Generative Adversarial Networks for Style Transfer (LIVE)

Generative Adversarial Networks for Style Transfer (LIVE)


Author: Kevin Mason

55 thoughts on “Generative Adversarial Networks for Style Transfer (LIVE)

  1. Please do a video showing how to do this: https://www.theverge.com/2016/3/21/11275462/facial-transfer-donald-trump-george-bush-video You could go one step further and change the voice too! xD

  2. Can GANs be used to create a 8 bit pixel avatar given the input of your photo? Are DiscoGANs, the way to go for this? Or any better suggestions? Apart from this, could you think of any GAN applications for game content?

  3. Could you please make a video how to install python 3 + cv2 + tenserflow all in one I just dont get it to work

  4. Could you also use keras for this type of things, by which I mean building completely novel NN structures?

  5. I get this error while running through the code. Any ideas? (Maybe it needs Pytorch and I cannot install it under Windows?)
    22 #gausian mixture model for generating data
    > 23 from data_gmm import GMM_distribution, sample_GMM, plot_GMM
    24 #analyzing data
    25 from data_utils import shuffle, iter_data

    ImportError: No module named 'data_gmm'

  6. I kept getting error for "matplotlib" and "os" the i tried to import em, even though i tested this for both anaconda2 and anaconda3 help please!

  7. "Stop reading textbooks. Textbooks should read themselves to us." Siraj you are awesome! Siraj continues …
    "Let me tell you why you're here. You're here because you know something. What you know you can't explain, but you feel it. You've felt it your entire life, that there's something wrong with the world. You don't know what it is, but it's there, like a splinter in your mind, driving you mad. It is this feeling that has brought you to me. Do you know what I'm talking about?"

  8. Great video ! But maybe you should try to focus a little more on what it is that we want to show with this experiment. I think that it is very clear what the paper does : transferring a class A object's style to an object of class B. However, with the Gaussian Mixture Model, it is not very clear what we really want to do. The consequence of that is that it is hard to understand and interpret the result. Even though after 1h of live stream you might be tired and it can be hard to stay focused, I think you should take the time to interpret the results. Not the first time someone tells you that, just sayin' ! Great work anyway, thanks.

  9. Hej Siraj! Love your content buddy! Quick question. Can you use GAN in anyway for analyzing market data, like prediction or stock movement?

  10. Hi Siraj, I got inspired by your videos and started learning machine learning. As you always say having Cheat Sheet is very helpful so I developed this android app "Code Cheat Sheets" Link :
    https://play.google.com/store/apps/details?id=com.radioboxlabs.codescheatsheets&hl=en
    which contains cheet sheets of many different programming languages and libraries such as R,Python,Nodejs and libraries Numpy,Matplotlib,Scikit-Learn,Pandas,Keras,Scipy. I am sharing it here because it might be helpful to others.
    Siraj you rock man. Keep making great stuffs.Thanks

  11. Hello Siraj,

    First of all thank you very much for starting youtube channel and sharing cutting edge tech (Machine Learning-AI) stuffs and teaching lots of us.

    Although recently I was trying to do some hobby project (Volume of Liquid present inside Liquor Bottle) for this I wanted to use ML and TensorFlow and thought of identifying Bottle Shape (Cylindric, Cuboid, 3D Hexagon, etc.) and later using some simple Image Processing computing Volume based on the Height Width of bottle. Although I found that this kind of approach is bit unstable and not so accurate. Can you please suggest some Model for solving such problem.
    Note that in order recognise 3D Shape I used inception-v3 model and retrain it with the pre-classified dataset of Cylindric, Cuboid, Hexagon Bottles.

    Thank you in advance. Any help in this would be appretiated.

    Thanks,
    Jinay

  12. Daum man! You moving up! I remember when i subbed to you not long ago, you had under 10k views 🙂 been sharing the channel with everyone i know 🙂

  13. I would love to see someone apply this to the Pokemon Go dataset on Kaggle to generate new Pokemon characters.

    https://www.kaggle.com/abcsds/pokemongo

  14. 21:31 Sigma is the standard deviation. sqrt(variance) therefore 22:08 it was actually the variance. sigma squared is variance. I am a statistics student, trust me 😀

  15. Don't look at your eye crust in public. Don't groom yourself in public.
    You're less manic in this video. Thank you.

