python Ftrl optimizer does not distribute classes correctly

Question

I have a task: to create RNN: 1 LSTM layer, softmax activation function, FTRL optimizer. I follow the example and the adam optimizer is used there. With this optimizer, I get this graph: Adam optimizer

I change the optimizer to FTRL and get this graph: FTRL optimizer

Who can explain what the problem is? How can I get a class distribution graph like with the adam optimizer?

Code:

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
from sklearn.datasets import make_classification
from tensorflow.python.keras.layers import Dense
from sklearn.metrics import accuracy_score



def plot_scatter(pred):
    plt.figure()
    plt.scatter(range(0, len(pred)), pred)
    plt.xlabel('x')
    plt.ylabel('y')
    plt.show()


train_size = 5000
n_samples = train_size + 100
n_features = 10
n_classes = 5
x_data, y_data = make_classification(
    n_samples=n_samples,
    n_features=n_features,
    n_informative=10,
    n_redundant=0,
    n_classes=n_classes
)


x_train = x_data[0:train_size]
y_train = y_data[0:train_size]
x_test = x_data[train_size:]
y_test = y_data[train_size:]


model = tf.keras.Sequential()


model.add(tf.keras.layers.LSTM(200,  
                               return_sequences=True,  
                               input_shape=(n_features, 1))  
          )

model.add(tf.keras.layers.LSTM(200))  
model.add(Dense(n_classes, activation='softmax'))

model.compile(optimizer='Ftrl',
              loss='sparse_categorical_crossentropy'
              )

x_train = np.expand_dims(x_train, axis=2)
x_test = np.expand_dims(x_test, axis=2)


model.fit(
    x_train,
    y_train,
    validation_data=(x_test, y_test),
    epochs=10
)

if __name__ == '__main__':
    y_pred = model.predict(x_test)
    classes = np.argmax(y_pred, axis=1)
    print(accuracy_score(y_test, classes))
    plot_scatter(classes)