Multiclass Dataset Imbalance - keras

Question

from tensorflow.keras.preprocessing.image import ImageDataGenerator
import tensorflow as tf
train_path = 'Skin/Train'
test_path = 'Skin/Test'
train_gen = ImageDataGenerator(rescale=1./255)
train_generator = train_gen.flow_from_directory(train_path,target_size=
(300,300),batch_size=30,class_mode='categorical')
model = tf.keras.models.Sequential([
# Note the input shape is the desired size of the image 300x300 with 3 bytes color
# This is the first convolution
tf.keras.layers.Conv2D(16, (3,3), activation='relu', input_shape=(600, 450, 3)),
tf.keras.layers.MaxPooling2D(2, 2),
# The second convolution
tf.keras.layers.Conv2D(32, (3,3), activation='relu'),
tf.keras.layers.MaxPooling2D(2,2),
# The third convolution
tf.keras.layers.Conv2D(64, (3,3), activation='relu'),
tf.keras.layers.MaxPooling2D(2,2),
# The fourth convolution
tf.keras.layers.Conv2D(64, (3,3), activation='relu'),
tf.keras.layers.MaxPooling2D(2,2),
# The fifth convolution
tf.keras.layers.Conv2D(64, (3,3), activation='relu'),
tf.keras.layers.MaxPooling2D(2,2),
# Flatten the results to feed into a DNN
tf.keras.layers.Flatten(),
# 512 neuron hidden layer
tf.keras.layers.Dense(512, activation='relu'),
tf.keras.layers.Dense(9, activation='softmax')
])
from tensorflow.keras.optimizers import RMSprop
model.compile(loss='categorical_crossentropy',
optimizer=RMSprop(lr=0.001),
metrics=['acc'])
history = model.fit_generator(
train_generator,
steps_per_epoch=8,
epochs=15,
verbose=2, class_weight = ? )
I have issue in achieving accuracy, I am training a 9 classes dataset, in which class 1 , 4, and 5 have only 100, 96, 90 images while the remaining classes are having above 500 images. Due to which I am unable to achieve higher accuracy as the weights are skewed toward images which are higher in number. I want that during training all classes are considered equal i.e 500. Would be appreciated if I could upsample the classes via tensorflow or any keras function code. rather than manually upsampling or downsampling the images in folders.

Answer 1

You can, instead, use a class_weight argument in your fit method.
For upsampling, you need a lot of manual work, that's inevitable.
Assuming you have an output with shape (anything, 9), and you know the totals of each class:
totals = np.array([500,100,500,500,96,90,.......])
totalMean = totals.mean()
weights = {i: totalMean / count for i, count in enumerate(totals)}
model.fit(....., class_weight = weights)