[英]Dynamic switching of dropout in Keras/Tensorflow
我正在 Tensorflow 中构建一个强化学习算法,我希望能够在一次调用session.run()
动态关闭然后打开 dropout。
理由:我需要(1)做一个没有辍学的前向传递来计算目标; (2) 对生成的目标进行训练。 如果我在对session.run()
不同调用中执行这两个步骤,则一切正常。 但我想通过一次调用session.run()
(使用tf.stop_gradients(targets)
)来完成。
在尝试了几个没有取得很大成功的解决方案后,我找到了一个解决方案,我用一个变量替换了 Keras 使用的learning_phase占位符(因为占位符是张量并且不允许赋值)并使用自定义层将该变量设置为 True 或假如所愿。 此解决方案显示在下面的代码中。 分别获取m1
或m2
的值(例如,运行sess.run(m1, feed_dict={ph:np.ones((1,1))})
按预期工作,没有错误。但是,获取的值m3
,或同时获取m1
和m2
的值,有时有效,有时无效(并且错误消息不提供信息)。
你知道我做错了什么或做我想做的更好的方法吗?
编辑:代码显示了一个玩具示例。 实际上,我只有一个模型,我需要运行两次前向传球(一次关闭退出,另一次打开退出)和一次向后传球。 我想在不返回 python 的情况下完成所有这一切。
from tensorflow.keras.layers import Dropout, Dense, Input, Layer
from tensorflow.python.keras import backend as K
from tensorflow.keras import Model
import tensorflow as tf
import numpy as np
class DropoutSwitchLayer(Layer):
def __init__(self, stateful=True, **kwargs):
self.stateful = stateful
self.supports_masking = True
super(DropoutSwitchLayer, self).__init__(**kwargs)
def build(self, input_shape):
self.lph = tf.Variable(True, dtype=tf.bool, name="lph", trainable=False)
K._GRAPH_LEARNING_PHASES[tf.get_default_graph()] = self.lph
super(DropoutSwitchLayer, self).build(input_shape)
def call(self, inputs, mask=None):
data_input, training = inputs
op = self.lph.assign(training[0], use_locking=True)
# ugly trick here to make the layer work
data_input = data_input + tf.multiply(tf.cast(op, dtype=tf.float32), 0.0)
return data_input
def compute_output_shape(self, input_shape):
return input_shape[0]
dropout_on = np.array([True], dtype=np.bool)
dropout_off = np.array([False], dtype=np.bool)
input_ph = tf.placeholder(tf.float32, shape=(None, 1))
drop = Input(shape=(), dtype=tf.bool)
input = Input(shape=(1,))
h = DropoutSwitchLayer()([input, drop])
h = Dense(1)(h)
h = Dropout(0.5)(h)
o = Dense(1)(h)
m = Model(inputs=[input, drop], outputs=o)
m1 = m([input_ph, dropout_on])
m2 = m([input_ph, dropout_off])
m3 = m([m2, dropout_on])
sess = tf.Session()
K.set_session(sess)
sess.run(tf.global_variables_initializer())
编辑 2: Daniel Möller 下面的解决方案在使用Dropout
层时有效,但如果在LSTM
层内使用LSTM
呢?
input = Input(shape=(1,))
h = Dense(1)(input)
h = RepeatVector(2)(h)
h = LSTM(1, dropout=0.5, recurrent_dropout=0.5)(h)
o = Dense(1)(h)
为什么不制作一个单一的连续模型?
#layers
inputs = Input(shape(1,))
dense1 = Dense(1)
dense2 = Dense(1)
#no drop pass:
h = dense1(inputs)
o = dense2(h)
#optionally:
o = Lambda(lambda x: K.stop_gradient(x))(o)
#drop pass:
h = dense1(o)
h = Dropout(.5)(h)
h = dense2(h)
modelOnlyFinalOutput = Model(inputs,h)
modelOnlyNonDrop = Model(inputs,o)
modelBothOutputs = Model(inputs, [o,h])
选择一项进行培训:
model.fit(x_train,y_train) #where y_train = [targets1, targets2] if using both outputs
事实证明,Keras 支持开箱即用的我想做的事情。 在调用Dropout/LSTM 层时使用训练参数,结合 Daniel Möller 构建模型的方法(谢谢!),可以解决问题。
在下面的代码中(只是一个玩具示例), o1
和o3
应该等于并且不同于o2
from tensorflow.keras.layers import Dropout, Dense, Input, Lambda, Layer, Add, RepeatVector, LSTM
from tensorflow.python.keras import backend as K
from tensorflow.keras import Model
import tensorflow as tf
import numpy as np
repeat = RepeatVector(2)
lstm = LSTM(1, dropout=0.5, recurrent_dropout=0.5)
#Forward pass with dropout disabled
next_state = tf.placeholder(tf.float32, shape=(None, 1), name='next_state')
h = repeat(next_state)
# Use training to disable dropout
o1 = lstm(h, training=False)
target1 = tf.stop_gradient(o1)
#Forward pass with dropout enabled
state = tf.placeholder(tf.float32, shape=(None, 1), name='state')
h = repeat(state)
o2 = lstm(h, training=True)
target2 = tf.stop_gradient(o2)
#Forward pass with dropout disabled
ph3 = tf.placeholder(tf.float32, shape=(None, 1), name='ph3')
h = repeat(ph3)
o3 = lstm(h, training=False)
loss = target1 + target2 - o3
opt = tf.train.GradientDescentOptimizer(0.1)
train = opt.minimize(loss)
sess = tf.Session()
K.set_session(sess)
sess.run(tf.global_variables_initializer())
data = np.ones((1,1))
sess.run([o1, o2, o3], feed_dict={next_state:data, state:data, ph3:data})
这个怎么样 :
class CustomDropout(tf.keras.layers.Layer):
def __init__(self):
super(CustomDropout, self).__init__()
self.dropout1= Dropout(0.5)
self.dropout2= Dropout(0.1)
def call(self, inputs):
if xxx:
return self.dropout1(inputs)
else:
return self.dropout2(inputs)
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.