我正在使用Python（3.6.7）中的TensorFlow（1.12.0）编写基本的神经网络。
一切似乎都可以单独运行，但是运行model()时出现此错误消息。
F ./tensorflow/core/framework/tensor.h:663] Check failed: new_num_elements == NumElements() (96 vs. 60000)
Aborted
据我所知，我在代码中的任何地方都没有使用数字96。
我在下面提供了我的代码：
model.py：
import numpy as np
import tensorflow as tf
from preprocessing import getDataset
from parameter import initializeParameters
from forward import forwardPropagation
from cost import computeCost
def model(numberOfIterations = 10000):
"""
Architecture:
W1, b1 W2, b2 W3, b3 W4, b4
X(numberOfFeatures, numberOfImages) ---> L1(25, numberOfImages) ---> L2(25, numberOfImages) ---> L3(10, numberOfImages) ---> H(10,numberOFImages)
relu relu relu sigmoid
W1(25, numberOfFeatures)
b1(25,1)
W2(25, 25)
b2(25, 1)
W3(10, 25)
b3(10, 1)
W4(10, 10)
b4(10, 1)
"""
# get all dataset information
data = getDataset()
(numberOfFeatures, numberOfImages) = data["X_train"].get_shape()
# initalize parameters whilst breaking symmetry
parameters = initializeParameters(numberOfFeatures, numberOfImages)
# construct forward propagation graph
(hypothesis, Z4) = forwardPropagation(data["X_train"], parameters)
# define cost tensor
cost = computeCost(Z4, data["Y_train"])
# create optimizer
optimizer = tf.train.AdamOptimizer().minimize(cost)
# define session object
with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
# train network
for iteration in range(numberOfIterations):
sess.run(optimizer)
if iteration % 10 == 0:
print("After iteration", iteration, "cost =", sess.run(cost))
model(numberOfIterations = 50)
preprocessing.py：
import numpy as np
import tensorflow as tf
#import matplotlib.pyplot as plt
def getDataset(datasetInfo = True, exampleImage = False):
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
numberOfImages_train = np.shape(x_train)[0]
numberOfImages_test = np.shape(x_test)[0]
imageDimension = (np.shape(x_train)[1], np.shape(x_train)[2])
if datasetInfo:
print("Number of training data images:", numberOfImages_train)
print("Number of testing data images:", numberOfImages_test)
print("Image dimensions:", imageDimension)
# if exampleImage:
# randIndex = np.random.randint(0, numberOfImages)
# plt.imshow(X_mat[:,:,:, randIndex])
data = {
"X_train": tf.convert_to_tensor(np.reshape(x_train, (numberOfImages_train, -1)).T, dtype = tf.float64),
"Y_train": tf.convert_to_tensor(y_train.T),
"X_test": tf.convert_to_tensor(np.reshape(x_test, (numberOfImages_test, -1)).T, dtype = tf.float64),
"Y_test": tf.convert_to_tensor(y_test.T)
}
return data
parameter.py：
import numpy as np
import tensorflow as tf
def initializeParameters(numberOfFeatures, numberOfImages):
W1 = tf.Variable(np.random.randn(25, numberOfFeatures) )
b1 = tf.Variable(np.zeros((25,1)))
W2 = tf.Variable(np.random.randn(25, 25))
b2 = tf.Variable(np.zeros((25,1)))
W3 = tf.Variable(np.random.randn(10, 25))
b3 = tf.Variable(np.zeros((10,1)))
W4 = tf.Variable(np.random.randn(10, 10))
b4 = tf.Variable(np.zeros((10,1)))
parameters = {
"W1": W1,
"b1": b1,
"W2": W2,
"b2": b2,
"W3": W3,
"b3": b3,
"W4": W4,
"b4": b4
}
return parameters
forward.py：
import numpy as np
import tensorflow as tf
def forwardPropagation(X, parameters):
Z1 = tf.add( tf.matmul(parameters["W1"], X) , parameters["b1"])
A1 = tf.nn.relu(Z1)
Z2 = tf.add( tf.matmul(parameters["W2"],A1) , parameters["b2"])
A2 = tf.nn.relu(Z2)
Z3 = tf.add( tf.matmul(parameters["W3"],A2) , parameters["b3"])
A3 = tf.nn.relu(Z3)
Z4 = tf.add( tf.matmul(parameters["W4"],A3) , parameters["b4"])
hypothesis = tf.sigmoid(Z4)
return (hypothesis, Z4)
cost.py：
import numpy as np
import tensorflow as tf
def computeCost(hypothesis, labels):
onehotlabels = tf.cast(tf.one_hot(labels, depth = 10, axis = 0), tf.float64)
losses = tf.nn.sigmoid_cross_entropy_with_logits(labels = onehotlabels, logits = hypothesis)
cost = tf.reduce_mean(losses)
return cost
完整输出：
Number of training data images: 60000