  16. Variable inference/Repeat/fully_connected_1/weights already exists, disallowed. Did you mean to set reuse=True in VarScope?
    How do i fix this error please help me out

  17. Is there any GAN I can play with? and give input, and receive images of my description, for example "Erase the human from the picture"

  18. I was attempting to run the demo IPython notebook, but I ran into this issue:

    ValueError: Variable inference/Repeat/fully_connected_1/weights already exists, disallowed. Did you mean to set reuse=True in VarScope? Originally defined at:

    File "c:usersuserappdatalocalprogramspythonpython35libsite-packagestensorflowpythonframeworkops.py", line 1269, in _init_
    self._traceback = _extract_stack()
    File "c:usersuserappdatalocalprogramspythonpython35libsite-packagestensorflowpythonframeworkops.py", line 2506, in create_op
    original_op=self._default_original_op, op_def=op_def)
    File "c:usersuserappdatalocalprogramspythonpython35libsite-packagestensorflowpythonframeworkop_def_library.py", line 767, in apply_op
    op_def=op_def)

    FROM:

    ValueError Traceback (most recent call last)
    <ipython-input-14-0f0e7a66854c> in <module>()
    9
    10 #2 more generators (decoders)
    —> 11 rec_z = inference_network(p_x, latent_dim, n_layer_inf, n_hidden_inf, eps_dim )
    12 rec_x = generative_network(q_z, input_dim , n_layer_gen, n_hidden_gen, eps_dim )
    13

    <ipython-input-7-2f1f7b2e4e74> in inference_network(x, latent_dim, n_layer, n_hidden, eps_dim)
    25 with tf.variable_scope("inference"):
    26 h = x
    —> 27 h = slim.repeat(h, n_layer, slim.fully_connected, n_hidden, activation_fn=tf.nn.relu)
    28 z = slim.fully_connected(h, latent_dim, activation_fn=None, scope="q_z")
    29 return z

    I am running this on Windows 10 using Python 3 and I have all of my dependencies installed, including the files from ChunyuanLI/DiscoGAN. Thank you!

  19. I'd like to ask what's the detailed idea of style transfer between Xdata and Zdata in this example?
    I've watched this video and downloaded the code, but I haven't got the clue yet.(I've run the code but it only gives me the initial pdf files of X and Z)

  20. 4:46 Does evil exist, and if so, can one detect and measure it? Rhetorical question, Morty. The answer's yes, you just have to be a genius.

  21. Best stuff ive ever seen!! so i write my graduation work on gans for fashion and feel so frustrated about reading all this papers. In ur video i found a lot of answers about research pipeline in this area. Thank you dude for that stuff!!!!

  22. Like you're content, about to start binge watching hope it helps me make this model. Let's say I want to use the same model to generate a rap style model, bit problematic I know to get "the whole" to make sense but bare with me.

    So I first remove chorus from some lyrics I same from various songs for a rapper. Reason being a chorus to me is a unique feature of a song so you could in general to train a model to generate a good chorus… anyway encode the text, possibly also add sentiment analysis to keep to the mood of the song and then have it only return a verse of a song as an output.
    Whether it makes sense or not have edits passed back to be learnt/fit from and also have the discriminator learn we don't like this kind of output we want more like this.

    Keep this cycle of refitting till it makes something sensible. It sounds hard but honestly do rappers say anything sensible these days 🙂 come on surely we can make one as good then slap on a dope beat and boom you got a nice song.

    Even for beats from my perspective instead of audio analysis turn it to a music sheet and have it know how beats tend to go and i'm not a major in music but it would i guess be categorized for different instruments in that sense then have it encode music sheets that a discriminator could then wonder "is this a beat from a dope ass song? I wonder… might be, might be" then decode that to actual beats… have some python script then that can read music sheets to beats and then hopefully tadaa you got something.

    Doing this as my research for getting good with machine learning so feel free to use the idea freely but if you can include me in the development process, I'm good with python and I know tend to know how machine learning models works and have implemented a few simple ones.

  23. I'm lost. I know linear algebra and calculus and some Python, Javascript etc.., so I need some roadmap to reach the content of this video because the complexity of its content is overwhelming. From choosing a TensorFlow keras caffe and many other frameworks, to "omg I have AMD graphics card I cannot use it". Or for example I install Docker, then install TF container. Well there is actually a Jupyter Notebook, where are the console, and why is a jupiter notebook? and I DO NOT understand what the heck is happening on that virtual machine or what is supposed to do there or if is it compatible with my graphic card…. LOST. Which machine learning roadmap should I take to learn it with a daily exercises?

Leave a Reply

Your email address will not be published. Required fields are marked *