如何在张量流中实现提前停止
Posted
技术标签:
【中文标题】如何在张量流中实现提前停止【英文标题】：how to implement early stopping in tensorflow 【发布时间】：2018-03-07 19:18:08 【问题描述】：def train():
# Model
model = Model()
# Loss, Optimizer
global_step = tf.Variable(1, dtype=tf.int32, trainable=False, name='global_step')
loss_fn = model.loss()
optimizer = tf.train.AdamOptimizer(learning_rate=TrainConfig.LR).minimize(loss_fn, global_step=global_step)
# Summaries
summary_op = summaries(model, loss_fn)
with tf.Session(config=TrainConfig.session_conf) as sess:
# Initialized, Load state
sess.run(tf.global_variables_initializer())
model.load_state(sess, TrainConfig.CKPT_PATH)
writer = tf.summary.FileWriter(TrainConfig.GRAPH_PATH, sess.graph)
# Input source
data = Data(TrainConfig.DATA_PATH)
loss = Diff()
for step in xrange(global_step.eval(), TrainConfig.FINAL_STEP):
mixed_wav, src1_wav, src2_wav, _ = data.next_wavs(TrainConfig.SECONDS, TrainConfig.NUM_WAVFILE, step)
mixed_spec = to_spectrogram(mixed_wav)
mixed_mag = get_magnitude(mixed_spec)
src1_spec, src2_spec = to_spectrogram(src1_wav), to_spectrogram(src2_wav)
src1_mag, src2_mag = get_magnitude(src1_spec), get_magnitude(src2_spec)
src1_batch, _ = model.spec_to_batch(src1_mag)
src2_batch, _ = model.spec_to_batch(src2_mag)
mixed_batch, _ = model.spec_to_batch(mixed_mag)
# Initializae our callback.
#early_stopping_cb = EarlyStoppingCallback(val_acc_thresh=0.5)
l, _, summary = sess.run([loss_fn, optimizer, summary_op],
feed_dict=model.x_mixed: mixed_batch, model.y_src1: src1_batch,
model.y_src2: src2_batch)
loss.update(l)
print('step-\td_loss=:2.2f\tloss='.format(step, loss.diff * 100, loss.value))
writer.add_summary(summary, global_step=step)
# Save state
if step % TrainConfig.CKPT_STEP == 0:
tf.train.Saver().save(sess, TrainConfig.CKPT_PATH + '/checkpoint', global_step=step)
writer.close()
我有这个神经网络代码，可以将音乐与 .wav 文件中的声音分开。 如何引入提前停止算法来停止火车部分？我看到一些谈论 ValidationMonitor 的项目。有人可以帮我吗？
【问题讨论】：
从最新的 TensorFlow 文档（测试版）中，可以使用自定义回调实现提前停止。 tensorflow.org/beta/guide/keras/… 好吧，我提供的链接直接指向一个示例回调类，EarlyStoppingAtMinLoss。该类的一个实例可以在训练期间作为回调传递给模型，并在训练期间用于在损失停止减少时提前停止。该示例给出了类的实现，以及在训练期间如何使用它。此外，文档中还提到了一个额外的回调，“tf.keras.callbacks.EarlyStopping 提供了更完整和通用的实现”。这是回调：tensorflow.org/versions/r2.0/api_docs/python/tf/keras/callbacks/…
【参考方案1】：
这是我的早期停止的实现，你可以适应它：
可以在训练过程的某些阶段应用提前停止，例如在每个 epoch 结束时。具体来说;就我而言；我监控每个时期的测试（验证）损失，并且在 20 时期 (self.require_improvement= 20) 之后测试损失没有改善后，训练被中断。
您可以将最大时期设置为 10000 或 20000 或任何您想要的 (self.max_epochs = 10000)。
self.require_improvement= 20
self.max_epochs = 10000
这是我使用提前停止的训练功能：
def train(self):
# training data
train_input = self.Normalize(self.x_train)
train_output = self.y_train.copy()
#===============
save_sess=self.sess # this used to compare the result of previous sess with actual one
# ===============
#costs history :
costs = []
costs_inter=[]
# =================
#for early stopping :
best_cost=1000000
stop = False
last_improvement=0
# ================
n_samples = train_input.shape[0] # size of the training set
# ===============
#train the mini_batches model using the early stopping criteria
epoch = 0
while epoch < self.max_epochs and stop == False:
#train the model on the traning set by mini batches
#suffle then split the training set to mini-batches of size self.batch_size
seq =list(range(n_samples))
random.shuffle(seq)
mini_batches = [
seq[k:k+self.batch_size]
for k in range(0,n_samples, self.batch_size)
]
avg_cost = 0. # The average cost of mini_batches
step= 0
for sample in mini_batches:
batch_x = x_train.iloc[sample, :]
batch_y =train_output.iloc[sample, :]
batch_y = np.array(batch_y).flatten()
