or list as variables because path and it's just you all know it's not a good practice but you see that a lot um okay so neural network was the network except for learning which networks are good for which use case this was all i knew had to know about new networks and then choosing parameters and stuff, there's always libraries to help you with that. So the network actually was not the problem. Unexpectedly because it's the magic, the black magic part from the outside. They are not hard to train because it all depends on the data. And I think they're also not smart if we look at this. For example, this is the Google Translate. And if you, these people who actually have the hobby of it, I'm not one of them, I just found this online. If you choose Somali, throw a lot of ak at there, it will translate a real sentence. And I think, but this is a very good example to understand how networks actually work, because they're not smart. They just, you just found a nice spot of correlations. They said, oh yeah, there's many aks, and ak is actually a word in Somali, and it comes to something like this. Also, where does this text come from? People were claiming, oh, Google, do you use the Bible to learn from us? No, we don't do that. But actually, I don't believe them because is this even documented on what Google Translate ever learned translating Somali into English? I don't know. Maybe. Maybe not. Because this is not going on since yesterday. It's like 10, 15 years. I don't know when Google Translate started. Networks are easy to fool. I found this, which is, I think, mostly entertaining. So here, you see, if you don't know the MNIST, who knows the MNIST data set? Okay, half of you. The MNIST data set is something you very often use in beginners' training. So basically, have a network learn how to read handwritten numbers. And something, Emilian, just wrote this blog post, and he just put a chicken here. And basically, the network was very sure the chicken is a five. So it was not, I don't know, maybe 10%, most likely a five. No, it was like more than 95% sure. It's 99.9% sure it's a five. So networks are easy to fool as well. Another nice example is here. On the left side, we see a panda. The network says, I'm 57% sure it's a panda. Then we put some noise in there, in the picture. and the result is 99.3% confidence this is a given okay and this is something we don't see as humans it's also like one of the big challenges if you talk to people building cars autonomous cars how can we prevent people fooling our autonomous systems with stuff like that or like something lights and stuff so because these networks don't work as we humans do they are very different the next assumption is data is the new oil and i would claim these people are rather right because oil is limited data is not limited it can be more data than atoms so um i put two here you can look into it i think you should also like rather think it's you know the new nuclear power without the waste um hopefully um yeah okay and then if you look at how much what's what what was the expectation if you start with neural nets the expectations like that yeah we have the data anywhere that they've easy okay we build some pipelines then we have a huge network problem so you see here this is like the expectations many people have on how much resources are required to get the network to do something with deep learning. So I get the data, data's around anyway. We have some pipelines. Oh, that's a big problem. And then applying and serving the models. Once we have the models, it's easy to serve. And I want to go, but this is like a good machine learning development example, but we'll do it later. We'll show it on time. Reality shows, and also everything I did was, what was the biggest problem? It was always like having nice, clean data and having them system. The network, actually, I can just take it from a research paper because there's, of course, building a network, experimenting with networks is also a lot of work. But we don't have to look into that because there's many great scientists publishing their work and we can just basically build and continue what they produce and openly make available to us. And I'm really grateful for that. But for me, from the perspective, building something with deep learning, the network was never a real problem. But serving models, updating in production is a problem in many people. And there's a lot of talks about this as well. So the perspective is totally different. It's not the magic AI network, which is the challenge. It's the data mostly and having models in production. Apart, there's also other problems or things you have to consider. Ethics, because once we have something in a network, you can also look into, okay, is it discriminating anyone? Is it not only from humans, but also are there some assumptions which are just self-fulfilling because the network has learned on data? There's also a lot of talks on that. Just go to pyvideo.org. There's a lot of talks about this on Python Conference. I think a very good read is this from Rachel Thomas from Fast.ai. And she's writing what do machine learning practitioners actually do. So she wrote the post. What does it work? It's like a free hard blog post. It's like maybe one hour read. And I think this is really valuable input. It's also like a little bit opinionated because she also says, okay, why is the CEO of Google, like in the Google conference announcing, yeah, TensorFlow, deep learnings, all this. Yeah, we have to also think about it. Google also like self-cloud computing. So stuff like that. There's a really good read. I have, this is like the German version. So German readers read fit real quick. There's an English version coming. It's an interview with Margaret Bowden, which I can really recommend. It was in the Zeit magazine recently. Margaret Bowden is actually, she studied psychology for 50 years ago. And she started for experiments to look into the AI space. And she was giving interviews. She wrote a lot of books on the whole thing. And this is like the translation. And she said, okay, many think there's like this superhuman artificial intelligence is near and coming. For example, Ray Kurzweil. Do you believe them? And she says, I ask him, I value him. They know each other. They do all of this for 40 years. They're like this. And he really believes it, she says. And what do you think? I think this is all, I mean, I really value his work. He's a