top-5精度是否可能低于top-1精度？
在我的训练结果中：
我的代码：
import keras
from keras import datasets
from keras import layers
from keras import models
print(f'keras版本={keras.__version__}')
#載入 cifar10 資料庫的訓練資料，並自動分為『訓練組』及『測試組』
(X_train, y_train),(X_test, y_test)= datasets.cifar10.load_data()
#建立FNN model
model = models.Sequential()
#建立input層
#此例若沒加Flatten Layer則會出錯（dense_數字，此數字後面會變動，其餘都一樣）：
#ValueError: Error when checking target: expected dense_30 to have 3 dimensions, but got array with shape (60000, 1)
model.add(layers.Flatten())
#建立隱藏層（此例中此隱藏層只有1個神經元）
#可以寫得較完整：model.add(layers.Dense(input_dim=28*28,units=2))
#units的舊式寫法為output_dim：model.add(layers.Dense(input_dim=28*28,output_dim=2))
model.add(layers.Dense(units=1))
#建立輸出層（輸出神經元至少>=10）
model.add(layers.Dense(units=10))
import functools
#若將k=5改成k=3就會變成計算Top-3
top5_acc = functools.partial(keras.metrics.top_k_categorical_accuracy, k=5)
#此名稱可自訂，當training時顯示用的
top5_acc.__name__ = 'top5_acc'
# 最佳化設定
#若loss設定為categorical_crossentropy則會出錯：
#ValueError: Error when checking target: expected dense_2 to have shape (10,) but got array with shape (1,)
model.compile(loss='sparse_categorical_crossentropy', optimizer='sgd', metrics=['accuracy',top5_acc])
# 開始進行training
model.fit(
x=X_train,
y=y_train,
epochs=1,
verbose=1,
validation_split=0.1
)
#若沒有呼叫
#input_shape = [28,28,1]
#build(input_shape)
#的話，唯有在fit()之後才能呼叫summary()
#否則會出錯：
#This model has not yet been built. Build the model first by calling build() or calling fit() with some data.
#Or specify input_shape or batch_input_shape in the first layer for automatic build.
#model.build(input_shape)
model.summary()
# 開始進行testing
test_loss,top_1_accuracy, top_5_accuracy = model.evaluate(X_test, y_test)
print(f'test_loss={test_loss}')
print(f'top_1_accuracy={top_1_accuracy}')
print(f'top_5_accuracy={top_5_accuracy}')
是否存在设置错误或概念错误？
谢谢〜
====================================
将精度更改为categorical_accuracy后：
import keras
from keras import datasets
from keras import layers
from keras import models
print(f'keras版本={keras.__version__}')
#載入 cifar10 資料庫的訓練資料，並自動分為『訓練組』及『測試組』
(X_train, y_train),(X_test, y_test)= datasets.cifar10.load_data()
#建立FNN model
model = models.Sequential()
#建立input層
#此例若沒加Flatten Layer則會出錯（dense_數字，此數字後面會變動，其餘都一樣）：
#ValueError: Error when checking target: expected dense_30 to have 3 dimensions, but got array with shape (60000, 1)
model.add(layers.Flatten())
#建立隱藏層（此例中此隱藏層只有1個神經元）
#可以寫得較完整：model.add(layers.Dense(input_dim=28*28,units=2))
#units的舊式寫法為output_dim：model.add(layers.Dense(input_dim=28*28,output_dim=2))
model.add(layers.Dense(units=1))
#建立輸出層（輸出神經元至少>=10）
model.add(layers.Dense(units=10))
import functools
#若將k=5改成k=3就會變成計算Top-3
top5_acc = functools.partial(keras.metrics.top_k_categorical_accuracy, k=5)
#此名稱可自訂，當training時顯示用的
top5_acc.__name__ = 'top5_acc'
# 最佳化設定
#若loss設定為categorical_crossentropy則會出錯：
#ValueError: Error when checking target: expected dense_2 to have shape (10,) but got array with shape (1,)
model.compile(loss='sparse_categorical_crossentropy', optimizer='sgd', metrics=['categorical_accuracy',top5_acc])
# 開始進行training
model.fit(
x=X_train,
y=y_train,
epochs=1,
verbose=1,
validation_split=0.1
)
#若沒有呼叫
#input_shape = [28,28,1]
#build(input_shape)
#的話，唯有在fit()之後才能呼叫summary()
#否則會出錯：
#This model has not yet been built. Build the model first by calling build() or calling fit() with some data.
#Or specify input_shape or batch_input_shape in the first layer for automatic build.
