keras image generator

walkthrough on how to use keras image generator. Easy steps.

let’s use the dataset : https://www.kaggle.com/cactus3/basicshapes

it is the perfect dataset to practice keras image generator

1. Create a dataframe:

base_path = '/kaggle/input/basicshapes/shapes/'
def get_filenames(shape):
    path = os.path.join(base_path, shape)
    return [os.path.join(path, img) for img in os.listdir(path)]
c = pd.DataFrame({
    'img_path' : get_filenames('circles'),
    'ishape'    : 'circle'
})

s = pd.DataFrame({
    'img_path' : get_filenames('squares'),
    'ishape'    : 'square'
})

t = pd.DataFrame({
    'img_path' : get_filenames('triangles'),
    'ishape'    : 'triangle'
})

df = pd.concat([c, s, t]).reset_index(drop = True)
df.head()

make sure it outputs df similar to this

| img_path | ishape |
| --- | ----------- |
| /kaggle/input/basicshapes/shapes/circles/drawi... | circle |
| /kaggle/input/basicshapes/shapes/triangles/drawi... | triangle|
| /kaggle/input/basicshapes/shapes/squares/drawi...  | square |

2. import the required modules

import tensorflow as tf 
import tensorflow.keras as keras 
from tensorflow.keras import Sequential
from tensorflow.keras.utils import plot_model, to_categorical
from tensorflow.keras.layers import Dense, Input, Dropout, BatchNormalization, Flatten
from sklearn.metrics import classification_report, confusion_matrix
from sklearn.model_selection import train_test_split
from tensorflow.keras.preprocessing.image import ImageDataGenerator
import tensorflow_addons as tfa

3. split the df into train and test so we can create a two different data generators.

x = df.drop(['ishape'], axis = 1)
y = df.ishape

x_train, x_dev, y_train, y_dev = sample(x, y, 0.3, True)

y_train.shape, y_dev.shape
((210,), (90,))

4. create a data generator function so its easy to create multiple different data generators easily in one line of code

def create_data_generators(df, batch_size = 16, datagen_args = {
     'rotation_range'     : 20,
     'width_shift_range'  : 0.2,
     'height_shift_range' : 0.2,
     'rescale'            : 1./255,
     'vertical_flip'      : True,
     'horizontal_flip'    : True
    }):
    
    datagen = ImageDataGenerator(**datagen_args)
    datagenerator = datagen.flow_from_dataframe(
                                            df,
                                            base_path,
                                            x_col = 'img_path',
                                            y_col = 'ishape',
                                            target_size = (28,28),
                                            class_mode='categorical',
                                            batch_size = batch_size, 
                                            shuffle=True
                                            )
    return datagenerator

if you want to know more about the parameters look them up here : https://www.tensorflow.org/api_docs/python/tf/keras/preprocessing/image/ImageDataGenerator

5. join the x_train and y_train into one df and x_dev and y_dev into another df in order to create a image data generator

train_df = pd.concat([x_train, y_train], axis = 1)
dev_df = pd.concat([x_dev, y_dev], axis = 1)


train_datagen_args = {
     'rescale'            : 1./255,
     'zoom_range'         : 0.1, 
     'vertical_flip'      : True,
     'horizontal_flip'    : True
    }
train_gen = create_data_generators(train_df, datagen_args = train_datagen_args)
dev_gen = create_data_generators(dev_df, datagen_args = {'rescale': 1./255})

if successful then it should log something like this

Found 210 validated image filenames belonging to 3 classes.
Found 90 validated image filenames belonging to 3 classes.

6. finally you can feed that train_gen into your model using model.fit_generator in keras