sample size is not divisible by batch_size in multimodal model

Question

I need help in code debugging, given X1_train_scaled_.shape=(2960642, 17, 1), X2_train_padded.shape=(414, 17, 1), X1_test_.shape=(1214889, 17, 1), X2_test_padded.shape=(190, 17, 1),

• two modalities have different sizes,
• in training I need to create batches for first modality(X1) making samples in each batch is same as X2_train_padded, and in the last batch, randomly select samples from X2_train_padded making it same size as the remainder of X1_train_scaled_,
• in validation, I need to create batches for first modality(X1) making samples in each batch is same as X2_test_padded, and in the last batch, randomly select samples from X2_test_padded making it same size as the remainder of X1_test_, but I got IndexError: index 190 is out of bounds for axis 0 with size 190 error, can anyone help me solve it? Below is my code:

# Concatenate the features from both modalities
merged_features = concatenate([facial_model.output, mouse_.output])

# Joint Learning
joint_layer = Dense(2, activation='relu', name='dense_hidden_final1')(merged_features)
joint_layer = Dropout(0.5)(joint_layer)
joint_layer = Dense(32, activation='relu', name='dense_hidden_final2')(joint_layer)
joint_layer = Dropout(0.5)(joint_layer)

# Output Layer
output_layer = Dense(1, activation='sigmoid', name='output')(joint_layer)

# Combined Model
model = Model(inputs=[facial_model.input, mouse_.input], outputs=output_layer)

# Compile the model
model.compile(
    optimizer='adam',
    loss='binary_crossentropy', 
    metrics=['accuracy', Precision(), Recall(), RootMeanSquaredError()]
)

# Print summary
# model.summary()

# Define custom generator
def custom_generator(X1, X2, y, batch_size):
    total_samples = min(len(X1), len(X2))  # Ensure we use the minimum length
    batches_per_epoch = total_samples // batch_size
    
    while True:
        for i in range(batches_per_epoch):
            start_idx = i * batch_size
            end_idx = (i + 1) * batch_size
            yield ([X1[start_idx:end_idx], X2[start_idx:end_idx]], y[start_idx:end_idx])
        
        # Handle the last batch
        if total_samples % batch_size != 0:
            start_idx = batches_per_epoch * batch_size
            end_idx = total_samples
            # print(f"Debug: last batch - start_idx: {start_idx}, end_idx: {end_idx}")
            
            # Randomly select samples from the second modality to match the size of the first modality
            selected_indices = np.random.choice(len(X2), size=batch_size - (end_idx - start_idx), replace=True)
            # print(f"Debug: selected_indices: {selected_indices}")
            
            yield ([X1[start_idx:end_idx], X2[selected_indices]], y[start_idx:end_idx])

# Training
batch_size = X2_train_padded.shape[0]  

history_joint = model.fit(
    custom_generator(X1_train_scaled_, X2_train_padded, y1_trainE, batch_size),
    steps_per_epoch=len(X1_train_scaled_) // batch_size + 1,  # +1 to include the last batch
    epochs=10,
    callbacks=[stop, best],
    validation_data=([X1_test_, X2_test_padded], y1_testE),
    verbose=1
)