I am running different datasets with different numbers of clients on federated learning code. However, when working with some datasets, I can set the value of NUM_CLIENTS to a maximum of 20. When the client value exceeds 20, I encounter the error below. Where am I making a mistake?

ValueError                                Traceback (most recent call last)
<ipython-input-19-3e647171f427> in <cell line: 18>()
     35 
     36 
---> 37         client_model.fit(datagen(client[0],client[1], batch_size=batch_size, epochs=epochs),
     38                          epochs=5, steps_per_epoch=steps_per_epoch)
     39 

1 frames
/usr/local/lib/python3.10/dist-packages/keras/src/engine/data_adapter.py in __init__(self, x, y, sample_weight, batch_size, steps_per_epoch, initial_epoch, epochs, shuffle, class_weight, max_queue_size, workers, use_multiprocessing, model, steps_per_execution, distribute, pss_evaluation_shards)
   1273 
   1274         if steps_per_epoch == 0:
-> 1275             raise ValueError(
   1276                 "Unexpected value for `steps_per_epoch`. Received value is 0. "
   1277                 "Please check the docstring for `model.fit()` for supported "

ValueError: Unexpected value for `steps_per_epoch`. Received value is 0. Please check the docstring for `model.fit()` for supported values.

import tensorflow as tf
import numpy as np
from tensorflow.keras.preprocessing.image import ImageDataGenerator

# Define the number of clients and the number of training rounds
NUM_CLIENTS = 10
NUM_ROUNDS = 3


# Define the clients and their data
clients = []
for i in range(NUM_CLIENTS):
    client_data = train_paths[i * (len(train_paths) // NUM_CLIENTS):(i + 1) * (len(train_paths) // NUM_CLIENTS)]
    client_labels = train_labels[i * (len(train_labels) // NUM_CLIENTS):(i + 1) * (len(train_labels) // NUM_CLIENTS)]
    clients.append((client_data, client_labels))

# Federated learning loop
for round_num in range(NUM_ROUNDS):

    # Select clients
    selected_client_indices = np.random.choice(len(clients), size=int(NUM_CLIENTS * 0.5), replace=False)
    selected_clients = [clients[i] for i in selected_client_indices]

    # Transmit the global model to the selected clients
    for client in selected_clients:
        client_model = tf.keras.models.clone_model(model)
        client_model.set_weights(model.get_weights())

        # Compile the client model
        client_model.compile(optimizer=Adam(learning_rate=0.0001),
        loss='sparse_categorical_crossentropy',
        metrics=['sparse_categorical_accuracy'])

        steps_per_epoch = int(len(client[0]) / 20)


        client_model.fit(datagen(client[0],client[1], batch_size=batch_size, epochs=epochs),
                         epochs=5, steps_per_epoch=steps_per_epoch)

        # Train locally
        steps_per_epoch = int(len(client[0]) / 20)

        # Aggregate the model
        new_weights = []
        for layer_index in range(len(model.get_weights())):
            new_layer_weights = np.mean([client_model.get_weights()[layer_index], model.get_weights()[layer_index]], axis=0)
            new_weights.append(new_layer_weights)
        model.set_weights(new_weights)


batch_size = 32
steps = int(len(test_paths)/batch_size)
y_pred = []
y_true = []
for x,y in tqdm(datagen(test_paths, test_labels, batch_size=batch_size, epochs=1), total=steps):
    pred = model.predict(x)
    pred = np.argmax(pred, axis=-1)
    for i in decode_label(pred):
        y_pred.append(i)
    for i in decode_label(y):
        y_true.append(i)

#     # Evaluate the global model
#     test_loss, test_acc = model.evaluate(test_paths, test_labels,)
#     print('Round {}: Test accuracy = {}'.format(round_num, test_acc))

# # Fine-tune the model
# model.fit(test_paths, test_labels, epochs=1, batch_size=32)

# Deploy the model
model.save('my_model.h5')

I want to increase the number of clients to 50