Tflearn 中的分类器质量差？
Poor quality classifier in Tflearn?
我是机器学习的新手，正在尝试 TFlearn，因为它很简单。
我正在尝试制作一个我觉得很有趣的基本分类器。
我的objective是训练系统预测一个点在哪个方向。
例如，如果我输入两个二维坐标 (50,50) 和 (51,51)，系统必须预测方向是 NE（东北）。
如果我输入 (50,50) 和 (49,49)，系统必须预测方向是 SW（西南）
输入： X1,Y1,X2,Y2,标签
输出： 0 到 8。对于 8 个方向。
所以这是我写的小代码，
from __future__ import print_function
import numpy as np
import tflearn
import tensorflow as tf
import time
from tflearn.data_utils import load_csv
#Sample input 50,50,51,51,5
data, labels = load_csv(filename, target_column=4,
categorical_labels=True, n_classes=8)
my_optimizer = tflearn.SGD(learning_rate=0.1)
net = tflearn.input_data(shape=[None, 4])
net = tflearn.fully_connected(net, 32) #input 4, output 32
net = tflearn.fully_connected(net, 32) #input 32, output 32
net = tflearn.fully_connected(net, 8, activation='softmax')
net = tflearn.regression(net,optimizer=my_optimizer)
model = tflearn.DNN(net)
model.fit(data, labels, n_epoch=100, batch_size=100000, show_metric=True)
model.save("direction-classifier.tfl")
我面临的问题是，即使在我通过了大约 4000 万个输入样本后，系统准确率仍低至 20%。
我将输入限制为 40-x-60 和 40-y-60
我无法理解我是否过度拟合了样本，因为在总共 4000 万个输入的整个训练期间准确率一直不高
为什么这个简单示例的准确性如此低？
编辑：
我降低了学习率并减小了批量大小。但是，结果仍然相同，但准确性很差。
我已经包含了前 25 个步骤的输出。
--
Training Step: 100000 | total loss: 6.33983 | time: 163.327s
| SGD | epoch: 001 | loss: 6.33983 - acc: 0.0663 -- iter: 999999/999999
--
Training Step: 200000 | total loss: 6.84055 | time: 161.981ss
| SGD | epoch: 002 | loss: 6.84055 - acc: 0.1568 -- iter: 999999/999999
--
Training Step: 300000 | total loss: 5.90203 | time: 158.853ss
| SGD | epoch: 003 | loss: 5.90203 - acc: 0.1426 -- iter: 999999/999999
--
Training Step: 400000 | total loss: 5.97782 | time: 157.607ss
| SGD | epoch: 004 | loss: 5.97782 - acc: 0.1465 -- iter: 999999/999999
--
Training Step: 500000 | total loss: 5.97215 | time: 155.929ss
| SGD | epoch: 005 | loss: 5.97215 - acc: 0.1234 -- iter: 999999/999999
--
Training Step: 600000 | total loss: 6.86967 | time: 157.299ss
| SGD | epoch: 006 | loss: 6.86967 - acc: 0.1230 -- iter: 999999/999999
--
Training Step: 700000 | total loss: 6.10330 | time: 158.137ss
| SGD | epoch: 007 | loss: 6.10330 - acc: 0.1242 -- iter: 999999/999999
--
Training Step: 800000 | total loss: 5.81901 | time: 157.464ss
| SGD | epoch: 008 | loss: 5.81901 - acc: 0.1464 -- iter: 999999/999999
--
Training Step: 900000 | total loss: 7.09744 | time: 157.486ss
| SGD | epoch: 009 | loss: 7.09744 - acc: 0.1359 -- iter: 999999/999999
--
Training Step: 1000000 | total loss: 7.19259 | time: 158.369s
| SGD | epoch: 010 | loss: 7.19259 - acc: 0.1248 -- iter: 999999/999999
--
Training Step: 1100000 | total loss: 5.60177 | time: 157.221ss
| SGD | epoch: 011 | loss: 5.60177 - acc: 0.1378 -- iter: 999999/999999
--
Training Step: 1200000 | total loss: 7.16676 | time: 158.607ss
| SGD | epoch: 012 | loss: 7.16676 - acc: 0.1210 -- iter: 999999/999999
--
Training Step: 1300000 | total loss: 6.19163 | time: 163.711ss
| SGD | epoch: 013 | loss: 6.19163 - acc: 0.1635 -- iter: 999999/999999
--
Training Step: 1400000 | total loss: 7.46101 | time: 162.091ss
| SGD | epoch: 014 | loss: 7.46101 - acc: 0.1216 -- iter: 999999/999999
