This is the repository for the group 4 for the seminar of Deep Learning for Computer Vision in ETSETB:
https://github.com/imatge-upc/telecombcn-2016-dlcv/
And the group is formed by:
Manel Baradad, Míriam Bellver, Martí Cervià, Hector Esteban, Carlos Roig
The slides of the project can be seen here:
https://docs.google.com/presentation/d/1daS4M7e5Grk6Ytqk2kdapNonDVCPKwiv5HxscIA8UQI/edit?usp=sharing
We have performed 5 different tasks, described in the following lines:
TASK 1: ARCHITECTURE
- Build network for classification problem
- Study memory requirements and computational loads for different layers
Main idea: use MNIST and start with small network, test different layers and architectures
TASK 2: TRAINING
- Study impact in performance of DATA AUGMENTATION, batches size, batch normalization
- Training validation curves
- Overfitting
TASK 3: VISUALIZATION
- Visualize filter responses from own and also pretrained net
- t-SNE
- Off-the-shelf AlexNet
TASK 4: TRANSFER LEARNING
- Train network on CIFAR10 and fine-tune for Terrassa Buildings 900
- Off-the-shelf convnet
TASK 5: OPEN PROJECT
- We have played with neural style, changing the features that encode the style and see the results.
DATASETS:
We have worked with three different Datasets:
MNIST CIFAR10 Terrassa-building-900 (https://imatge.upc.edu/web/resources/terrassa-buildings-900)