def test(model: tf.keras.Model,
train_data: List[Tuple[int, int, int]],
test_data: List[Tuple[int, int, int]],
topk_data: TopkData,
loss_object=None,
batch=512) -> None:
"""
通用测试流程。
:param model: 模型
:param train_data: 训练集
:param test_data: 测试集
:param topk_data: 用于topk评估数据
:param loss_object: 损失函数，默认为BinaryCrossentropy
:param batch: 批数量
"""
if loss_object is None:
loss_object = tf.keras.losses.BinaryCrossentropy()
train_ds, test_ds = prepare_ds(train_data, test_data, batch)
train_loss, train_auc, train_precision, train_recall = _evaluate(
model, train_ds, loss_object)
test_loss, test_auc, test_precision, test_recall = _evaluate(
model, test_ds, loss_object)
log(-1, train_loss, train_auc, train_precision, train_recall, test_loss,
test_auc, test_precision, test_recall)
topk(topk_data, get_score_fn(model))
def train(model: tf.keras.Model, train_data: List[Tuple[int, int, int]], test_data: List[Tuple[int, int, int]],
topk_data: TopkData = None, optimizer=None, epochs=100, batch=512):
if optimizer is None:
optimizer = tf.keras.optimizers.Adam()
train_ds, test_ds = prepare_ds(train_data, test_data, batch)
loss_mean_metric = tf.keras.metrics.Mean()
auc_metric = tf.keras.metrics.AUC()
precision_metric = tf.keras.metrics.Precision()
recall_metric = tf.keras.metrics.Recall()
loss_object = tf.keras.losses.BinaryCrossentropy()
if topk_data:
score_fn = get_score_fn(model)
def reset_metrics():
for metric in [loss_mean_metric, auc_metric, precision_metric, recall_metric]:
tf.py_function(metric.reset_states, [], [])
def update_metrics(loss, label, score):
loss_mean_metric.update_state(loss)
auc_metric.update_state(label, score)
precision_metric.update_state(label, score)
recall_metric.update_state(label, score)
def get_metric_results():
return loss_mean_metric.result(), auc_metric.result(), precision_metric.result(), recall_metric.result()
@tf.function
def train_batch(ui, label):
with tf.GradientTape() as tape:
score = model(ui, training=True)
loss = loss_object(label, score) + sum(model.losses)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
update_metrics(loss, label, score)
@tf.function
def test_batch(ui, label):
score = model(ui)
loss = loss_object(label, score) + sum(model.losses)
update_metrics(loss, label, score)
for epoch in range(epochs):
epoch_start_time = time.time()
reset_metrics()
for ui, label in train_ds:
train_batch(ui, label)
train_loss, train_auc, train_precision, train_recall = get_metric_results()
reset_metrics()
for ui, label in test_ds:
test_batch(ui, label)
test_loss, test_auc, test_precision, test_recall = get_metric_results()
log(epoch, train_loss, train_auc, train_precision, train_recall, test_loss, test_auc, test_precision, test_recall)
if topk_data:
topk(topk_data, score_fn)
print('epoch_time=', time.time() - epoch_start_time, 's', sep='')
def _train_graph(model, train_ds, test_ds, topk_data, optimizer, loss_object, epochs):
score_fn = get_score_fn(model)
@tf.function
def train_batch(ui, label):
with tf.GradientTape() as tape:
score = tf.squeeze(model(ui, training=True))
loss = loss_object(label, score) + sum(model.losses)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
for epoch in range(epochs):
for ui, label in train_ds:
train_batch(ui, label)
train_loss, train_auc, train_precision, train_recall = _evaluate(model, train_ds, loss_object)
test_loss, test_auc, test_precision, test_recall = _evaluate(model, test_ds, loss_object)
log(epoch, train_loss, train_auc, train_precision, train_recall, test_loss, test_auc, test_precision, test_recall)
topk(topk_data, score_fn)
def on_epoch_end(self, epoch, logs=None):
log(epoch, logs['loss'], logs['AUC'], logs['Precision'],
logs['Recall'], logs['val_loss'], logs['val_AUC'],
logs['val_Precision'], logs['val_Recall'])
topk(self.topk_data, self.score_fn)
def on_epoch_end(self, epoch, logs=None):
log(epoch, logs['loss'], logs['auc'], logs['precision'], logs['recall'],
logs['val_loss'], logs['val_auc'], logs['val_precision'], logs['val_recall'])
topk(self.topk_data, self.score_fn)
