This repository contains code for training training deep learning model using k-cross validation for image data.
We expect to have proir knowledge of keras and tensorflow. Also see keras documentation for Sequential model and Compilation
You can use any version of the packages but we have used the following versions:
| Packages | Versions |
|---|---|
| Keras | 2.1.3 |
| Tensorflow | 1.8.0 |
| Numpy | 15.4.0 |
| Matplotlib | 3.0.2 |
We assume that your dataset is in the form of images. These images are located in every class-folder of training. Before getting to this point of work, your code might look like this
# Load entire dataset
path_dataset_train='/home/activity/train'
# Design model
model.Sequential()
model.add(Conv2D(filters, kernel_size, strides=(1, 1))
.....
# Compile model with your optimizers
model.compile(optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
In python headers where packages are imported code this line
from cross_Validation import training_crossVal
then pass following parameters for training of the model initialized above.
dataset_train='/home/abdulrehman/images/train'
trained_model=training_crossVal(kvalidation_splits=7,train_batch_size=2100,model_train=model,epochs=15,image_directory_path=dataset_train)
On every epoch the dataset fed to the model will be train_batch_size/kvalidation_splits. So if you have total training dataset of 10000 images, train_batch_size=2100, kvalidation_splits=7. In every epoch, generator will pick 2100 images with labels randomly and 7 sub-epochs will run having batch of train_batch_size/kvalidation_splits=300 for training and validation each.
- Total epochs will be kvalidation_splits x epochs
- Total epochs for above example are
7x15=105 epochs
The epochs are great in number but the batch size taken for each epoch becomes small and in each epoch data is rotated randomly. It enhances the training capability and time for training is same for the model if other generator with no-sub epochs are used.
This project is under GNU General Public License v3.0 see License file