The project is now finished ! I would like to conclude this blog by summarizing the results and my experience : We started by simple models for naive generation, trying to understand which architecture would generate sharper images. Of course we could not expect to come out with interesting results without using the captions, but… Continue reading Conclusion

## Conditional LSGAN

Since my first attempts weren't very satisfying, I decided to move forward to a more sophisticated and better performing model : Least-Squares GAN. Generative Adversarial Networks have already showed their performance on image generation. Specifically, DCGANs showed impressive performance in unsupervised learning for image generation. The architecture of a traditional GAN is the following : A… Continue reading Conditional LSGAN

## Fully-Convolutional Generator

As discussed in the previous post, I wanted to explore the performance of a Fully-Convolutional model. The idea is that pooling layers and the encoding layer may result in a loss of information that is too difficult to recover. Given only the image latent representation, recovering its details is an under-determined problem ! Therefore, this… Continue reading Fully-Convolutional Generator

## Convolutional Auto-Encoder

First Model My first attempt for this class project is to implement a Convolutional Auto-Encoder, which a modified form of the following architecture. The specific activation functions I used (Tanh and Elu) result from various tests I did over small train set sizes. Exponential linear unit is useful for dropping out useless information while keeping […]

## Introduction

This blog is the presentation of my results for the final project of IFT6266 course on Deep Learning given at Université de Montréal in winter 2017. The project is to generate the middle region of images conditioned on the outside border and a caption describing the image. To perform this task, I will use different… Continue reading Introduction