Note: I had previously done a video on DCGAN but with this GAN playlist I had several things I wanted to make better in that tutorial. This video should be better and also fit better to this playlist :) Also if you have recommendations on GANs you think would make this into an even better resource for people wanting to learn about GANs let me know in the comments below and I'll try to do it! I learned a lot and was inspired to make these GAN videos by the GAN specialization on coursera which I recommend. Below you'll find both affiliate and non-affiliate links, the pricing for you is the same but a small commission goes back to the channel if you buy it through the affiliate link. affiliate: bit.ly/2OECviQ non-affiliate: bit.ly/3bvr9qy Timestamps: 0:00 - Introduction 0:26 - Quick Paper Recap 4:31 - Implementation of Discriminator 9:38 - Implementation of Generator 15:27 - Weight initialization and test model 19:09 - Setup of training 31:36 - Training on MNIST 32:20 - Modifications to CelebA dataset 33:52 - Training on CelebA and ending
21:22 - Please, use transforms.Resize((IMAGE_SIZE, IMAGE_SIZE)) instead of transforms.Resize(IMAGE_SIZE) because transforms.Resize(IMAGE_SIZE) resizes image PROPORTIONALY. For example, celebrity image (3, 178, 218) is resized to (3, 72, 64) by transforms.Resize(IMAGE_SIZE). This doesn't influence MNIST, because its images are squares. However, for CelebA dataset DCGAN shows no errors, but doesn't converge.
Hi, I need to know why proportionally changing the image harms the training? intuitively it looks like that resizing while keeping the ratio should work better.
Great video. Can wait for the next video in the series. Can you please also create videos relating to the matrix for evaluating GANs like Inception Score?
I actually used the DCGAN you implemented in your previous video as a foundation for some work I was doing. I haven't watched this video yet (I plan to) but I wanted to ask how big is the improvement between this one and the older one? If it is pretty significant I would look into refactoring my current implementation.
Not much difference, so if you followed the previous tutorial it's mostly the same, and there's not much gain in watching this one. I felt I wanted to clarify some explanations and improve the video quality mostly and adapt it better to the GAN playlist. For the actual changes/differences (none are major): 1. I didn't use soft labels in the new video (didn't have to) 2. Cleaned up some code particularly for the model architectures 3. Used weight initialization similar to what they did in the paper, i.e mean 0 and std 0.02 for all weights (skipped this in previous) 4. Used bias=False for conv layers since we used BatchNorm they are spurious parameters 5. Trained both on MNIST and CelebA dataset Next video I will implement WGAN and WGAN-GP which is probably going to be a much more useful to watch. My goal in future videos for the playlist is to implement more advanced architectures so that we can reach closer to sota performance
I had a doubt, in the discriminator, we are taking the features_d as 64, and in the first layer, we map the channel_img(=3) to feature_d(=64), but in the actual diagram of generator, we map channel_img(=3) to 128 as the discriminator is just the ooposite of generator. So in the first layer of discriminator, shouldn't it be features_d*2 instead of features_d?
Thanks for video, may I ask something? When I trained the generator with celebA dataset and monitored fake images on Tensorboard. I got fake images as color noise that is kind of different as you got real fake face images. do you know why? thx in advance
I cropped the images and it worked. I hope this solves your problem: transforms = transforms.Compose( [ transforms.Resize(IMAGE_SIZE), transforms.CenterCrop(IMAGE_SIZE), # THIS LINE IS ADDED transforms.ToTensor(), transforms.Normalize( [0.5 for _ in range(CHANNELS_IMG)], [0.5 for _ in range(CHANNELS_IMG)]), ] ) Another version of this model: pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html
@@ahmetburakyldrm8518 You have to crop the image to be square. So add transforms.CenterCrop(IMAGE_SIZE) Or use transforms.Resize((IMAGE_SIZE, IMAGE_SIZE)) instead of transforms.Resize(IMAGE_SIZE), because transforms.Resize(IMAGE_SIZE) resizes images PROPORTIONALLY.
how can you save the generated fake images to a separate folder during training rather than displaying them in a grid, so i can access each one individually?
