Styjun

I want to change my class_weight during training in Keras.

I used fit_generator and Callback method like below.

model.fit_generator(
 decoder_generator(x_train, y_train),
 steps_per_epoch=len(x_train),
 epochs=args.epochs,
 validation_data=decoder_generator(x_valid, y_valid),
 validation_steps=len(x_valid),
 callbacks=callback_list,
 class_weight=class_weights,
 verbose=1)

And

class Valid_checker(keras.callbacks.Callback):
 def __init__(self, model_name, patience, val_data, x_length):
 super().__init__()
 self.best_score = 0
 self.patience = patience
 self.current_patience = 0 
 self.model_name = model_name
 self.validation_data = val_data
 self.x_length = x_length


 def on_epoch_end(self, epoch, logs=):
 X_val, y_val = self.validation_data
 y_predict, x_predict = model.predict_generator(no_decoder_generator(X_val, y_val), steps=len(X_val))
 y_predict = np.asarray(y_predict)
 x_predict = np.asarray(x_predict) 

decoder_generator and no_decoder_generator are just custom generator.

I want to change the class weight every time the epoch ends. Is it possible? Then how can I do?

My data is imbalanced data, and overfitting is continued for one class.

At the end of the epoch I want to increase the weight for classes with low accuracy by calculating the accuracy by class.

How can I do?

How about a simple approach like looping over one epoch at a time ?

for i in range(args.epochs):
 class_weights = calculate_weights()
 model.fit_generator(
 decoder_generator(x_train, y_train),
 steps_per_epoch=len(x_train),
 epochs=1,
 validation_data=decoder_generator(x_valid, y_valid),
 validation_steps=len(x_valid),
 callbacks=callback_list,
 class_weight=class_weights,
 verbose=1)

There is no straight forward way to use different class weights for each epoch in fit_generator. You can incorporate early stopping by checking the value of model.stop_training

Sample

import numpy as np
from keras.models import Sequential
from keras.layers import Input, Dense
from keras.models import Model
from keras.callbacks import Callback

class Valid_checker(Callback):
 def __init__(self):
 super().__init__()
 self.model = model
 self.n_epoch = 0

 def on_epoch_end(self, epoch, logs=): 
 self.n_epoch += 1
 if self.n_epoch == 8:
 self.model.stop_training = True

def decoder_generator():
 while True: 
 for i in range(10):
 yield np.random.rand(10,5), np.random.randint(3,size=(10,3))


inputs = Input(shape=(5,))
outputs = Dense(3, activation='relu')(inputs)
model = Model(inputs=inputs, outputs=outputs)
model.compile(optimizer='rmsprop',
 loss='categorical_crossentropy',
 metrics=['accuracy'])

for i in range(10): 
 model.fit_generator(generator=decoder_generator(),
 class_weight=0:1/3, 1:1/3, 2:1/3,
 steps_per_epoch=10, 
 epochs=1,
 callbacks=[Valid_checker()])
 if model.stop_training:
 break