How Are Gans Used in Fashion

Using GAN for manner

How to utilize GAN to improve way industry ?

How does GAN learn design

INTRODUCTION

Hullo anybody! If you opened this commodity, probably, y'all are interested in such topics like: fashion, motorcar learning, deep learning, artificial intelligence and customization.

Present each internet or real life user wants to get a production specifically designed for his or her needs and wants. How better fashion industry and move it to the "full customized" land ? One of the ideas is using GAN to generate apparel past its clarification.

So, permit's begin!

CONTENT

1. Introduction
2. GAN understanding
3. GAN for fashion
4. Conclusion

GAN agreement

DCGAN construction

On the picture above is a construction of such type of GAN equally DCGAN or deep convolutional generative adversarial network.

How does it piece of work ?

First of all, we need to define a task. Our task is to generate images past text descriptions. It means we will have some encoded input text and we will expect an epitome. You lot may retrieve why don't we apply just transposed convolution or some kind of autoencoder with modified input shape ? Of grade we tin, at that place is no limit for imagination. But our task will be much more circuitous considering nosotros will need to compute loss for output paradigm and existent image(MSE for instance). It's harder to brand than understand GAN principle.

Permit'south brand it unproblematic. How to understand GAN architecture and training idea ?

On the in a higher place image is shown an analogy with real life example. Imagine a situation: there is a criminal who wants to make faux money of really proficient quality and there is a law officer with an aim to distinguish real and false money. In our globe in 90's information technology was a kind of popular idea.

So, how do they interact ? I volition make this story closer to ML point of view. Let's write information technology footstep by stride.

Plan of activity for initial interaction:

1. Counterfeiter take read some material (generator input noise/encoded instruction)
2. Counterfeiter made a false nib (generator output)
3. Police officer received this bill and example of a existent ane(discriminator ii inputs)
4. Police officer decides information technology'due south a faux (discriminator output = 0)

It was just a forrard pass. How to make generator and discriminator be cleverer ?

For instance counterfeiter made this :)

Outset of all we need to evaluate discriminator (police officer). Officeholder needs to make two things: know how do the real money look similar (real loss) and be skilful in distinguishing false ones (fake loss). In our case police officeholder mark (or may exist some daily bonus for crime identifying) will be calculated like: (Loss (Real money|Officer decision) + Loss(Simulated money|Officer decision)) * ane/ii .

Now some math staff and we volition continue our story.

We will use BCE Loss (binary cross entropy) with the general formula on a picture below.

Cross entropy in general view

Let's imagine law officeholder may tell us a probability of is it a real or false (0–100% or from 0 to ane). And officer tells u.s.a.: real money: 0.ix real and 0.i imitation (with ground truth i real and 0 false) and fake coin 0.2 existent and 0.8 simulated (with ground truth 0 real and 1 faux). Now let'south calculate the loss.

Police officer loss = 1/2 * ((-log 0.9 * i-log 0.i * 0)+(-log 0.2 * 0— log 0.8 * 1)) = 1/2 * (0.04 + 0.09) = 0.065 .

And it's kind of a practiced loss! Groovy, officer!

Merely what to do with counterfeiter ? He/she also needs to be good in making fake coin (in real life of course not, just we are mathematicians and may make any analogies for better understanding). In our case nosotros even want to make counterfeiter really skillful in making simulated money! Let'south assistance him/her !

We will tell counterfeiter the result of the expertise of the police officer (loss) and will close the eyes of the officer for some fourth dimension (discriminator weights freezing).

Now nosotros will learn counterfeiter (generator). We flip the labels and make generator fake data every bit real one and nosotros will employ frozen discriminator to piece of work for generator purpose ! Kind of fob.

And so now we train counterfeiter to make better coin. For example the event of counterfeiter is on a movie beneath.

Generator loss = -log 0.1 * i-log 0.9 * 0 = 1 (information technology'due south a kind huge loss and we tell it to the generator and update information technology's weights). And side by side deportment repeat with forrad pass as described earlier.

The plan of actions for grooming (metaphoric):

1. Evaluate police officer power to know real money (real loss with existent data)
2. Evaluate police officeholder ability to distinguish imitation money (fake loss with imitation data)
3. Summate medium mistake in evaluation for the officer (general loss as a one-half sum of real and fake losses)
4. Amend officer and counterfeiter noesis with new general evaluation upshot (dorsum propagation for the whole model)
5. Close the eyes of the officer (freeze discriminator model)
6. Make counterfeiter think he/she does real coin by using "blind" officer expertise (flip labels and calculate generator loss BCE and update generator)
7. Open the eyes, officeholder! Counterfeiter now can exercise better fake bills! (unfreeze discriminator and brand forwards pass again)

Returning to the machine learning. Nosotros are interested to make generator exist awesome, yes ? We don't need a cool discri minator every bit our work result (in the nigh of the cases). We demand to brand fake information!

Now let'southward work with the fashion example.

GAN for manner

How exercise the results wait like ?

On the images above are shown examples when trained GAN has an input as a text and the result is an image. It is a kind of prototype with 128x128x3 images with low quality.

To make this project existent were taken 15 GB of images with corresponding text description for each one (gender, fashion, color, description etc.)

Images data was trimmed to 128x128x3 (RGB) images and text data was transformed with TFIDF technique to 256 values vectors.

In some cases model guesses the main idea like colour or men/women/unisex blazon. What virtually some new data (text description) ?

Some kind of a horror movie. We encounter black and carmine colors, some kind of human standing (may exist man) and information technology's all.

GANs are extremely hungry for data and slowly in training. Exactly generator is a slowly learner and to make this model information technology was spent 72 hours with GeForce GTX 1070 Ti for 300000 epochs and 12000 images with corresponding text captures.

What to do next ? DCGAN has its ain limit. There is no better results than with 128x128x3 pictures. Model is needed to be more complex.

In this example is ameliorate to utilise ProGAN or progressive growing generative adversarial network. About it — in the next commodity.

CONCLUSION

GAN — is a practiced technology to make neural networks be able to "imagine" similar humans can do similar somebody hear "bluish jeans" in imagination appears some kind of images and it'southward not like database search, it's always a process of creation and fantasying.

Only architecture must be really complex to make images in a skilful resolution. For example in ProGAN the networks is training footstep by step from images 4x4x3 with no sense to images 1024x1024x3 (RGB) with good quality images and deep sense.

Thank you for reading! I hope you found something interesting in this article and probably improved your knowledge in GANs and find new ideas for developing.

With the all-time wishes,

Bondarenko K., machine learning engineer.

Share this post