Initializing the val_loss with not Inf value #12803

def batch_generator(batch_size, sequence_length): 
    for i in range(batch_size):
       x_batch[i] = batch_x_train_scaled[idx:idx+sequence_length]
       y_batch[i] = batch_y_train_scaled[idx:idx+sequence_length] 
 yield (x_batch, y_batch)
I use also a ModelCheckpoint to save the best-trained model
callback_checkpoint = ModelCheckpoint(filepath=path_checkpoint, 
monitor='val_loss', verbose=1,save_weights_only=False, save_best_only=True)
Also, every time I want to resume the training, first I load the last saved model:
if True:
#         model.load_weights(path_checkpoint)
        model = load_model(path_checkpoint)
    except Exception as error:
        print("Error trying to load checkpoint.")
        print(error)
What is the Problem? Every time I resume the training, a new batch file is loaded first into the fit_generator and the model would use the last save weights. the loss value is reset to inf. Therefore, at the end of the first resumed training epoch, no matter how good or bad is the training outcomes, the model reports that the val_loss is improved from inf to some number and thus overwrites the new weights. The problem is that sometimes the new weights are not as optimal as the previous ones (due to the fact that this time the model is using new batch data for training) and thus I will lose some optimum weights.
What I have done so far to fix this issue?
Approach one (unsuccessful): defining custom loss function:
def my_loss(y_true, y_pred):
    train_loss = binary_crossentropy(y_true, y_pred)
    validation_loss = 2*binary_crossentropy(y_true, y_pred)
    temp=tf.keras.backend.cast(validation_loss,'float16')
    if temp>1:  # update 1 to last best val_loss before resume training
        validation_loss=validation_loss+np.inf
# validation_loss=np.inf
    return tf.keras.backend.in_train_phase(train_loss, validation_loss)
model.compile(loss=my_loss, optimizer=optimizer)
results of approach one:
Error:
---> 12     if temp>1:
TypeError: Using a `tf.Tensor` as a Python `bool` is not allowed. Use `if t is not None:` instead of `if t:` to test if a tensor is defined, and use TensorFlow ops such as tf.cond to execute subgraphs conditioned on the value of a tensor.
Approach two (unsuccessful): defining a custom callback to save model:
best_val_loss = 1 # update 1 to last best val_loss before resume training
def saveModel(epoch,logs):
    val_loss = logs['val_loss']
    if val_loss < best_val_loss:
        best_val_loss=val_loss
        model.save('my_model.hdf5')
my_callback = LambdaCallback(on_epoch_end=saveModel)
results of approach two:
UnboundLocalError: local variable 'best_val_loss' referenced before assignment
Approach three(unsuccessful): defining a custom callback to save model:
best_val_loss = 1 # update 1 to last best val_loss before resume training
def saveModel(epoch,logs,best_val_loss):
    val_loss = logs['val_loss']
    if val_loss < best_val_loss:
        best_val_loss=val_loss
        model.save('my_model.hdf5')
my_callback = LambdaCallback(on_epoch_end=saveModel)
results of approach three:
TypeError: saveModel() missing 1 required positional argument: 'best_val_loss'