L18.6: A DCGAN for Generating Face Images in PyTorch -- Code Example

Sebastian's books: sebastianrasch...
Slides: sebastianrasch...
Code: github.com/ras...
-------
This video is part of my Introduction of Deep Learning course.
Next video: • L19.0 RNNs & Transform...
The complete playlist: • Intro to Deep Learning...
A handy overview page with links to the materials: sebastianrasch...
-------
If you want to be notified about future videos, please consider subscribing to my channel: / sebastianraschka

[@ryanjeon3620]

Great video! Subscribed and following. Question- What are your thoughts on using a DCGAN to generate synthetic images of rare classes in an unbalanced computer vision training dataset?

[@kafaayari]

As always, great lecture Mr. Raschka! I have a question that you may have an idea. Suppose I want to generate some images of faces for a specific race but I have a limited dataset for it (let's say only 100 face images from this race) Therefore, can we use a pretrained encoder, decoder trained on a large general image dataset (or encoder-decoder trained on a large face dataset) and utilize transfer learning to be used for GAN for my purpose?

[@SebastianRaschka]

Huh, that's a good question. I actually never tried transfer-learning for GANs. I think it should generally work though; most StyleGAN projects are based on transfer learning from bigger datasets I think.

[@kafaayari]

@@SebastianRaschka I understand and will check them. Thank you very much professor.

[@deepexplorationcode9456]

Hi Sebastian, First thanks for the Great videos, I have learned a lot. I have also now understand how to make and train a DCGAN model in PyTorch. I would like to ask you two questions. My first question: According to the code that you put in github, you divided the dataset in three parts: Training, validation and testing. What is the percentage you take for each part compared to the complete dataset, and how you divided it. My second question: Is it possible to make a video on how to train an instance segmentation from scratch in PyTorch, because until now I haven't found any tutorials. For example a simple example by using centernet model or centermask model.

[@SebastianRaschka]

Thanks, I am glad to hear you got something useful out of these videos! Regarding the train, validation, test splits, there is no universal solution. Usually, you want to have the training set as large as possible, but you also want to have a large test set so that the performance estimates are accurate. Also, you don't want the validation set to be too small, because then then that performance estimate during training and tuning becomes too unreliable. 70/10/20, 75/5/20, 80/5/15, etc. could all be legit choices Regarding the video suggestion: that's an interesting one and I will keep it in mind. Unfortunately, it's currently so busy that I will probably not get to it in the short term, but I will note it down!

[@deepexplorationcode9456]

@@SebastianRaschka Thanks a lot for your answer, and I hope that in the future you will make other great videos about Deep learning by using pytorch.

[@zigzag4273]

Hey Sebastian. Hope you're well. I've got a question. How would you read or advise your student to go through a technical/tech book (like the ones you recommended for this course). I ask this cause I'm currently like at L15. So I'm kinda struggling to keep up cause of all the new info and trying to do some supplementary reading from the books and reading the books kinda take some time to understand.

[@SebastianRaschka]

I think there is no magic bullet, unfortunately. Often, there is some struggle involved when working through a textbook. This can be good though, in terms of getting the best bang for the buck (shows that the info is not redundant) but yeah, it can be frustrating. Personally, when I read papers or books, I look at the figures and main equations first to get a feeling of what is being covered. Then, I read through it and take notes about things that are new or I don't understand (but I am not trying to summarize everything). Then, after reading I organize my notes and go back to passages that were not clear. Sometimes, I need to re-read the whole text to get a better understanding.

[@mikhaeldito]

Dear Dr. Raschka, Thank you for sharing the videos and notebooks for this course. Are you planning to share the assignments as well? That would really help to measure my learning. Thanks in advance.

[@SebastianRaschka]

Sorry, I may adapt and reuse certain assignments in the future and don't want to share them online for that reason.