TensorFlow 2.0是对1.x版本做了一次大的瘦身,Eager Execution默认开启,并且使用Keras作为默认高级API,
这些改进大大降低的TensorFlow使用难度。
本文主要记录了一次曲折的使用Keras+TensorFlow2.0的BatchNormalization的踩坑经历,这个坑差点要把TF2.0的新特性都毁灭殆尽,如果你在学习TF2.0的官方教程,不妨一观。
问题的产生
从教程[1]https://www.tensorflow.org/alpha/tutorials/images/transfer_learning?hl=zh-cn(讲述如何Transfer Learning)说起:
IMG_SHAPE = (IMG_SIZE, IMG_SIZE, 3)
# Create the base model from the pre-trained model MobileNet V2
base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE,
include_top=False,weights='imagenet')
model = tf.keras.Sequential([
base_model,
tf.keras.layers.GlobalAveragePooling2D(),
tf.keras.layers.Dense(NUM_CLASSES)
])
简单的代码我们就复用了MobileNetV2的结构创建了一个分类器模型,接着我们就可以调用Keras的接口去训练模型:
model.compile(optimizer=tf.keras.optimizers.RMSprop(lr=base_learning_rate),
loss='sparse_categorical_crossentropy',
metrics=['sparse_categorical_accuracy'])
model.summary()
history = model.fit(train_batches.repeat(),
epochs=20,
steps_per_epoch = steps_per_epoch,
validation_data=validation_batches.repeat(),
validation_steps=validation_steps)
输出的结果看,一起都很完美:
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
mobilenetv2_1.00_160 (Model) (None, 5, 5, 1280) 2257984
_________________________________________________________________
global_average_pooling2d (Gl (None, 1280) 0
_________________________________________________________________
dense (Dense) (None, 2) 1281
=================================================================
Total params: 2,259,265
Trainable params: 1,281
Non-trainable params: 2,257,984
_________________________________________________________________
Epoch 11/20
581/581 [==============================] - 134s 231ms/step - loss: 0.4208 - accuracy: 0.9484 - val_loss: 0.1907 - val_accuracy: 0.9812
Epoch 12/20
581/581 [==============================] - 114s 197ms/step - loss: 0.3359 - accuracy: 0.9570 - val_loss: 0.1835 - val_accuracy: 0.9844
Epoch 13/20
581/581 [==============================] - 116s 200ms/step - loss: 0.2930 - accuracy: 0.9650 - val_loss: 0.1505 - val_accuracy: 0.9844
Epoch 14/20
581/581 [==============================] - 114s 196ms/step - loss: 0.2561 - accuracy: 0.9701 - val_loss: 0.1575 - val_accuracy: 0.9859
Epoch 15/20
581/581 [==============================] - 119s 206ms/step - loss: 0.2302 - accuracy: 0.9715 - val_loss: 0.1600 - val_accuracy: 0.9812
Epoch 16/20
581/581 [==============================] - 115s 197ms/step - loss: 0.2134 - accuracy: 0.9747 - val_loss: 0.1407 - val_accuracy: 0.9828
Epoch 17/20
581/581 [==============================] - 115s 197ms/step - loss: 0.1546 - accuracy: 0.9813 - val_loss: 0.0944 - val_accuracy: 0.9828
Epoch 18/20
581/581 [==============================] - 116s 200ms/step - loss: 0.1636 - accuracy: 0.9794 - val_loss: 0.0947 - val_accuracy: 0.9844
Epoch 19/20
581/581 [==============================] - 115s 198ms/step - loss: 0.1356 - accuracy: 0.9823 - val_loss: 0.1169 - val_accuracy: 0.9828
Epoch 20/20
581/581 [==============================] - 116s 199ms/step - loss: 0.1243 - accuracy: 0.9849 - val_loss: 0.1121 - val_accuracy: 0.9875
然而这种写法还是不方便Debug,我们希望可以精细的控制迭代的过程,并能够看到中间结果,所以我们训练的过程改成了这样:
optimizer = tf.keras.optimizers.RMSprop(lr=base_learning_rate)
train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')
@tf.function
def train_cls_step(image, label):
with tf.GradientTape() as tape:
predictions = model(image)
loss = tf.keras.losses.SparseCategoricalCrossentropy()(label, predictions)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
train_accuracy(label, predictions)
for images, labels in train_batches:
train_cls_step(images,labels)
重新训练后,结果依然很完美!