Not working as per GitHub code you uploaded for MNIST dataset. Discriminator getting strong(as loss remains 0 after Epoch [0/5] Batch 100/469. How to fix it? Thanks Edited: It works after uncomment the BN. But I wonder why some of the steps showing blank black images? That weird. It did works when training going on and opening the tensorboard to observe simultaneously
This fixed it for me, uncommenting 'nn.BatchNorm2d(out_channels),' in both the Discriminator and Generator block methods (model.py). Also fixed the 'transforms.Resize((IMAGE_SIZE, IMAGE_SIZE))' line, as mentioned below (train.py). Thanks!
Hello Aladdin! As usual amazing videos. Have been reading a few blogs on the same topic. I see this statement "train the Discriminator first and then the generator" makes sense since you don't want the Discriminator to strong. What I fail to get it is in code, where do we find that. Per epoch both the Generator and Discriminator get trained in the same loop. I hope I am clear. Do people mean for an epoch do Discriminatior first and then Generator?
Hey Sai thanks for the kind words! I have not seen people train the Discriminator and Generator separately in the way you describe. What I think they could be referring to in the article you read is that for architectures like WGAN they want to train the Discriminator more (5 update steps vs 1 update step). This has the intuition that the Discriminator is leading the Generator, because it needs to have some understanding of what a real image actually looks like in order to provide valuable feedback. In more state of the art architectures like ProGAN, StyleGAN they kind of move away from these "tricks" and provide a new way of training GANs. I don't think exactly how many update steps you train the Generator vs Discriminator matters too much. Will make videos on those in the future:)
Hello Aladdin! First of all, thank you for the amazing content! This is the best resource on ML that I have found on the internet. I cloned your repo and ran the train.py wo changing anything and I am having a problem: the D loss goes quickly to zero and the G loss grows at the same rate (already in the first epoch). What is the problem here? And what things could I try to solve this? Thank you for this content again!
Thank you so much for the help! I need to save my generated images for a project and was wondering if you had any code or recommendations for saving the images individually?
Not sure if this is still needed for you, you'll need to import save_image and then in the if statement where we create the grid of images you'll add a couple of lines using the sav_image function to write the image to a png or jpeg, should look something like this: #import statement from torchvision.utils import save_image #at the end of the if statement generating the grid of images place this (or similar code) save_gen_img = fake[0] save_image(save_gen_img, "images/%d.png" % batch_idx, normalize=True) # note that you will need to create a new folder in the project called images to have it write properly using this.
Thanks for the video, I have two questions after watching it. How can I generate multiple fake images at once and save them? For different size datasets, it seems that you start training without changing other parts of the network, is this correct?? Looking forward to your answer!
Thanks for this video, I tried running the code but I got this error: File "C:\Python39\lib\site-packages\torch n\modules\module.py", line 1130, in __getattr__ raise AttributeError("'{}' object has no attribute '{}'".format( AttributeError: 'Generator' object has no attribute 'parameter'
I didn't understand what that super(Generator, self).__init__() is doing, super().__init__() is basically used to initialize the parent class but why have we given it these arguments?
Very good. Thanks for your useful videos. I want to train a cycle gan but my colab session was crashed just at first batch of data. Do you have any idea for its reason?
Thank you very much. I just had the CELEBA version running. It is awesome! I work on Google Colab and faced a bunch of problems with the Dataset, which I eventually solved by generating 4 zip files with a size of approx. 50K pics in them and read them with an IterativeDataloader inspired by the web page medium.com/speechmatics/how-to-build-a-streaming-dataloader-with-pytorch-a66dd891d9dd of David McLeod. Why 4 zip files? I found if they are bigger than 2^16, they span several disks and can not be handled by the python zipfile library. Additionally I took over some code lines from the Pytorch DCGAN tutorial (eg. model initialization, transforms.CenterCrop) to get convergence. I would not wonder, if this need came up because I had first introduced some mistakes into the code. Maybe, these informations help others. I have learnt a huge amount!!!
Thanks for the details and not copying code in chunks. Please keep it up! Also, please note that the target dataset is MNIST, not CelebA at the beginning as I was confused with the 64x64 dimension of the generator/discriminator. Also, I think it would be better to use one writer for real and fake image, I'm not sure what are the reasons behind using two separate writers.
Thank you for this awesome video! I have a problem that if I want to save every generated fake image (the image of each grid) individually, how can I do it?
Doesn't line 68 loss_disc.backwards() cause the generator to train at the same time unless you freeze the weights? I was under the impression the generator is not supposed to train at the same time as the discriminator because then you are making the gen worse to improve the disc.