--
Training Step: 1500000 | total loss: 7.78055 | time: 158.468ss
| SGD | epoch: 015 | loss: 7.78055 - acc: 0.1122 -- iter: 999999/999999
--
Training Step: 1600000 | total loss: 6.03101 | time: 158.251ss
| SGD | epoch: 016 | loss: 6.03101 - acc: 0.1103 -- iter: 999999/999999
--
Training Step: 1700000 | total loss: 5.59769 | time: 158.083ss
| SGD | epoch: 017 | loss: 5.59769 - acc: 0.1182 -- iter: 999999/999999
--
Training Step: 1800000 | total loss: 5.45591 | time: 158.088ss
| SGD | epoch: 018 | loss: 5.45591 - acc: 0.0868 -- iter: 999999/999999
--
Training Step: 1900000 | total loss: 6.54951 | time: 157.755ss
| SGD | epoch: 019 | loss: 6.54951 - acc: 0.1353 -- iter: 999999/999999
--
Training Step: 2000000 | total loss: 6.18566 | time: 157.408ss
| SGD | epoch: 020 | loss: 6.18566 - acc: 0.0551 -- iter: 999999/999999
--
Training Step: 2100000 | total loss: 4.95146 | time: 157.572ss
| SGD | epoch: 021 | loss: 4.95146 - acc: 0.1114 -- iter: 999999/999999
--
Training Step: 2200000 | total loss: 5.97208 | time: 157.279ss
| SGD | epoch: 022 | loss: 5.97208 - acc: 0.1277 -- iter: 999999/999999
--
Training Step: 2300000 | total loss: 6.75645 | time: 157.201ss
| SGD | epoch: 023 | loss: 6.75645 - acc: 0.1507 -- iter: 999999/999999
--
Training Step: 2400000 | total loss: 7.04119 | time: 157.346ss
| SGD | epoch: 024 | loss: 7.04119 - acc: 0.1512 -- iter: 999999/999999
--
Training Step: 2500000 | total loss: 5.95451 | time: 157.722ss
| SGD | epoch: 025 | loss: 5.95451 - acc: 0.1421 -- iter: 999999/999999
正如我在上面的评论中所讨论的，这里是使用 a MLP helper class I created 训练多层感知器 classifier 模型的代码。 class 使用 TensorFlow 实现并遵循 scikit-learn 拟合、预测、评分接口。
基本思想是生成一个随机的起点和终点，然后使用字典根据方向创建标签。我使用 np.unique 在生成的数据中查找 class 标签的数量，因为它可能会有所不同（某些方向可能会丢失）。当起点和终点相同时，我还包含了一个空字符串标签。
代码
使用下面的代码，我能够在某些 运行 上实现 100% 的交叉验证准确率。
将 numpy 导入为 np
来自 sklearn.model_selection 导入 ShuffleSplit
从 TFANN 导入 MLPC
#Dictionary to lookup direction ()
DM = {(-1, -1):'SW', (-1, 0):'W', (-1, 1):'NW', (0, 1):'N',
( 1, 1):'NE', ( 1, 0):'E', ( 1, -1):'SE', (0, -1):'S',
( 0, 0):''}
NR = 4096 #Number of rows in sample matrix
A1 = np.random.randint(40, 61, size = (NR, 2)) #Random starting point
A2 = np.random.randint(40, 61, size = (NR, 2)) #Random ending point
A = np.hstack([A1, A2]) #Concat start and end point as feature vector
#Create label from direction vector
Y = np.array([DM[(x, y)] for x, y in (A2 - A1).clip(-1, 1)])
NC = len(np.unique(Y)) #Number of classes
ss = ShuffleSplit(n_splits = 1)
trn, tst = next(ss.split(A)) #Make a train/test split for cross-validation
#%% Create and train Multi-Layer Perceptron for Classification (MLPC)
l = [4, 6, 6, NC] #Neuron counts in each layer
mlpc = MLPC(l, batchSize = 64, maxIter = 128, verbose = True)
mlpc.fit(A[trn], Y[trn])
s1 = mlpc.score(A[trn], Y[trn]) #Training accuracy
s2 = mlpc.score(A[tst], Y[tst]) #Testing accuracy
s3 = mlpc.score(A, Y) #Total accuracy
print('Trn: {:05f}\tTst: {:05f}\tAll: {:05f}'.format(s1, s2, s3))
结果
这是上述代码在我机器上的示例运行：
Iter 1 2.59423236 (Batch Size: 64)
Iter 2 2.25392553 (Batch Size: 64)
Iter 3 2.02569708 (Batch Size: 64)
...
