from tensorflow import keras import numpy as np from keras_radam import RAdam # Build toy model with RAdam optimizer model = keras.models.Sequential() model.add(keras.layers.Dense(input_shape=(17,), units=3)) model.compile(RAdam(), loss='mse') # Generate toy data x = np.random.standard_normal((4096 * 30, 17)) w = np.random.standard_normal((17, 3)) y = np.dot(x, w) # Fit model.fit(x, y, epochs=5)

from keras_radam import RAdam RAdam(total_steps=10000, warmup_proportion=0.1, min_lr=1e-5)