Генератор проверки TensorFlow не вычисляет val_loss после первой эпохи ⇐ Python

[Anonymous]

Я пытаюсь обучить модель, используя генераторы для наборов данных обучения и проверки. Каждая эпоха потребляет все пакеты набора проверочных данных (верно?). Первая эпоха работает нормально, но начиная со второй эпохи выдает ошибку, сообщающую, что невозможно вычислить val_loss. Я подозреваю, что это может быть связано с тем, что набор проверочных данных не сбрасывается, но я не уверен. Как это исправить?
ошибка
Cell In[13], line 43
41 #treinar modelo
42 now = datetime.now()
---> 43 model.fit(train_dataset, epochs=5, verbose=1, validation_data=val_dataset,
44 callbacks=[printbatch(), EarlyStopping(monitor='val_loss', patience=50, verbose=0),
45 ModelCheckpoint(os.path.join(output_path, "models", "dnn_model_" + str(tam) + "_" + str(now.day) + "." +
46 str(now.month) + "." + str(now.year) + "_" + "_{epoch:02d}-{val_loss:.2f}.keras"),
47 monitor='val_loss', verbose=0, save_best_only=True, save_weights_only=False, mode='auto'),
48 reset_val_callback])
49 model.save(os.path.join(output_path,"dnn_model_" + str(tam) + "_" + str(now.day) + "." +str(now.month) + "." + str(now.year) + "_.keras"))
...
325 f"Reason: {e}"
326 )
327 return file_path

KeyError: 'Failed to format this callback filepath: "modelos/models/dnn_model_150_11.12.2024__{epoch:02d}-{val_loss:.2f}.keras". Reason: \'val_loss\''

также
/home/igor/anaconda3/envs/tcc/lib/python3.9/site-packages/keras/src/callbacks/early_stopping.py:155: UserWarning: Early stopping conditioned on metric `val_loss` which is not available. Available metrics are: accuracy,loss
current = self.get_monitor_value(logs)

мой код
def data_generator(file_paths, batch_size, files_per_batch, tam, value, val_split):
# Pegando todos os arquivos do diretório
files = sorted([os.path.basename(ii) for ii in glob.glob(f"{file_paths}/*.dimacs")])
cut_index = int(len(files) * 0.9) # Índice para os primeiros 90% dos arquivos
files = files[:cut_index] # Mantém apenas os primeiros 90% para usar no treino
# Dividir aleatoriamente 20% dos arquivos para validação e 80% para treinamento
files = files[:1000]
files_t, files_v = train_test_split(files, test_size=val_split, random_state=42)

# Gerador de treinamento
def create_generator(file_list):
def generator():
idx = 0
remaining_batch_input = remaining_batch_labels = []
while idx < len(file_list):
batch_files = file_list[idx: idx + files_per_batch]
res, clique, labels = ler_arquivos(batch_files, file_paths, tam, value)
combined_input, labels = combinar_entrada(res, clique, tam, labels, remaining_batch_input, remaining_batch_labels)
remaining_batch_input = remaining_batch_labels = []
num_samples = len(combined_input)
num_batches = num_samples // batch_size
for batch_idx in range(num_batches):
batch_input = combined_input[batch_idx * batch_size: (batch_idx + 1) * batch_size]
batch_labels = labels[batch_idx * batch_size: (batch_idx + 1) * batch_size]
x_batch, y_batch = estruturar_entrada(batch_input, batch_labels)
yield x_batch, y_batch
idx += files_per_batch
remaining_samples = num_samples % batch_size
if remaining_samples > 0:
remaining_batch_input = combined_input[-remaining_samples:]
remaining_batch_labels = labels[-remaining_samples:]
return generator

return create_generator(files_t)(), create_generator(files_v)()

train_generator_func, val_generator_func = data_generator(file_paths, batch_size, files_per_batch, tam, False, 0.2)
train_dataset = tf.data.Dataset.from_generator(lambda : train_generator_func, output_signature=output_signature_type(batch_size, tam))
val_dataset = tf.data.Dataset.from_generator(lambda : val_generator_func, output_signature=output_signature_type(batch_size, tam))

# Prefetch para melhorar a performance
train_dataset = train_dataset.prefetch(tf.data.AUTOTUNE)
val_dataset = val_dataset.prefetch(tf.data.AUTOTUNE)

# Instanciando o callback para reiniciar o gerador de validação
reset_val_callback = ResetValidationGenerator(val_generator_func=val_generator_func, batch_size=batch_size, tam=tam)

#treinar modelo
now = datetime.now()
model.fit(train_dataset, epochs=5, verbose=1, validation_data=val_dataset,
callbacks=[printbatch(), EarlyStopping(monitor='val_loss', patience=50, verbose=0),
ModelCheckpoint(os.path.join(output_path, "models", "dnn_model_" + str(tam) + "_" + str(now.day) + "." +
str(now.month) + "." + str(now.year) + "_" + "_{epoch:02d}-{val_loss:.2f}.keras"),
monitor='val_loss', verbose=0, save_best_only=True, save_weights_only=False, mode='auto'),
reset_val_callback])

Я создал обратный вызов, чтобы попытаться сбросить набор данных, но все равно получаю ту же ошибку.
class ResetValidationGenerator(tf.keras.callbacks.Callback):
def __init__(self, val_generator_func, batch_size, tam):
super(ResetValidationGenerator, self).__init__()
self.val_generator_func = val_generator_func
self.batch_size = batch_size
self.tam = tam
self.val_dataset = None

def on_epoch_end(self, epoch, logs=None):
# Aqui você pode reiniciar o gerador de validação
print(f"Reiniciando o gerador de validação após a época {epoch + 1}")

# Reinicia o gerador de validação chamando a função que cria o gerador
self.val_dataset = tf.data.Dataset.from_generator(
lambda: self.val_generator_func(),
output_signature=output_signature_type(self.batch_size, self.tam)
).prefetch(tf.data.AUTOTUNE)


Подробнее здесь: https://stackoverflow.com/questions/792 ... irst-epoch