Iter 12 1.53575111 (Batch Size: 64)
Iter 13 1.47963311 (Batch Size: 64)
Iter 14 1.42776408 (Batch Size: 64)
...
Iter 83 0.23911642 (Batch Size: 64)
Iter 84 0.22893350 (Batch Size: 64)
Iter 85 0.23644384 (Batch Size: 64)
...
Iter 94 0.21170238 (Batch Size: 64)
Iter 95 0.20718799 (Batch Size: 64)
Iter 96 0.21230888 (Batch Size: 64)
...
Iter 126 0.17334313 (Batch Size: 64)
Iter 127 0.16970796 (Batch Size: 64)
Iter 128 0.15931854 (Batch Size: 64)
Trn: 0.995659 Tst: 1.000000 All: 0.996094
原来是优化器导致了所有问题。移除自定义优化器后，损失开始正确下降，准确率提高到 99%
必须修改以下两行。
my_optimizer = tflearn.SGD(learning_rate=0.1)
net = tflearn.regression(net,optimizer=my_optimizer)
替换为
时
net = tflearn.regression(net)
取得了完美的结果。
我是机器学习的新手，正在尝试 TFlearn，因为它很简单。
我正在尝试制作一个我觉得很有趣的基本分类器。 我的objective是训练系统预测一个点在哪个方向。
例如，如果我输入两个二维坐标 (50,50) 和 (51,51)，系统必须预测方向是 NE（东北）。
如果我输入 (50,50) 和 (49,49)，系统必须预测方向是 SW（西南）
输入： X1,Y1,X2,Y2,标签
输出： 0 到 8。对于 8 个方向。
所以这是我写的小代码，
from __future__ import print_function
import numpy as np
import tflearn
import tensorflow as tf
import time
from tflearn.data_utils import load_csv
#Sample input 50,50,51,51,5
data, labels = load_csv(filename, target_column=4,
categorical_labels=True, n_classes=8)
my_optimizer = tflearn.SGD(learning_rate=0.1)
net = tflearn.input_data(shape=[None, 4])
net = tflearn.fully_connected(net, 32) #input 4, output 32
net = tflearn.fully_connected(net, 32) #input 32, output 32
net = tflearn.fully_connected(net, 8, activation='softmax')
net = tflearn.regression(net,optimizer=my_optimizer)
model = tflearn.DNN(net)
model.fit(data, labels, n_epoch=100, batch_size=100000, show_metric=True)
model.save("direction-classifier.tfl")
我面临的问题是，即使在我通过了大约 4000 万个输入样本后，系统准确率仍低至 20%。
我将输入限制为 40-x-60 和 40-y-60
我无法理解我是否过度拟合了样本，因为在总共 4000 万个输入的整个训练期间准确率一直不高
为什么这个简单示例的准确性如此低？
编辑： 我降低了学习率并减小了批量大小。但是，结果仍然相同，但准确性很差。 我已经包含了前 25 个步骤的输出。
--
Training Step: 100000 | total loss: 6.33983 | time: 163.327s
| SGD | epoch: 001 | loss: 6.33983 - acc: 0.0663 -- iter: 999999/999999
--
Training Step: 200000 | total loss: 6.84055 | time: 161.981ss
| SGD | epoch: 002 | loss: 6.84055 - acc: 0.1568 -- iter: 999999/999999
--
Training Step: 300000 | total loss: 5.90203 | time: 158.853ss
| SGD | epoch: 003 | loss: 5.90203 - acc: 0.1426 -- iter: 999999/999999
--
Training Step: 400000 | total loss: 5.97782 | time: 157.607ss
| SGD | epoch: 004 | loss: 5.97782 - acc: 0.1465 -- iter: 999999/999999
--
Training Step: 500000 | total loss: 5.97215 | time: 155.929ss
| SGD | epoch: 005 | loss: 5.97215 - acc: 0.1234 -- iter: 999999/999999