但是,这时候我们想对比一下Finetune和重头开始训练的差别,所以把构建模型的代码改成了这样:
base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE,
include_top=False,weights=None)
使得模型的权重随机生成,这时候训练结果就开始抽风了,Loss不下降,Accuracy稳定在50%附近游荡:
Step #10: loss=0.6937199831008911 acc=46.5625%
Step #20: loss=0.6932525634765625 acc=47.8125%
Step #30: loss=0.699873685836792 acc=49.16666793823242%
Step #40: loss=0.6910845041275024 acc=49.6875%
Step #50: loss=0.6935917139053345 acc=50.0625%
Step #60: loss=0.6965731382369995 acc=49.6875%
Step #70: loss=0.6949992179870605 acc=49.19642639160156%
Step #80: loss=0.6942993402481079 acc=49.84375%
Step #90: loss=0.6933775544166565 acc=49.65277862548828%
Step #100: loss=0.6928421258926392 acc=49.5%
Step #110: loss=0.6883170008659363 acc=49.54545593261719%
Step #120: loss=0.695658802986145 acc=49.453125%
Step #130: loss=0.6875559091567993 acc=49.61538314819336%
Step #140: loss=0.6851695775985718 acc=49.86606979370117%
Step #150: loss=0.6978713274002075 acc=49.875%
Step #160: loss=0.7165156602859497 acc=50.0%
Step #170: loss=0.6945627331733704 acc=49.797794342041016%
Step #180: loss=0.6936900615692139 acc=49.9305534362793%
Step #190: loss=0.6938323974609375 acc=49.83552551269531%
Step #200: loss=0.7030564546585083 acc=49.828125%
Step #210: loss=0.6926192045211792 acc=49.76190185546875%
Step #220: loss=0.6932414770126343 acc=49.786930084228516%
Step #230: loss=0.6924526691436768 acc=49.82337188720703%
Step #240: loss=0.6882281303405762 acc=49.869789123535156%
Step #250: loss=0.6877702474594116 acc=49.86249923706055%
Step #260: loss=0.6933954954147339 acc=49.77163314819336%
Step #270: loss=0.6944763660430908 acc=49.75694274902344%
Step #280: loss=0.6945018768310547 acc=49.49776840209961%
我们将predictions的结果打印出来,发现batch内每个输出都是一模一样的:
0 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32)
1 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32)
2 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32)
3 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32)
4 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32)
5 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32)
6 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32)
7 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32)
8 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32)
9 = tf.Tensor([0.51352817 0.48647183], shape=(2,), dtype=float32)
只是修改了初始权重,为何会产生这样的结果?
问题排查
实验1
是不是训练不够充分,或者learning rate设置的不合适?
经过几轮调整,发现无论训练多久,learning rate变大变小,都无法改变这种结果
实验2
既然是权重的问题,是不是权重随机初始化的有问题,把初始权重拿出来统计了一下,一切正常
实验3
这种问题根据之前的经验,在导出Inference模型的时候BatchNormalization没有处理好会出现这种一个batch内所有结果都一样的问题。但是如何解释训练的时候为什么会出现这个问题?而且为什么Finetue不会出现问题呢?只是改了权重的初始值而已呀
按照这个方向去Google的一番,发现了Keras的BatchNormalization确实有很多issue,其中一个问题是在保存模型的是BatchNormalzation的moving mean和moving variance不会被保存[6]https://github.com/tensorflow/tensorflow/issues/16455,而另外一个issue提到问题就和我们问题有关系的了:
[2] https://github.com/tensorflow/tensorflow/issues/19643
[3] https://github.com/tensorflow/tensorflow/issues/23873
最后,这位作者找到了原因,并且总结在了这里:
[4] https://pgaleone.eu/tensorflow/keras/2019/01/19/keras-not-yet-interface-to-tensorflow/
实验3.1
改用model.fit的写法进行训练,在最初的几个epoch里面,我们发现好的一点的是training accuracy已经开始缓慢提升了,但是validation accuracy存在原来的问题。而且通过model.predict_on_batch()拿到中间结果,发现依然还是batch内输出都一样。
Epoch 1/20
581/581 [==============================] - 162s 279ms/step - loss: 0.6768 - sparse_categorical_accuracy: 0.6224 - val_loss: 0.6981 - val_sparse_categorical_accuracy: 0.4984
Epoch 2/20
581/581 [==============================] - 133s 228ms/step - loss: 0.4847 - sparse_categorical_accuracy: 0.7684 - val_loss: 0.6931 - val_sparse_categorical_accuracy: 0.5016
Epoch 3/20
581/581 [==============================] - 130s 223ms/step - loss: 0.3905 - sparse_categorical_accuracy: 0.8250 - val_loss: 0.6996 - val_sparse_categorical_accuracy: 0.4984
