telecombcn-dl / dlcv05

DLCV-05

Our team was composed by:

Albert Jiménez, Marc Górriz, Adrià Romero, Michelle Compri, Denjj Osele

Project

[Project Slides] (https://github.com/telecombcn-dl/dlcv05/blob/master/Project%20Presentation.pdf)

[Presentation] (https://docs.google.com/presentation/d/1mflcMsA4rgJHat2hwqVu6zJjPiipFmea3BJUwwSxVtI/edit#slide=id.g14a9129f29_0_0)

Task 1: Architecture

Build your own network to solve a classification task.

Script mnist_cnn.py:

• Options added at the beginning of the script
• We can save and load trained models
• We have the value of the loss & accuracy at each epoch
• Save total time computed

Script cifar10_cnn.py:

• Options added at the beginning of the script
• We can save and load trained models
• We have the value of the loss & accuracy at each epoch
• Save total time computed

Script mnist_cnn_3layers.py:

• Custom architecture proposed

If you are saving the model be careful when setting the paths and the name not to overwrite!

Task 2: Training

Objectives:

Study the impact in performance of:

• Data augmentation.
• Sizes of the training batches.
• Batch normalization

Overfitting:

• Force an overfitting problem.
• Investigate if regularization (eg. drop out) reduces/solves it.

Task 3: Visualization

Objective: Visualize filter responses

• There is the code to visualize the value of the weight as well as the output of every filter in our custom architecture on mnist database.

Task 4: Transfer Learning

(Experimental code ... not working properly when fine tuning)

Train a network over CIFAR-10 and fine-tune over Terrassa Buildings 900.

Off the Shelf VGG-16

• Freeze weight in all layers but the last one, and replace it with a softmax to solve Terrassa Buildings 900.

Task 5: Open Project

Adquire knwoledge and insights about what is happening inside the deepdream network.

• Deepdream modify the images it is given as an input enhancing some features depending on the layer we choose to boost.
• Lower layers focus on simpler features (edges, orientation, shapes)
• Higher layers focus on concrete objects that have been seen during training