--
Training Step: 600000 | total loss: 6.86967 | time: 157.299ss
| SGD | epoch: 006 | loss: 6.86967 - acc: 0.1230 -- iter: 999999/999999
--
Training Step: 700000 | total loss: 6.10330 | time: 158.137ss
| SGD | epoch: 007 | loss: 6.10330 - acc: 0.1242 -- iter: 999999/999999
--
Training Step: 800000 | total loss: 5.81901 | time: 157.464ss
| SGD | epoch: 008 | loss: 5.81901 - acc: 0.1464 -- iter: 999999/999999
--
Training Step: 900000 | total loss: 7.09744 | time: 157.486ss
| SGD | epoch: 009 | loss: 7.09744 - acc: 0.1359 -- iter: 999999/999999
--
Training Step: 1000000 | total loss: 7.19259 | time: 158.369s
| SGD | epoch: 010 | loss: 7.19259 - acc: 0.1248 -- iter: 999999/999999
--
Training Step: 1100000 | total loss: 5.60177 | time: 157.221ss
| SGD | epoch: 011 | loss: 5.60177 - acc: 0.1378 -- iter: 999999/999999
--
Training Step: 1200000 | total loss: 7.16676 | time: 158.607ss
| SGD | epoch: 012 | loss: 7.16676 - acc: 0.1210 -- iter: 999999/999999
--
Training Step: 1300000 | total loss: 6.19163 | time: 163.711ss
| SGD | epoch: 013 | loss: 6.19163 - acc: 0.1635 -- iter: 999999/999999
--
Training Step: 1400000 | total loss: 7.46101 | time: 162.091ss
| SGD | epoch: 014 | loss: 7.46101 - acc: 0.1216 -- iter: 999999/999999
--
Training Step: 1500000 | total loss: 7.78055 | time: 158.468ss
| SGD | epoch: 015 | loss: 7.78055 - acc: 0.1122 -- iter: 999999/999999
--
Training Step: 1600000 | total loss: 6.03101 | time: 158.251ss
| SGD | epoch: 016 | loss: 6.03101 - acc: 0.1103 -- iter: 999999/999999
--
Training Step: 1700000 | total loss: 5.59769 | time: 158.083ss
| SGD | epoch: 017 | loss: 5.59769 - acc: 0.1182 -- iter: 999999/999999
--
Training Step: 1800000 | total loss: 5.45591 | time: 158.088ss
| SGD | epoch: 018 | loss: 5.45591 - acc: 0.0868 -- iter: 999999/999999
--
Training Step: 1900000 | total loss: 6.54951 | time: 157.755ss
| SGD | epoch: 019 | loss: 6.54951 - acc: 0.1353 -- iter: 999999/999999
--
Training Step: 2000000 | total loss: 6.18566 | time: 157.408ss
| SGD | epoch: 020 | loss: 6.18566 - acc: 0.0551 -- iter: 999999/999999
--
Training Step: 2100000 | total loss: 4.95146 | time: 157.572ss
| SGD | epoch: 021 | loss: 4.95146 - acc: 0.1114 -- iter: 999999/999999
--
Training Step: 2200000 | total loss: 5.97208 | time: 157.279ss
| SGD | epoch: 022 | loss: 5.97208 - acc: 0.1277 -- iter: 999999/999999
--
Training Step: 2300000 | total loss: 6.75645 | time: 157.201ss
| SGD | epoch: 023 | loss: 6.75645 - acc: 0.1507 -- iter: 999999/999999
--