I think there is a problem with the interpretation of figure 13:46. If noise is a 100-dimensional vector it should be drawn as a 1x1 square with 100 channels depth. Now as seen in figure, I would translate it as a 1 (x),100(y),z(1) vector (VECTOR OF ONE CHANNEL AND 100 FEATURES) . So in 11:59, you reject the figure You should reshape the 1,1,1,100 to 1,100,1,1 to work as it says to figure! Another question is why the discriminator does not have Dense layers?
Trying to copy and paste the code here but I'm getting the error that there is 'No Module named model'. I have PyTorch installed and to my knowledge this error should not be happening? Is the model being used here from PyTorch?
Thanks much for the video .... I am trying to adopt this implementation to text classification using SGANs...I would appreciate your help in this regard or further lead to a similar implementation for text classification.
The real images input to discriminator (from dataloader) still has range from 0 to 1 right ? We did not change it to [-1, 1] before feeding it to discriminator even though the generator will output value between [-1, 1]
When we are applying transform to our dataset using transform = transforms (where transforms is our custom transform) we are normalizing our input data between -1 and 1 because transforms has a step of transform.normalize(). Hope this helps!
When you print the results on tensorboard for every step..is it runs automaticaly?...I mean that when i start training and open tensorboard i need to click refresh button on browser tab to see the next step progress.
Hey Aladdin, great video. I had a quick question, in the fixed noise for the generator, what is the 32 for? Initially I presumed that each image in the batch would receive a 100x1x1 noise tensor for generation, a.k.a. fixed_noise = (batch_size, noise_dim, 1, 1). If the question is a little confusing, I can definitely elaborate more. Thanks.
I think I may have found the answer to the question, but I believe we are showing 32 images on Tensorboard for each epoch, so 32 is the image count for that (testing). Hopefully this is correct.
Awesome video bro, one more suggestion... you can use labels that comes with celeba dataset and implement cgan, the training in cgan is much more faster and produces better results... And after sufficient training, since the generator is conditional, so you could generate the girl of ur dream 😁
Thanks for the good video, but I have a question, why the generated numbers are not equal to the real numbers even though they definitely can be identified as numbers.
In a GAN, the generated images are supposed to match the distribution of the real images. In a autoencoder, the output looks just like the input, but in a GAN we are just making something that looks like it could have been part of the dataset. Does that help at all?
26:01 - Training GANS is not unsupervised. When you use the BCE criterion to calculate loss for the Generator and Discriminator, you pass in the torch.zeros_like() and torch.ones_like() as the labels to the disc_real and disc_fake predictions. We can use the labels from the Dataloader instead but we do need labels to calculate the loss and optimize the weights i.e. supervised learning.
I think you have a good point, maybe a better term is self-supervised. I'm honestly not sure if the distinction between unsupervised and self-supervised is perfect either but you're right that it's not unsupervised by definition. With GANs you do not need a labelled dataset, you just need a dataset and the labels are inferred by themselves. Do you find it better to use self-supervised in this context?
@@AladdinPersson You are right GANs don't specifically need a labelled dataset, just a set of labels you provide to both criterion. By that definition I think self supervised can be an appropriate term. I'd also want to thank you here for all your tutorials. They have really helped me not just learn pytorch but get the proper intuition behind Machine Learning approaches. Keep it up!
"Unsupervised learning is a type of machine learning in which the algorithm is not provided with any pre-assigned labels or scores for the training data." - from wikipedia It's not that you can't have labels in unsupervised learning, you just can't have them ahead of time.
@@AladdinPersson There is more than one definition of self-supervised learning... but GANs (in general) don't qualify against any definition I'm aware of. Some people restrict self-supervised learning to be only robotics tasks. Yann LeCun uses the phrase to talk about autoencoders as well. In either case, we are looking for situations where labels are obtained by hiding a bit of the data (or making use of naturally hidden data). So in NLP, we could hide one word from a sentence, and then discovering that hidden word from context is the task... In robotics, often the hidden information is contained in the physics of the world... So based on a picture, how soft is the object in the picture? Then check softness with a robotic hand to see how to update the weights. In a autoencoder (to be fair, you can train a pix2pix to do this as well) automatic colorization is a great example of self-supervised learning.