Epoch 4/20
581/581 [==============================] - 131s 225ms/step - loss: 0.3113 - sparse_categorical_accuracy: 0.8660 - val_loss: 0.6935 - val_sparse_categorical_accuracy: 0.5016
但是,随着训练的深入,结果出现了逆转,开始变得正常了(tf.function的写法是无论怎么训练都不会变化,幸好没有放弃治疗)(追加:其实这里还是有问题的,继续看后面,当时就觉得怪怪的,不应该收敛这么慢)
Epoch 18/20
581/581 [==============================] - 131s 226ms/step - loss: 0.0731 - sparse_categorical_accuracy: 0.9725 - val_loss: 1.4896 - val_sparse_categorical_accuracy: 0.8703
Epoch 19/20
581/581 [==============================] - 130s 225ms/step - loss: 0.0664 - sparse_categorical_accuracy: 0.9748 - val_loss: 0.6890 - val_sparse_categorical_accuracy: 0.9016
Epoch 20/20
581/581 [==============================] - 126s 217ms/step - loss: 0.0631 - sparse_categorical_accuracy: 0.9768 - val_loss: 1.0290 - val_sparse_categorical_accuracy: 0.9031
通多model.predict_on_batch()拿到的结果也和这个Accuracy也是一致的
实验3.2
通过上一个实验,我们验证了确实如果只通过Keras的API去训练,是正常。更深层的原因是什么呢?是不是BatchNomalization没有update moving mean和moving variance导致的呢?答案是Yes
我们分别在两中训练方法前后,打印 moving mean和moving variance的值:
def get_bn_vars(collection):
moving_mean, moving_variance = None, None for var in collection:
name = var.name.lower()
if "variance" in name:
moving_variance = var
if "mean" in name:
moving_mean = var
if moving_mean is not None and moving_variance is not None:
return moving_mean, moving_variance
raise ValueError("Unable to find moving mean and variance")
mean, variance = get_bn_vars(model.variables)
print(mean)
print(variance)
我们发现,确实如果使用model.fit()进行训练,mean和variance是在update的(虽然更新的速率看着有些奇怪),但是对于tf.function那种写法这两个值就没有被update
那这里我们也可以解释为什么Finetune不会出现问题了,因为imagenet训练的mean, variance已经是一个比较好的值了,即使不更新也可以正常使用
实验3.3
是不是改成[4]里面说的方法构建动态的Input_Shape的模型就OK了呢?
class MyModel(Model):
def __init__(self):
super(MyModel, self).__init__()
self.conv1 = Conv2D(32, 3, activation='relu')
self.batch_norm1=BatchNormalization()
self.flatten = Flatten()
self.d1 = Dense(128, activation='relu')
self.d2 = Dense(10, activation='softmax')
def call(self, x):
x = self.conv1(x)
x = self.batch_norm1(x)
x = self.flatten(x)
x = self.d1(x)
return self.d2(x)
model = MyModel()
#model.build((None,28,28,1))
model.summary()
@tf.functiondef train_step(image, label):
with tf.GradientTape() as tape:
predictions = model(image)
loss = loss_object(label, predictions)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
train_loss(loss)
train_accuracy(label, predictions)
模型如下:
Model: "my_model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) multiple 320
_________________________________________________________________
batch_normalization_v2 (Batc multiple 128
_________________________________________________________________
flatten (Flatten) multiple 0
_________________________________________________________________
dense (Dense) multiple 2769024
_________________________________________________________________
dense_1 (Dense) multiple 1290
=================================================================
Total params: 2,770,762
Trainable params: 2,770,698
Non-trainable params: 64
从Output Shape看,构建模型没问题
跑了一遍MINST,结果也很不错!
以防万一,我们同样测试了一下mean和variance是否被更新,然而结果出乎意料,并没有!
也就是说[4]里面说的方案在我们这里并不可行
实验3.4
既然我们定位问题是在BatchNormalization这里,所以就想到BatchNormalization的training和testing时候行为是不一致的,在testing的时候moving mean和variance是不需要update的,那么会不会是tf.function的这种写法并不会自动更改这个状态呢?
查看源码,发现BatchNormalization的call()存在一个training参数,而且默认是False
Call arguments:
inputs: Input tensor (of any rank).
training: Python boolean indicating whether the layer should behave in
training mode or in inference mode.
- `training=True`: The layer will normalize its inputs using the
mean and variance of the current batch of inputs.
- `training=False`: The layer will normalize its inputs using the
mean and variance of its moving statistics, learned during training.
