我正在尝试使用CSV输入在Tensorflow中实现MNIST图像分类。准确度非常低，约为10％。我在输出端使用3个完全连接的层和softmax熵。我将mnist数据集从here转换为csv，以使用以下代码作为输入：
import idx2numpy
import numpy as np
# Reading
filename = 'train-images'
flext1 = '.idx3-ubyte'
flext2 = '.csv'
ndarr = idx2numpy.convert_from_file(filename+flext1).reshape(60000, 784)
np.savetxt(filename,ndarr,delimiter=',')
然后我对标签进行了单热编码，如下所示：
import tensorflow as tf
import numpy as np
import pandas as pd
trainimages=pd.read_csv("train-images.csv",header=None,delimiter=',').values
trainlabels=pd.read_csv("train-labels.csv",header=None,delimiter=',').values
testimages=pd.read_csv("t10k-images.csv",header=None,delimiter=',').values
testlabels=pd.read_csv("t10k-labels.csv",header=None,delimiter=',').values
label_ph=tf.placeholder(tf.int64,shape=(None,1))
reshapeval=tf.Variable(100)
onehotencoding=tf.one_hot(label_ph,depth=10)
datalabels=tf.reshape(onehotencoding,shape=(reshapeval,10))
with tf.Session() as sess:
onehottrainlabels=sess.run(datalabels,feed_dict={label_ph:trainlabels,reshapeval:trainlabels.shape[0]})
onehottestlabels=sess.run(datalabels,feed_dict={label_ph:testlabels,reshapeval:testlabels.shape[0]})
print("datalabels:",onehottrainlabels)
print("datalabels:",onehottestlabels)
#sess.run(datalabels)
np.savetxt("mnist_train_labels_onehot.csv",onehottrainlabels,delimiter=',')
np.savetxt("mnist_test_labels_onehot.csv",onehottestlabels,delimiter=',')
我尝试使用如上所示生成的csv文件来尝试和训练图像分类模型，如下所示：
import tensorflow as tf
import pandas as pd
import numpy as np
trainimages=pd.read_csv("train-images.csv",header=None,delimiter=',').values
trainlabels=pd.read_csv("mnist_train_labels_onehot.csv",header=None,delimiter=',').values
testimages=pd.read_csv("t10k-images.csv",header=None,delimiter=',').values
testlabels=pd.read_csv("mnist_test_labels_onehot.csv",header=None,delimiter=',').values
#%%
train_num=len(trainimages)
n_epochs=400
BATCH_SIZE = 50
N_FEATURES = 785
n_inputs =28*28
n_hidden1=300
n_hidden2=100
n_outputs = 10
learning_rate=0.01
X=tf.placeholder(tf.float32,shape=(None,n_inputs),name='X')
y=tf.placeholder(tf.int64,shape=(None,10),name='y')
print("stage1")
with tf.name_scope("dnn"):
hidden1=tf.contrib.layers.fully_connected(X,n_hidden1,scope="hidden1")
hidden2=tf.contrib.layers.fully_connected(hidden1,n_hidden2,scope="hidden2")
logits=tf.contrib.layers.fully_connected(hidden2,n_outputs,scope="outputs",activation_fn=None)
with tf.name_scope("loss"):
xentropy =tf.nn.softmax_cross_entropy_with_logits(labels=y,logits=logits)
loss =tf.reduce_mean(xentropy,name="loss")
loss_summary=tf.summary.scalar("log_loss",loss)
with tf.name_scope("train"):
optimizer=tf.train.GradientDescentOptimizer(learning_rate)
training_op=optimizer.minimize(loss)
with tf.name_scope("eval"):
correct=tf.equal(tf.argmax(y, 1), tf.argmax(logits, 1))#tf.nn.in_top_k(logits,y,1)
accuracy=tf.reduce_mean(tf.cast(correct,tf.float32))
init =tf.global_variables_initializer()
saver =tf.train.Saver()
print("stage 2")
with tf.Session() as sess:
sess.run(init)
print("stage 3")
for epoch in range(n_epochs):
#print("stage 4")
for iteration in range(train_num//BATCH_SIZE):
#print("stage 5")
ind = np.random.choice(trainimages.shape[0],BATCH_SIZE)
x_train_batch = trainimages[ind]
y_train_batch = trainlabels[ind]
sess.run(training_op,feed_dict={X:x_train_batch,y:y_train_batch})
#print("stage 6")
if epoch%20==0:
#print(x_train_batch, y_train_batch)
acc_train=sess.run(accuracy,feed_dict={X:x_train_batch,y:y_train_batch})
acc_test=sess.run(accuracy,feed_dict={X:testimages,y:testlabels})
print(epoch, "Train accuracy:", acc_train,"Test accuracy:",acc_test)
输出精度从未增加约0.15。如果有人可以就我在这里做错的事情向我提出建议，我将非常感激。
作为参考，上面的图形结构基于Tensorflow动手机器学习书中给出的示例。以下是原始代码，其准确度为98％。当然我已经修改了一点代码以便能够接受csv输入。
import tensorflow as tf
import numpy as np
import pandas as pd
import random
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/")
#%%
tf.reset_default_graph()
print("I reset graph")
n_inputs =28*28
n_hidden1=300
n_hidden2=100
n_outputs = 10
learning_rate=0.01
X=tf.placeholder(tf.float32,shape=(None,n_inputs),name='X')