genius. She says, I would not label many people as a genius. But he's just like crazy, and the prediction is crazy. And this is basically what all the media buzz is about. And this is also I have to point out because this is also something if you're in the AI and deep learning space, you have to point it out because there's a lot of stuff like this going on. I really had to think about what to do about this tweet from Siemens. Because it's like a sponsored tweet and you see like a nice movie, there's Uli. Uli is working with his team at the Vitualienmarkt in Munich, like center city, super top spot in Munich. He's working there and teaches. And with his team, he teaches machines to think. I said, well, no. Who came up with that? And I really thought, should I comment, should I not? So I came up with this. I said, machines don't work. They don't think. Sorry. Machines, yeah, work, but they don't think. Maybe Uli wants to explain this to his Siemens colleagues in marketing. And actually, Uli saw this tweet. He replied. He said, yeah, you're right. um we do software and we implement this but this is like the one of the biggest issues in the public perception that ai is not an android it's not human-like it's super specialists on very specific tasks as something to consider um maybe this is not easy to sell to a customer but actually when we talk to customers we always have to say okay yeah you know we have read all this great stuff and there's like all the ideas we have in our mind from Hollywood and movies and marketing and people who don't have a clue about AI and deep learning making all these assumptions and putting all these pictures in your head and usually it takes some time to really guess okay this is what you can do this can there you can apply and you don't really have to worry about your workers getting unemployed because maybe they will just move to a different space and we have some awesome tools so the summaries Are neural networks great and useful? Of course they are, but we have to give it the right perspective. So this is an ongoing talk, so there's going to be constant updates, but most of the conference talks will be published. I'm quite surprised we even have five minutes for Q&A left. And if you have, yeah, thank you, five minutes for coming in. If you have any questions, we also have a booth here. Clinics Week is one of the community sponsors of this conference. Just come talk to me. There's a hiring. And if you want to have a comic version of your picture or any picture, I can also happily provide it to you if you want. Okay. Thank you very much. Thank you very much, Alexander, for this great talk. We see pandas and noise are given. so first question thanks Alex do you know the Prisma photo editor Prisma photo editor yeah I know it and the other one is also for videos already I think it's Aristo or yeah there's some apps who use this technology same technology what do you think about this you can comment on this actually I've never tried them in depth but I know friends who use them so yeah it's there I mean it's for a while the style transfer paper I think it was like two and a half years ago I think what I probably didn't point out today is like how to judge which is good or not because we were quite impressed probably from the fabricated voices which were not production ready but easy to say but we also have and also like the text actually artwork is way easier to fabricate because there is no definition I can tell you this is the wrong German sentence with grammars and stuff so weird, you have to be way more accurate, and basically I can throw any painting, or many paintings which look like something, to you and say, yeah, yeah, it's art, I mean, the space is open. More questions? Thank you for the talk. My question is how much do you think of the style transfer is actually cherry-picking? What again? How much is cherry-picked? How many of the examples we see? I cannot really, I didn't count but i just saw like a lot of papers on github and i'm just i mean i mean my intuition just like alarmed me you see you always see like the same examples so of course i mean you have to look a lot of stuff is published by students who put some work in it and probably they just have some simple goals maybe they just want to share their work um they would just want to see this works in another library so i don't want to really judge them for doing this because still they provides something and they're to free knowledge but the thing is and it's also like more like a general problem I also see working on the conferences it's way harder to make people talk about problems and stuff this is not working for me this was also like a big issues um because I had these results especially in the tech space luckily I know a lot of experts who can I can ask Kate is these are my results and I said yeah yeah that's okay um even people working full-time have similar thoughts, especially in the German space for the text. So I can, if you work in the space, I just can encourage you, in any space, also talk about stuff that is not working because it's usually more helpful than the glamorous stuff for your people listening to your talks. More questions? So I do have one question. No, there's no There's no GUI. No, that's no problem. I was wondering about the eGPU you mentioned in the beginning. How hard is it to set one up? Is it worth it? Actually, my main problem actually was PyTorch. Because PyTorch and eGPU, you have to compile it on the Mac, and that was the hardest thing. For example, if you use a Linux machine, you have a wheel. You can just install it. On a Mac, it was a little problem. Also, we have to be aware of the commits. Like, PyTorch is moving all the time, but the commits go to the master. And I had versions compiling, and then I did an update, got the newest versions, and stuff was failing. Because also, PyTorch is a very complex project. And some other libraries, I think Cafe, were actually not compiling. So it was not actually PyTorch, but something PyTorch uses made the trouble. So I checked the right commits. But I'm not sure about, because PyTorch 1 beta is released now. I think maybe even on the Mac you can get a compiled version. I know it is on the agenda, but if you want to hear more about the latest stuff in PyTorch, Adam Patzke is speaking here at 2 o'clock. He's one of the co-developers of PyTorch. He can tell you all about it. Okay, great. Thank you very much, Jan. Thank you.