Tensorflow VGG Transfer Learning - Trainable

This is a Tensorflow implemention of VGG 16 and VGG 19 based on tensorflow-vgg, developed for Tensor r1.0

We have modified the implementation of tensorflow-vgg so that the model support different image formats (jpg,png,jpeg). Additionally we have developed a function In addition we have developed a function to perform model training using mini-batches and epochs. The project own us modify the original model and load the weights pre-trained in the required layer.

Download weights pre-trained:

The guiding project (tensorflow-vgg) for the development of this repository uses transfer learning with the weights obtained from the training of the VGG network with dataset Imaginet.

-OPTION A: To use the VGG networks, the npy files for VGG16 NPY or VGG19 NPY has to be downloaded. Uploaded in Mega.nz repositories.

-OPTION B: You can download the pre-trained weights by terminal. You should install Megatools and follow the guide in the website [link].

-OPTION C: You can also download the files from a repository in dropbox, by terminal:

            vgg16.npy[528MB]: wget https://www.dropbox.com/s/8a8rei66f72um4i/vgg16.npy
            vgg19.npy[548MB]: wget https://www.dropbox.com/s/691wtp4oq5ip38p/vgg19.npy

Project Files

The files test_vgg16.py, test_vgg19.py are simple examples, that can be executed only by cloning the project.
The file test_vgg19_trainable.py is a simple example of test and training of a CNN model. Also can be executed only by cloning the project
The file test_vgg19_all.py is a complete classification model that require other files and a dataset more elaborate.

Usage

Use this to build the VGG object

vgg = vgg19.Vgg19()
vgg.build(images)

vgg = vgg16.Vgg16()
vgg.build(images)

or by training

vgg = vgg19.Vgg19([path_vgg.npy], load_weight_fc)
vgg.build(images, train_mode)

The [path_vgg.npy] is the path where the file '.npy' is located. The load_weight_fc is of type bool, this variable own us reset the weights of the fully-connected layers, but can train the model with a new dataset. The images is a tensor with shape [None, 224, 224, 3]. The train_mode is of the type bool, this variable own us enable or disable the dropout layer.

Trick: the tensor can be a placeholder, a variable or even a constant.

Access to parameters

All the VGG layers (tensors) can then be accessed using the vgg object. For example, vgg.conv1_1, vgg.conv1_2, vgg.pool5, vgg.prob,etc. Even recover the weights of the model layers vgg.var_dict[(name, idx)], where the variable name is the name assigned to a layer and idx indicates the type of weights we want to recover, filters/weigths or biases.

The test_vgg16.py and test_vgg19.py contain the sample usage.

Load Dataset

The class for loading a dataset and generating sub sets for the minibatchs is in the file 'dataseTools.py', this class requires three input data: