object_classification_example.md

Augmenting a dataset for object classification

In this section, we will use a subset of the cats and dogs dataset to show how the library can be used for image augmentation for object classification. The dataset and the configuration files that we will employ for these examples is available at the datasets folder.

Folder-Folder-Linear

In this first example, we show how we can use the library for augmenting a dataset of images saved in folders and store the result in another folder. In this example, we use four augmentation techniques: equalize histogram, median blur, salt and pepper noise, and croping. The json file for this example (called cats_dogs_folder_folder_linear.json) is the following one. If you want to reproduce this example, you only need to change the value for the input_path and output_path parameters.

{
  "augmentation_techniques":[
    [
      "equalize_histogram",
      {
        
      }
    ],
    [
      "median_blur",
      {
        "kernel":"3"
      }
    ],
    [
      "salt_and_pepper",
      {
        "low":"0",
        "up":"25"
      }
    ],
    [
      "crop",
      {
        "startFrom":"TOPLEFT",
        "percentage":"0.9"
      }
    ]
  ],
  "generation_mode":"linear",
  "problem":"classification",
  "output_mode":"folders",
  "parameters":{
    "outputPath":"/home/joheras/cats_dogs/"
  },
  "annotation_mode":"folders",
  "input_path":"/home/joheras/Escritorio/Research/CLoDeSeAugmentor/docs/datasets/object_classification"
}

To run the augmentation process, the following command must be executed from the terminal.

clodsa cats_dogs_folder_folder_linear.json

Folder-Keras-Linear

In this second example, we show how we can use the library for augmenting a dataset of images saved in folders and use it to fed a Keras network. In this example, we use four augmentation techniques: equalize histogram, median blur, salt and pepper noise, and croping. The json file for this example (called cats_dogs_folder_keras_linear.json) is the following one. If you want to reproduce this example, you only need to change the value for the input_path.

{
  "augmentation_techniques":[
    [
      "equalize_histogram",
      {
        
      }
    ],
    [
      "median_blur",
      {
        "kernel":"3"
      }
    ],
    [
      "salt_and_pepper",
      {
        "low":"0",
        "up":"25"
      }
    ],
    [
      "crop",
      {
        "startFrom":"TOPLEFT",
        "percentage":"0.9"
      }
    ]
  ],
  "generation_mode":"linear",
  "problem":"classification",
  "output_mode":"folders",
  "parameters":{
    "batchSize": 32,
    "width": 64,
    "height": 64
  },
  "annotation_mode":"folders",
  "input_path":"/home/joheras/Escritorio/Research/CLoDeSeAugmentor/docs/datasets/object_classification"
}

In this case, to run the augmentation we need to work in a python file where we define and use the network. The code of this example is available in the project webpage.

We start by reading the configuration file that will be passed to the program as input.

ap = argparse.ArgumentParser()
ap.add_argument("-c", "--conf", required=True, help="path to configuration file")
args = vars(ap.parse_args())

conf = Conf(args["conf"])
# First, we read the parameters
problem = conf["problem"]
annotationMode = conf["annotation_mode"]
outputMode = conf["output_mode"]

if not(outputMode=="keras"):
    exit()

generationMode = conf["generation_mode"]
inputPath = conf["input_path"]
parameters = conf["parameters"]
augmentationTechniques = conf["augmentation_techniques"]

Now, we create the augmentor:

augmentor = createAugmentor(problem,annotationMode,outputMode,generationMode,inputPath,
                            parameters)

# Third, we load the techniques and add them to the augmentor
techniques = [createTechnique(technique,param) for (technique,param) in augmentationTechniques]

for technique in techniques:
    augmentor.addGenerator(Generator(technique))

Finally, the model is created and trained.

opt = SGD(lr=0.05)
model = MiniVGGNet.build(width=parameters["width"], height=parameters["height"], depth=3)
model.compile(loss="categorical_crossentropy", optimizer=opt,metrics=["accuracy"])

# Training the model
H = model.fit_generator(augmentor.applyAugmentation(),
                         steps_per_epoch= 32,
                         epochs=100, verbose=1)

In order to execute the above code, the python file must be invoked.

python keras_example.py -c cats_dogs_folder_keras_linear.json