所以,做了如下改进:
class MyModel(Model):
def __init__(self):
super(MyModel, self).__init__()
self.conv1 = Conv2D(32, 3, activation='relu')
self.batch_norm1=BatchNormalization()
self.flatten = Flatten()
self.d1 = Dense(128, activation='relu')
self.d2 = Dense(10, activation='softmax')
def call(self, x,training=True):
x = self.conv1(x)
x = self.batch_norm1(x,training=training)
x = self.flatten(x)
x = self.d1(x)
return self.d2(x)
model = MyModel()
#model.build((None,28,28,1))
model.summary()
@tf.functiondef train_step(image, label):
with tf.GradientTape() as tape:
predictions = model(image,training=True)
loss = loss_object(label, predictions)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
train_loss(loss)
train_accuracy(label, predictions)
@tf.functiondef test_step(image, label):
predictions = model(image,training=False)
t_loss = loss_object(label, predictions)
test_loss(t_loss)
test_accuracy(label, predictions)
结果显示,moving mean和variance开始更新啦,测试Accuracy也是符合预期
所以,我们可以确定问题的根源在于需要指定BatchNormalization是在training还是在testing!
实验3.5
3.4中方法虽然解决了我们的问题,但是它是使用构建Model的subclass的方式,而我们之前的MobileNetV2是基于更加灵活Keras Functional API构建的,由于无法控制call()函数的定义,没有办法灵活切换training和testing的状态,另外用Sequential的方式构建时也是一样。
[5]https://blog.keras.io/keras-as-a-simplified-interface-to-tensorflow-tutorial.html
[7]https://github.com/keras-team/keras/issues/7085
[8]https://github.com/keras-team/keras/issues/6752
从5[8]中,我了解到两个情况,
-
- tf.keras.backend.set_learning_phase()可以改变training和testing的状态;
-
- model.updates和layer.updates 存着old_value和new_value的Assign Op
所以我首先尝试:
tf.keras.backend.set_learning_phase(True)
结果,MobileNetV2构建的模型也可以正常工作了。
而且收敛的速度似乎比model.fit()还快了很多,结合之前model.fit()收敛慢的困惑,这里又增加的一个实验,在model.fit()的版本里面也加上这句话,发现同样收敛速度也变快了!1个epoch就能得到不错的结果了!
因此,这里又产生了一个问题model.fit()到底有没有设learning_phase状态?如果没有是怎么做moving mean和variance的update的?
第二个方法,由于教程中讲述的是如何在1.x的版本构建,而在eager execution模式下,似乎没有办法去run这些Assign Operation。仅做参考吧
update_ops = []
for assign_op in model.updates:
update_ops.append(assign_op))
#但是不知道拿到这些update_ops在eager execution模式下怎么处理呢?
结论
总结一下,我们从[4]找到了解决问题的启发点,但是最终证明[4]里面的问题和解决方法用到我们这里并不能真正解决问题,问题的关键还是在于Keras+TensorFlow2.0里面我们如何处理在training和testing状态下行为不一致的Layer;以及对于model.fit()和tf.funtion这两种训练方法的区别,最终来看model.fit()里面似乎包含很多诡异的行为。
最终的使用建议如下:
- 在使用model.fit()或者model.train_on_batch()这种Keras的API训练模型时,也推荐手动设置tf.keras.backend.set_learning_phase(True),可以加快收敛
- 如果使用eager execution这种方法,
- 1)使用构建Model的subclass,但是针对call()设置training的状态,对于BatchNoramlization,Dropout这样的Layer进行不同处理
- 2)使用Functional API或者Sequential的方式构建Model,设置tf.keras.backend.set_learning_phase(True),但是注意在testing的时候改变一下状态
最后,为什么TF 2.0的教程里面没有提及这些?默认你已经精通Keras了吗?[捂脸哭]
感谢
感谢柏涛 帆月 应知老师提供的帮助
[1]https://www.tensorflow.org/alpha/tutorials/images/transfer_learning?hl=zh-cn
[2] https://github.com/tensorflow/tensorflow/issues/19643
[3] https://github.com/tensorflow/tensorflow/issues/23873
[4] https://pgaleone.eu/tensorflow/keras/2019/01/19/keras-not-yet-interface-to-tensorflow/
[5]https://blog.keras.io/keras-as-a-simplified-interface-to-tensorflow-tutorial.html
[6]https://github.com/tensorflow/tensorflow/issues/16455
[7]https://github.com/keras-team/keras/issues/7085
[8]https://github.com/keras-team/keras/issues/6752