#model.build(input_shape)
model.summary()
# 開始進行testing
test_loss,top_1_accuracy, top_5_accuracy = model.evaluate(X_test, y_test)
print(f'test_loss={test_loss}')
print(f'top_1_accuracy={top_1_accuracy}')
print(f'top_5_accuracy={top_5_accuracy}')
我知道了
===我添加softmax ===
import keras
from keras import datasets
from keras import layers
from keras import models
print(f'keras版本={keras.__version__}')
#載入 cifar10 資料庫的訓練資料，並自動分為『訓練組』及『測試組』
(X_train, y_train),(X_test, y_test)= datasets.cifar10.load_data()
#建立FNN model
model = models.Sequential()
#建立input層
#此例若沒加Flatten Layer則會出錯（dense_數字，此數字後面會變動，其餘都一樣）：
#ValueError: Error when checking target: expected dense_30 to have 3 dimensions, but got array with shape (60000, 1)
model.add(layers.Flatten())
#建立隱藏層（此例中此隱藏層只有1個神經元）
#可以寫得較完整：model.add(layers.Dense(input_dim=28*28,units=2))
#units的舊式寫法為output_dim：model.add(layers.Dense(input_dim=28*28,output_dim=2))
model.add(layers.Dense(units=1))
#建立輸出層（輸出神經元至少>=10）
model.add(layers.Dense(units=10,activation='softmax'))
import functools
#若將k=5改成k=3就會變成計算Top-3
top5_acc = functools.partial(keras.metrics.top_k_categorical_accuracy, k=5)
#此名稱可自訂，當training時顯示用的
top5_acc.__name__ = 'top5_acc'
# 最佳化設定
#若loss設定為categorical_crossentropy則會出錯：
#ValueError: Error when checking target: expected dense_2 to have shape (10,) but got array with shape (1,)
model.compile(loss='sparse_categorical_crossentropy', optimizer='sgd', metrics=['categorical_accuracy',top5_acc])
# 開始進行training
model.fit(
x=X_train,
y=y_train,
epochs=1,
verbose=1,
validation_split=0.1
)
#若沒有呼叫
#input_shape = [28,28,1]
#build(input_shape)
#的話，唯有在fit()之後才能呼叫summary()
#否則會出錯：
#This model has not yet been built. Build the model first by calling build() or calling fit() with some data.
#Or specify input_shape or batch_input_shape in the first layer for automatic build.
#model.build(input_shape)
model.summary()
# 開始進行testing
test_loss,top_1_accuracy, top_5_accuracy = model.evaluate(X_test, y_test)
print(f'test_loss={test_loss}')
print(f'top_1_accuracy={top_1_accuracy}')
print(f'top_5_accuracy={top_5_accuracy}')
我得到了：
================================================ ==
感谢@Thomas Schillaci的建议， 我发现一个没有任何错误：
import keras
from keras import datasets
from keras import layers
from keras import models
print(f'keras版本={keras.__version__}')
#載入 cifar10 資料庫的訓練資料，並自動分為『訓練組』及『測試組』
(X_train, y_train),(X_test, y_test)= datasets.cifar10.load_data()
#建立FNN model
model = models.Sequential()
#建立input層
#此例若沒加Flatten Layer則會出錯（dense_數字，此數字後面會變動，其餘都一樣）：
#ValueError: Error when checking target: expected dense_30 to have 3 dimensions, but got array with shape (60000, 1)
model.add(layers.Flatten())
#建立隱藏層（此例中此隱藏層只有1個神經元）
#可以寫得較完整：model.add(layers.Dense(input_dim=28*28,units=2))
#units的舊式寫法為output_dim：model.add(layers.Dense(input_dim=28*28,output_dim=2))
model.add(layers.Dense(units=1))
#建立輸出層（輸出神經元至少>=10）
model.add(layers.Dense(units=10,activation='softmax'))
import functools
#底下寫法會怪怪的，如：top1=99.64、top-5=0.0
#若將k=5改成k=3就會變成計算Top-3
#top5_acc = functools.partial(keras.metrics.top_k_categorical_accuracy, k=5)
#此名稱可自訂，當training時顯示用的
#top5_acc.__name__ = 'top5_acc'
# 最佳化設定
#若loss設定為categorical_crossentropy則會出錯：
#ValueError: Error when checking target: expected dense_2 to have shape (10,) but got array with shape (1,)
#model.compile(loss='sparse_categorical_crossentropy', optimizer='sgd', metrics=['categorical_accuracy',top5_acc])
#底下寫法會出錯：ValueError: '<lambda>' is not a valid scope name
#from keras import metrics
#top3 = lambda x,y: metrics.top_k_categorical_accuracy(x,y,3)
#top5 = lambda x,y: metrics.top_k_categorical_accuracy(x,y,5)
#model.compile(loss='sparse_categorical_crossentropy', optimizer='sgd', metrics=['accuracy',top3,top5])
#底下寫法會出錯：ValueError: None values not supported.