feed_dict=self.X: batch_x,self.Y:batch_y, self.is_train:True
_, cost,acc=self.sess.run([self.train_step, self.loss_, self.accuracy_],feed_dict=feed_dict)
avg_cost += cost *len(sample)/n_samples
print('epoch[] step [] train -- loss : , accuracy : '.format(epoch,step, avg_cost, acc))
step += 100
#cost history since the last best cost
costs_inter.append(avg_cost)
#early stopping based on the validation set/ max_steps_without_decrease of the loss value : require_improvement
if avg_cost < best_cost:
save_sess= self.sess # save session
best_cost = avg_cost
costs +=costs_inter # costs history of the validatio set
last_improvement = 0
costs_inter= []
else:
last_improvement +=1
if last_improvement > self.require_improvement:
print("No improvement found during the ( self.require_improvement) last iterations, stopping optimization.")
# Break out from the loop.
stop = True
self.sess=save_sess # restore session with the best cost
## Run validation after every epoch :
print('---------------------------------------------------------')
self.y_validation = np.array(self.y_validation).flatten()
loss_valid, acc_valid = self.sess.run([self.loss_,self.accuracy_],
feed_dict=self.X: self.x_validation, self.Y: self.y_validation,self.is_train: True)
print("Epoch: 0, validation loss: 1:.2f, validation accuracy: 2:.01%".format(epoch + 1, loss_valid, acc_valid))
print('---------------------------------------------------------')
epoch +=1
我们可以在这里恢复重要的代码：
def train(self):
...
#costs history :
costs = []
costs_inter=[]
#for early stopping :
best_cost=1000000
stop = False
last_improvement=0
#train the mini_batches model using the early stopping criteria
epoch = 0
while epoch < self.max_epochs and stop == False:
...
for sample in mini_batches:
...
#cost history since the last best cost
costs_inter.append(avg_cost)
#early stopping based on the validation set/ max_steps_without_decrease of the loss value : require_improvement
if avg_cost < best_cost:
save_sess= self.sess # save session
best_cost = avg_cost
costs +=costs_inter # costs history of the validatio set
last_improvement = 0
costs_inter= []
else:
last_improvement +=1
if last_improvement > self.require_improvement:
print("No improvement found during the ( self.require_improvement) last iterations, stopping optimization.")
# Break out from the loop.
stop = True
self.sess=save_sess # restore session with the best cost
...
epoch +=1
希望它能帮助某人:)。
【讨论】：
对不起这个愚蠢的问题，但是你如何调用火车功能？如何在调用时更改 self 参数？【参考方案2】：ValidationMonitor 被标记为已弃用。不推荐。但你仍然可以使用它。 这是一个如何创建的示例：
validation_monitor = monitors.ValidationMonitor(
input_fn=functools.partial(input_fn, subset="evaluation"),
eval_steps=128,
every_n_steps=88,
early_stopping_metric="accuracy",
early_stopping_rounds = 1000
)
你可以自己实现，这是我的实现：
if (loss_value < self.best_loss):
self.stopping_step = 0
self.best_loss = loss_value
else:
self.stopping_step += 1
if self.stopping_step >= FLAGS.early_stopping_step:
self.should_stop = True
print("Early stopping is trigger at step: loss:".format(global_step,loss_value))
run_context.request_stop()
【讨论】：
【参考方案3】：从 TensorFlow 版本开始，r1.10 早期停止钩子可用于 early_stopping.py 中的估算器 API（请参阅 github）。
例如tf.contrib.estimator.stop_if_no_decrease_hook（见docs）
【讨论】：
【参考方案4】：对于带有tf.keras 的自定义训练循环，您可以这样实现：
def main(early_stopping, epochs=50):
loss_history = deque(maxlen=early_stopping + 1)
for epoch in range(epochs):
fit(epoch)
loss_history.append(test_loss.result().numpy())
if len(loss_history) > early_stopping:
if loss_history.popleft() < min(loss_history):
print(f'\nEarly stopping. No validation loss '
f'improvement in early_stopping epochs.')
