migueldsw / tl-da-tf

Transfer Learning in Deep Architectures using TFLearn -> (Google)TensorFlow

TL-DA-TF

Transfer Learning in Deep Architectures using TFLearn -> (Google)TensorFlow

Networks

Adjusted networks for 28x28x3 image inputs. Code for pre training and training (fine tuning).

VGG-11 VGG-16

• VGG-11/16: 16 and 11 layers.

AlexNet

• AlexNet: 8 layers.

LeNet-5

• LeNet-5: 5 layers.

Datasets

Source and target datasets explored for transfer learning combinations.

MNIST

RGB channels included (28x28x3)

SVHN

Images cropped to 28x28x3

CIFAR-10

Images cropped to 28x28x3

tl-da-tf's Issues

Plot graphs

plot graphs for reports. Acc x Layer over 'base', 'selffer' and 'transfer'.

Implement experiments

Write the code to schedule, run and report all experiments.

• Experiments (main) code.
• Define source and target tasks (datasets and predict tasks).
• Evaluate basic model.
• Evaluate transferred model.

Clean VGG 16 code

Clean VGG 16 code(and project) for experiments. (vgg_net_mnist.py)

Experiments on CIFAR-10

CIFAR-10 ref: https://www.cs.toronto.edu/~kriz/cifar.html
tflearn implementation: https://github.com/tflearn/tflearn/blob/master/tflearn/datasets/cifar10.py

Transfer VGG training

Transfer VGG model training.
Create model for initial training 'started model' in a source domain(dataset) and another model that loads the started model 'loaded model' and continues the training proccess for a new target (dataset for prediction and evaluation).
Evaluate the transfered 'loaded model'.

• Implement 'Started Model'
• Implement 'Loaded Model'

Explore transfer possibilities

Which layers weights must/can be restored ? In VGG-11 and VGG-16 transfer.

Implement VGG-11

Implement VGG 11 weight layers.
Ref.: http://arxiv.org/pdf/1409.1556

Evaluate time cost and accuracy on MNIST and SVHN.