Deep learning series project

Here is the summary of the projects I did in the course CPSC8430 from the Clemson computer science department.

Image classification based on DNN and CNN

The DNN and CNN have been built with Tensorflow in order to classify the image from MINST dataset. Moreover, the PCA weights analysis, gradient observation, generalization analysis, and sensitivity analysis have been conducted. The report and codes can be found in my Github repository

Video caption with seq2seq model

In this project, a typical video caption task is fulfilled by the seq2seq model. Given a short video, the caption can be generated by the seq2seq model, which is built with Tensorflow. In addition, the Bahdanau attention and schedule sampling features are also added to improve the performance. The block diagrams have been shown in the following pics. The report and codes can be found in my Github repository

Generate pictures with GANs

In this project, three different forms of GANS, i.e., DCGAN, WGAN, and ACGAN, are applied to generate realistic images based on the cifar-10 image database. These networks are built with Tensorflow. The cifar-10 contains 50000training 32x32 RGB images with ten different classes, which are airplane, automobile, bird, cat, deer, dog, frog, horse, ship, and truck. The structure of GANs and ACGANs are shown as follows. The report and codes can be found in my Github repository