Training Step: 2400000 | total loss: 7.04119 | time: 157.346ss
| SGD | epoch: 024 | loss: 7.04119 - acc: 0.1512 -- iter: 999999/999999
--
Training Step: 2500000 | total loss: 5.95451 | time: 157.722ss
| SGD | epoch: 025 | loss: 5.95451 - acc: 0.1421 -- iter: 999999/999999
正如我在上面的评论中所讨论的，这里是使用 a MLP helper class I created 训练多层感知器 classifier 模型的代码。 class 使用 TensorFlow 实现并遵循 scikit-learn 拟合、预测、评分接口。
基本思想是生成一个随机的起点和终点，然后使用字典根据方向创建标签。我使用 np.unique 在生成的数据中查找 class 标签的数量，因为它可能会有所不同（某些方向可能会丢失）。当起点和终点相同时，我还包含了一个空字符串标签。
代码
使用下面的代码，我能够在某些 运行 上实现 100% 的交叉验证准确率。 将 numpy 导入为 np 来自 sklearn.model_selection 导入 ShuffleSplit 从 TFANN 导入 MLPC
#Dictionary to lookup direction ()
DM = {(-1, -1):'SW', (-1, 0):'W', (-1, 1):'NW', (0, 1):'N',
( 1, 1):'NE', ( 1, 0):'E', ( 1, -1):'SE', (0, -1):'S',
( 0, 0):''}
NR = 4096 #Number of rows in sample matrix
A1 = np.random.randint(40, 61, size = (NR, 2)) #Random starting point
A2 = np.random.randint(40, 61, size = (NR, 2)) #Random ending point
A = np.hstack([A1, A2]) #Concat start and end point as feature vector
#Create label from direction vector
Y = np.array([DM[(x, y)] for x, y in (A2 - A1).clip(-1, 1)])
NC = len(np.unique(Y)) #Number of classes
ss = ShuffleSplit(n_splits = 1)
trn, tst = next(ss.split(A)) #Make a train/test split for cross-validation
#%% Create and train Multi-Layer Perceptron for Classification (MLPC)
l = [4, 6, 6, NC] #Neuron counts in each layer
mlpc = MLPC(l, batchSize = 64, maxIter = 128, verbose = True)
mlpc.fit(A[trn], Y[trn])
s1 = mlpc.score(A[trn], Y[trn]) #Training accuracy
s2 = mlpc.score(A[tst], Y[tst]) #Testing accuracy
s3 = mlpc.score(A, Y) #Total accuracy
print('Trn: {:05f}\tTst: {:05f}\tAll: {:05f}'.format(s1, s2, s3))
结果
这是上述代码在我机器上的示例运行：
Iter 1 2.59423236 (Batch Size: 64)
Iter 2 2.25392553 (Batch Size: 64)
Iter 3 2.02569708 (Batch Size: 64)
...
Iter 12 1.53575111 (Batch Size: 64)
Iter 13 1.47963311 (Batch Size: 64)
Iter 14 1.42776408 (Batch Size: 64)
...
Iter 83 0.23911642 (Batch Size: 64)
Iter 84 0.22893350 (Batch Size: 64)
Iter 85 0.23644384 (Batch Size: 64)
...
Iter 94 0.21170238 (Batch Size: 64)
Iter 95 0.20718799 (Batch Size: 64)
Iter 96 0.21230888 (Batch Size: 64)
...
Iter 126 0.17334313 (Batch Size: 64)
Iter 127 0.16970796 (Batch Size: 64)
Iter 128 0.15931854 (Batch Size: 64)
Trn: 0.995659 Tst: 1.000000 All: 0.996094
原来是优化器导致了所有问题。移除自定义优化器后，损失开始正确下降，准确率提高到 99%
必须修改以下两行。
my_optimizer = tflearn.SGD(learning_rate=0.1)
net = tflearn.regression(net,optimizer=my_optimizer)
替换为
时net = tflearn.regression(net)
取得了完美的结果。