Number of testing data images: 10000
Image dimensions: (28, 28)
2018-12-03 12:35:50.912208: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2018-12-03 12:35:51.221523: F ./tensorflow/core/framework/tensor.h:663] Check failed: new_num_elements == NumElements() (96 vs. 60000)
Aborted
答案 0 :(得分：1)
您要喂养的批量大小为60,000，这非常大。我重现了该错误，但当我将训练图像的数量限制为255个时，该错误就消失了。 Tensorflow具有非常丰富的类和方法来处理此类问题。
对方法进行以下更改。 从getDataset（）方法仅返回numpy数组
def getDataset(datasetInfo = True, exampleImage = False):
# body of the method
data = {
"X_train": np.reshape(x_train, (numberOfImages_train, -1)).astype(np.float64),
"Y_train": y_train,
"X_test": np.reshape(x_test, (numberOfImages_test, -1)).astype(np.float64),
"Y_test": y_test
}
return data
内部模型方法进行以下更改
# Just to match the shape of X_train returned from getDataset method
(numberOfImages , numberOfFeatures ) = data["X_train"].shape
# Create iterator that gives you batch size of sample for at each batch for both training and test cases
train_data = tf.data.Dataset.from_tensor_slices((data["X_train"],data["Y_train"])).repeat().batch(BATCH_SIZE)
test_data = tf.data.Dataset.from_tensor_slices((data["X_test"],data["Y_test"])).repeat().batch(BATCH_SIZE)
train_itr = train_data.make_one_shot_iterator()
x_train, y_train = train_itr.get_next()
x_train = tf.transpose(x_train)
现在输入这些x_train和y_train来建立网络。
# initalize parameters whilst breaking symmetry
parameters = initializeParameters(numberOfFeatures, numberOfImages)
# construct forward propagation graph
(hypothesis, Z4) = forwardPropagation(x_train, parameters)
# define cost tensor
cost = computeCost(Z4, y_train)
我已经为getDataset和model方法提供了完整的代码
def getDataset(datasetInfo = True, exampleImage = False):
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
numberOfImages_train = np.shape(x_train)[0]
numberOfImages_test = np.shape(x_test)[0]
imageDimension = (np.shape(x_train)[1], np.shape(x_train)[2])
if datasetInfo:
print("Number of training data images:", numberOfImages_train)
print("Number of testing data images:", numberOfImages_test)
print("Image dimensions:", imageDimension)
# if exampleImage:
# randIndex = np.random.randint(0, numberOfImages)
# plt.imshow(X_mat[:,:,:, randIndex])
data = {
"X_train": np.reshape(x_train, (numberOfImages_train, -1)).astype(np.float64),
"Y_train": y_train,
"X_test": np.reshape(x_test, (numberOfImages_test, -1)).astype(np.float64),
"Y_test": y_test
}
return data
def model(numberOfIterations = 10000, BATCH_SIZE=32):
"""
Architecture:
W1, b1 W2, b2 W3, b3 W4, b4
X(numberOfFeatures, numberOfImages) ---> L1(25, numberOfImages) ---> L2(25, numberOfImages) ---> L3(10, numberOfImages) ---> H(10,numberOFImages)
relu relu relu sigmoid
W1(25, numberOfFeatures)
b1(25,1)
W2(25, 25)
b2(25, 1)
W3(10, 25)
b3(10, 1)
W4(10, 10)
b4(10, 1)
"""
# get all dataset information
data = getDataset()
(numberOfImages , numberOfFeatures ) = data["X_train"].shape
train_data = tf.data.Dataset.from_tensor_slices((data["X_train"],data["Y_train"])).repeat().batch(BATCH_SIZE)
test_data = tf.data.Dataset.from_tensor_slices((data["X_test"],data["Y_test"])).repeat().batch(BATCH_SIZE)
train_itr = train_data.make_one_shot_iterator()
x_train, y_train = train_itr.get_next()
x_train = tf.transpose(x_train)
# initalize parameters whilst breaking symmetry
parameters = initializeParameters(numberOfFeatures, numberOfImages)
# construct forward propagation graph
(hypothesis, Z4) = forwardPropagation(x_train, parameters)
# define cost tensor
cost = computeCost(Z4, y_train)
# create optimizer
optimizer = tf.train.AdamOptimizer().minimize(cost)
# define session object
with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
# train network
for iteration in range(numberOfIterations):
sess.run(optimizer)
if iteration % 10 == 0:
print("After iteration", iteration, "cost =", sess.run(cost))
model(numberOfIterations = 50)