break
在每个 epoch 结束时，验证损失都会在 collections.deque 中重新计算。假设early_stopping 设置为 3。每个 epoch，最后 4 个损失与最后三个损失进行比较。如果这3个损失没有改善，则循环中断。
这里是完整的代码：
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
import tensorflow_datasets as tfds
import tensorflow as tf
from collections import deque
data, info = tfds.load('iris', split='train',
as_supervised=True,
shuffle_files=True,
with_info=True)
dataset = data.shuffle(info.splits['train'].num_examples)
train_dataset = dataset.take(120).batch(4)
test_dataset = dataset.skip(120).take(30).batch(4)
model = tf.keras.Sequential([
tf.keras.layers.Dense(8, activation='relu'),
tf.keras.layers.Dense(16, activation='relu'),
tf.keras.layers.Dense(32, activation='relu'),
tf.keras.layers.Dense(info.features['label'].num_classes, activation='softmax')
])
loss_object = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
train_loss = tf.keras.metrics.Mean()
test_loss = tf.keras.metrics.Mean()
train_acc = tf.keras.metrics.SparseCategoricalAccuracy()
test_acc = tf.keras.metrics.SparseCategoricalAccuracy()
opt = tf.keras.optimizers.Adam(learning_rate=1e-3)
@tf.function
def train_step(inputs, labels):
with tf.GradientTape() as tape:
logits = model(inputs)
loss = loss_object(labels, logits)
gradients = tape.gradient(loss, model.trainable_variables)
opt.apply_gradients(zip(gradients, model.trainable_variables))
train_loss(loss)
train_acc(labels, logits)
@tf.function
def test_step(inputs, labels):
logits = model(inputs)
loss = loss_object(labels, logits)
test_loss(loss)
test_acc(labels, logits)
def fit(epoch):
template = 'Epoch :>2 Train Loss :.4f Test Loss :.4f ' \
'Train Acc :.2% Test Acc :.2%'
train_loss.reset_states()
test_loss.reset_states()
train_acc.reset_states()
test_acc.reset_states()
for X_train, y_train in train_dataset:
train_step(X_train, y_train)
for X_test, y_test in test_dataset:
test_step(X_test, y_test)
print(template.format(
epoch + 1,
train_loss.result(),
test_loss.result(),
train_acc.result(),
test_acc.result()
))
def main(early_stopping, epochs=50):
loss_history = deque(maxlen=early_stopping + 1)
for epoch in range(epochs):
fit(epoch)
loss_history.append(test_loss.result().numpy())
if len(loss_history) > early_stopping:
if loss_history.popleft() < min(loss_history):
print(f'\nEarly stopping. No validation loss '
f'improvement in early_stopping epochs.')
break
if __name__ == '__main__':
main(epochs=100, early_stopping=3)
这是输出：
Epoch 1 Train Loss 1.0368 Test Loss 0.9507 Train Acc 66.67% Test Acc 76.67%
Epoch 2 Train Loss 1.0013 Test Loss 0.9673 Train Acc 65.83% Test Acc 70.00%
Epoch 3 Train Loss 0.9582 Test Loss 1.0055 Train Acc 64.17% Test Acc 56.67%
Epoch 4 Train Loss 0.9116 Test Loss 0.8510 Train Acc 63.33% Test Acc 70.00%
Epoch 5 Train Loss 0.8401 Test Loss 0.8632 Train Acc 67.50% Test Acc 76.67%
Epoch 6 Train Loss 0.8114 Test Loss 0.7535 Train Acc 72.50% Test Acc 80.00%
Epoch 7 Train Loss 0.8105 Test Loss 0.8240 Train Acc 68.33% Test Acc 80.00%
Epoch 8 Train Loss 0.7956 Test Loss 0.7855 Train Acc 81.67% Test Acc 93.33%
Epoch 9 Train Loss 0.7740 Test Loss 0.8094 Train Acc 89.17% Test Acc 73.33%
Early stopping. No validation loss improvement in 3 epochs.
如您所见，最后一个最佳验证损失是在 epoch 6，然后是 3 次损失，没有任何改善。然后循环因此中断。
【讨论】：
以上是关于如何在张量流中实现提前停止的主要内容，如果未能解决你的问题，请参考以下文章
如何使用 sparse_softmax_cross_entropy_with_logits 在张量流中实现加权交叉熵损失