IFT6266 – H2017 DEEP LEARNING

Throughout the experiments, I observed that some hacks should be applied to get around training models faster or to get better outputs such as the famous GAN Hacks by Soumith Chintala or observing why we get the checkerboard effects when we use the transposed convolutions (as given in a blog post by Chris Olah). Some… Continue reading What else could be tried next?

Almost all the papers where we have a generative setting that is the application whose end result is generation of some image, employ the use of qualitative analysis. They select the best sample from their test results and publish the same. This process is subjective and as per the discretion of the academic writer. There… Continue reading Ways to analyse the results!

  In my final set of experiments, first I made a training set where I not only put the cropped out image (in which there is no central region) but also the original image. For generator, I use a 3072 vector based on the 32X32 cropped out region. The model reaches a zero error as… Continue reading DCGANs with captions – Final Experiments

In the previous post we saw that the 100 dim vector used to generate the central region did not give aesthetically pleasing results. So, now I use a 3072 dimension vector i.e. 32×32=3072 which the required size. Below are the results. These results show that at least now we can expect some colours in the… Continue reading Deep Convolutional GANs- Part 2

So, using GANs seems the obvious next step given its recent popularity and performance in generative tasks! I have trained a GAN model and it can be seen in this IPython Notebook. The following are the graphs which I obtain for the training loss over time for the generator and discriminator. The good thing is… Continue reading Deep Convolutional Conditional GANs

We are provided with captions for each of the images. These captions give us a brief description about what is happening in the image. For the image completion task these captions can be useful by providing overhead information. They can try to give additional information which the pixels alone in the image might fail to… Continue reading Which word embeddings to use?

Manifold Learning and T-SNE Plots: The images are high dimensional structures and as such not much information can be inferred from directly visualizing them. A method proposed by Hinton was t-SNE and recently a new version came out known as parametric t-SNE. One can see in the image below how the different images are portrayed… Continue reading Let us visualize the data manifold

I started off with setting up the environment and visualizing the images. The input is the image with the central 32X32 region removed and blacked out. The target is the removed 32X32 part from the original image. For now, I am only concentrating on the images. Later, I will use GANs and then also incorporate… Continue reading Experiments with Deep Convolutional Autoencoders