在 tensorflow 2.1.0 中强制执行急切

Question

我是 TensorFlow 和深度学习的新手。

我创建了一个自定义损失函数，但似乎在自定义损失函数中，未启用急切执行。下面是我的自定义损失函数（它不起作用）：

def custom_error_finder(y_actual,y_pred):
    print(tf.executing_eagerly())
    count = 0
    qw = tf.py_function((y_actual).numpy())
    ya = ((y_actual[0].numpy()).decode())
    yp = ((y_pred[0].numpy()).decode())
    for i,j in ya,yp:
        if i!=j:
            count = count+1
    mse = pow(count,2)/len(ya)
    return mse

让我难过的是，在这个函数之外，每当我运行 print(tf.executing_eagerly()) 时，它都会返回 true，但在函数内部，它会返回 false。

我已经尝试了所有我能找到的修复方法：

-在model.compile()函数中传递run_eagerly = True

-在编译函数后添加model.run_eagerly() = True

-在损失函数中运行tf.compat.v1.enable_eager_execution() 至少强制执行一次急切。

以上修复均无效。

Answer 1

我能够重现该问题，如下所示。您可以从here 下载我在程序中使用的数据集。我在程序中添加了print("tf.executing_eagerly() Results") 语句来跟踪更改。

代码 -

%tensorflow_version 2.x
import tensorflow as tf
print(tf.__version__)
import numpy as np
from numpy import loadtxt
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from tensorflow.keras import backend as K

print("tf.executing_eagerly() Results")
print("Before loading dataset :",tf.executing_eagerly())

# load pima indians dataset
dataset = np.loadtxt("/content/pima-indians-diabetes.csv", delimiter=",")

# split into input (X) and output (Y) variables
X = dataset[:,0:8]
Y = dataset[:,8]

# define model
model = Sequential()
model.add(Dense(12, input_dim=8, activation='relu'))
model.add(Dense(8, activation='relu'))
model.add(Dense(1, activation='sigmoid'))

print("After building model :",tf.executing_eagerly())

def weighted_binary_crossentropy(y_true, y_pred):
  print("In loss function :",tf.executing_eagerly())
  return K.mean(K.binary_crossentropy(y_pred, y_true))

# compile model
model.compile(loss=weighted_binary_crossentropy, optimizer='adam', metrics=['accuracy'])

print("After compiling model :",tf.executing_eagerly())

# Fit the model
model.fit(X, Y, epochs=1, batch_size=150, verbose=0)

# evaluate the model
scores = model.evaluate(X, Y, verbose=0)
print("%s: %.2f%%" % (model.metrics_names[1], scores[1]*100))

输出 -

2.2.0
tf.executing_eagerly() Results
Before loading dataset : True
After building model : True
After compiling model : True
In loss function : False
In loss function : False
In loss function : False
accuracy: 34.90%

解决方案 - 根据documentation。它提到，

run_eagerly - 指示模型是否应该急切运行的可设置属性。 急切地运行意味着您的模型将像 Python 代码一样逐步运行。您的模型可能会运行得更慢，但通过单步调用单独的层调用，您应该可以更轻松地对其进行调试。 默认情况下，我们会尝试将您的模型编译为静态图以提供最佳执行性能。

如果我们用run_eagerly = True 参数修改model.compile，我们可以解决这个问题。下图为修改后的model.compile代码，

model.compile(loss=weighted_binary_crossentropy, run_eagerly = True, optimizer='adam', metrics=['accuracy'])

固定代码 -

%tensorflow_version 2.x
import tensorflow as tf
print(tf.__version__)
import numpy as np
from numpy import loadtxt
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from tensorflow.keras import backend as K

print("tf.executing_eagerly() Results")
print("Before loading dataset :",tf.executing_eagerly())

# load pima indians dataset
dataset = np.loadtxt("/content/pima-indians-diabetes.csv", delimiter=",")

# split into input (X) and output (Y) variables
X = dataset[:,0:8]
Y = dataset[:,8]

# define model
model = Sequential()
model.add(Dense(12, input_dim=8, activation='relu'))
model.add(Dense(8, activation='relu'))
model.add(Dense(1, activation='sigmoid'))

print("After building model :",tf.executing_eagerly())

def weighted_binary_crossentropy(y_true, y_pred):
  print("In loss function :",tf.executing_eagerly())
  return K.mean(K.binary_crossentropy(y_pred, y_true))

# compile model
model.compile(loss=weighted_binary_crossentropy, run_eagerly = True, optimizer='adam', metrics=['accuracy'])

print("After compiling model :",tf.executing_eagerly())

# Fit the model
model.fit(X, Y, epochs=1, batch_size=150, verbose=0)

# evaluate the model
scores = model.evaluate(X, Y, verbose=0)
print("%s: %.2f%%" % (model.metrics_names[1], scores[1]*100))

输出 -

2.2.0
tf.executing_eagerly() Results
Before loading dataset : True
After building model : True
After compiling model : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
In loss function : True
accuracy: 34.90%

希望这能回答您的问题。快乐学习。