y=tf.placeholder(tf.int64,shape=(None),name='y')
with tf.name_scope("dnn"):
hidden1=tf.layers.dense(X,n_hidden1,name="hidden1",activation=tf.nn.relu)
hidden2=tf.layers.dense(hidden1,n_hidden2,name="hidden2",activation=tf.nn.relu)
logits=tf.layers.dense(hidden2,n_outputs,name="outputs")
with tf.name_scope("loss"):
xentropy =tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y,logits=logits)
loss =tf.reduce_mean(xentropy,name="loss")
#loss_summary=tf.summary.scalar("log_loss",loss)
with tf.name_scope("train"):
optimizer=tf.train.GradientDescentOptimizer(learning_rate)
training_op=optimizer.minimize(loss)
with tf.name_scope("eval"):
correct=tf.nn.in_top_k(logits,y,1)
accuracy=tf.reduce_mean(tf.cast(correct,tf.float32))
init =tf.global_variables_initializer()
saver =tf.train.Saver()
#==============================================================================
# from tensorflow.examples.tutorials.mnist import input_data
# mnist=input_data.read_data_sets("/tmp/data/")
#==============================================================================
n_epochs=400
batch_size=100
print("graph construction over i start train and test")
with tf.Session() as sess:
sess.run(init)
for epoch in range(n_epochs):
for iteration in range(mnist.train.num_examples//batch_size):
X_batch, y_batch = mnist.train.next_batch(batch_size)
#print(X_batch, y_batch)
sess.run(training_op,feed_dict={X:X_batch,y:y_batch})
print("iteration over")
if epoch%5==0:
acc_train=sess.run(accuracy,feed_dict={X:X_batch,y:y_batch})
acc_test=sess.run(accuracy,feed_dict={X: mnist.validation.images, y: mnist.validation.labels})
print(epoch, "Train accuracy:", acc_train,"Test accuracy:",acc_test)
save_path=saver.save(sess,"./my_model_final.ckpt")
其他信息：
在减少批量大小后，我打印了一些迭代的神经网络输出logits。在10次左右的迭代后，输出似乎饱和。我尝试将优化器更改为AdamOptimizer，但仍然无法提高准确性。下面显示了logits输出的示例。
iteration 7
Y_train: [[0 1 0 0 0 0 0 0 0 0]
[0 1 0 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 0 1 0]]
Logits: [[ 2.26569266e+17 3.05248736e+22 -2.50115512e+23 -4.39891557e+14
1.11553920e+23 3.07485097e+14 2.20944474e+17 4.47584162e+13
1.08036284e+23 4.83180834e+14]
[ 3.86803140e+17 8.75195504e+22 -7.17081147e+23 -2.68349530e+15
3.19824929e+23 8.67874780e+14 3.76970654e+17 4.52455091e+13
3.09735965e+23 3.87887019e+14]
[ 2.38161056e+17 3.84007123e+22 -3.14640962e+23 -6.05196997e+14
1.40332858e+23 3.16957379e+14 2.32241818e+17 4.01556080e+13
1.35906927e+23 4.00997575e+14]]
Softmax_out: [ 8.10290437e+22 2.32305379e+23 4.42593055e+21]
Accuracy: [False False False]
iteration 8
Y_train: [[1 0 0 0 0 0 0 0 0 0]
[0 1 0 0 0 0 0 0 0 0]
[0 0 0 1 0 0 0 0 0 0]]
Logits: [[ inf inf 1.66566635e+35 inf
-inf 1.03125672e+30 1.62685693e+35 6.78127837e+30
-4.95936289e+35 4.14895906e+30]
[ 1.38760797e+38 1.10565560e+38 3.23025301e+34 2.54545823e+38
-inf 1.99993203e+29 3.15498898e+34 1.31510396e+30
-9.61777107e+34 8.04613877e+29]
[ inf inf 2.54806070e+35 inf
-inf 1.57756931e+30 2.48869187e+35 1.03736909e+31
-7.58660917e+35 6.34688775e+30]]
Softmax_out: [ nan nan nan]
Accuracy: [ True False False]
iteration 9
Y_train: [[0 0 0 0 0 1 0 0 0 0]
[0 1 0 0 0 0 0 0 0 0]
[0 0 1 0 0 0 0 0 0 0]]
Logits: [[ nan nan nan nan nan nan nan nan nan nan]
[ nan nan nan nan nan nan nan nan nan nan]
[ nan nan nan nan nan nan nan nan nan nan]]
Softmax_out: [ nan nan nan]
Accuracy: [False False False]
答案 0 :(得分：0)
十类问题的准确率为10％是随机输出;我的猜测是，您拨打reshape的电话正在扰乱您的输入。例如，请考虑以下代码：
import tensorflow as tf
initial = tf.constant([[1, 2, 3, 4],[5, 6, 7, 8]])
reshaped = tf.reshape(initial, [4, 2])
with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
print(sess.run(initial))
print(sess.run(reshaped))
我将2x4张量转换为4x2张量，这听起来好像我正在调换它，但事实上输出是
[[1 2 3 4]
[5 6 7 8]]
[[1 2]
[3 4]
[5 6]
[7 8]]
我建议使用matplotlib来显示一些示例，看看它们是否仍然像数字。输出标签以确保它们仍然匹配也是一个好主意。
答案 1 :(得分：0)
我认为你应该像跟随一样构建层。
Convo有28输入+ relu + dropout（0.2）
Convo具有32个输入功能+ relu + dropout（0.2）
Convo具有32个输入功能+ relu + dropout（0.2）
flatten（）
密集或完全连接的图层，带有512个dendries + dropout（0.5）
密集或完全连接128个具有softmax功能的输出杂物
我使用自己的mnist类型数据集实现了99.7％的准确率。
检查出来：Text Classification using Keras
注意：我有36节课。 10位数和26个字母