#def top3(x,y):
# metrics.top_k_categorical_accuracy(x,y,3)
#def top5(x,y):
# metrics.top_k_categorical_accuracy(x,y,5)
#model.compile(loss='sparse_categorical_crossentropy', optimizer='sgd', metrics=['accuracy',top3,top5])
#底下寫法training時正確：accuracy: 0.1371 - top_k_categorical_accuracy: 0.4191 - acc_top3: 0.2147
#但testing時出錯：
# top_1_accuracy=0.15530000627040863
# top_3_accuracy=0.0
# top_5_accuracy=0.0
#而且會顯示：top_k_categorical_accuracy而非顯示top-5
#45000/45000 [==============================] - 2s 52us/step - loss: 7.9220
#- accuracy: 0.1136 - top_k_categorical_accuracy: 0.4338 - acc_top3: 0.3794 - val_loss: 2.2809
# - val_accuracy: 0.1470 - val_top_k_categorical_accuracy: 0.2550 - val_acc_top3: 0.2412
# def acc_top3(y_true, y_pred):
# return metrics.top_k_categorical_accuracy(y_true, y_pred, k=3)
# model.compile(loss='sparse_categorical_crossentropy',
# optimizer='adam',
# metrics=['accuracy', metrics.top_k_categorical_accuracy, acc_top3])
#底下寫法training時正確：accuracy: 0.1332 - top5_acc: 0.5029 - acc_top3: 0.3971
#但testing時出錯：
# top_1_accuracy=0.16220000386238098
# top_3_accuracy=0.08720000088214874
# top_5_accuracy=0.3935000002384186
#top5_acc = functools.partial(keras.metrics.top_k_categorical_accuracy, k=5)
#此名稱可自訂，當training時顯示用的
# top5_acc.__name__ = 'top5_acc'
# def acc_top3(y_true, y_pred):
# return metrics.top_k_categorical_accuracy(y_true, y_pred, k=3)
# model.compile(loss='sparse_categorical_crossentropy',
# optimizer='adam',
# metrics=['accuracy', top5_acc, acc_top3])
#底下寫法training時正確：acc: 0.1314 - acc_top2: 0.4392 - acc_top3: 0.4797 - acc_top4: 0.5172 - acc_top5: 0.5492
#但testing時出錯：
# top_1_accuracy=0.13410000503063202
# top_2_accuracy=0.03440000116825104
# top_3_accuracy=0.034699998795986176
# top_4_accuracy=0.03669999912381172
# top_5_accuracy=0.042399998754262924
# from keras.metrics import top_k_categorical_accuracy
# def acc_top2(y_true, y_pred):
# return top_k_categorical_accuracy(y_true, y_pred, k=2)
# def acc_top3(y_true, y_pred):
# return top_k_categorical_accuracy(y_true, y_pred, k=3)
# def acc_top4(y_true, y_pred):
# return top_k_categorical_accuracy(y_true, y_pred, k=4)
# def acc_top5(y_true, y_pred):
# return top_k_categorical_accuracy(y_true, y_pred, k=5)
# model.compile(loss='sparse_categorical_crossentropy',
# optimizer='adam',
# metrics=['acc', acc_top2, acc_top3, acc_top4, acc_top5])
#底下寫法training時正確：acc: 0.1314 - acc_top2: 0.4392 - acc_top3: 0.4797 - acc_top4: 0.5172 - acc_top5: 0.5492
#但testing時出錯：
# top_1_accuracy=0.15760000050067902
# top_2_accuracy=0.0
# top_3_accuracy=0.125900000333786
# top_4_accuracy=0.48910000920295715
# top_5_accuracy=0.6032000184059143
# from keras.metrics import top_k_categorical_accuracy
# def acc_top2(y_true, y_pred):
# return top_k_categorical_accuracy(y_true, y_pred, k=2)
# def acc_top3(y_true, y_pred):
# return top_k_categorical_accuracy(y_true, y_pred, k=3)
# def acc_top4(y_true, y_pred):
# return top_k_categorical_accuracy(y_true, y_pred, k=4)
# def acc_top5(y_true, y_pred):
# return top_k_categorical_accuracy(y_true, y_pred, k=5)
# model.compile(loss='sparse_categorical_crossentropy',
# optimizer='adam',
# metrics=['accuracy', acc_top2, acc_top3, acc_top4, acc_top5])
from keras.metrics import top_k_categorical_accuracy
def acc_top2(y_true, y_pred):
return top_k_categorical_accuracy(y_true, y_pred, k=2)
def acc_top3(y_true, y_pred):
return top_k_categorical_accuracy(y_true, y_pred, k=3)
def acc_top4(y_true, y_pred):
return top_k_categorical_accuracy(y_true, y_pred, k=4)
def acc_top5(y_true, y_pred):
return top_k_categorical_accuracy(y_true, y_pred, k=5)
model.compile(loss='sparse_categorical_crossentropy',
optimizer='adam',
metrics=['categorical_accuracy', acc_top2, acc_top3, acc_top4, acc_top5])
#底下看似正確，但卻有一次例外，其他次全對，但無法reproduce該次錯誤
# model.compile(loss='sparse_categorical_crossentropy',
# optimizer='adam',
# metrics=['accuracy', metrics.top_k_categorical_accuracy])
# 開始進行training
model.fit(
x=X_train,
y=y_train,
epochs=1,
verbose=1,
validation_split=0.1
)
#若沒有呼叫
#input_shape = [28,28,1]
#build(input_shape)
#的話，唯有在fit()之後才能呼叫summary()
#否則會出錯：
#This model has not yet been built. Build the model first by calling build() or calling fit() with some data.
#Or specify input_shape or batch_input_shape in the first layer for automatic build.
#model.build(input_shape)
model.summary()
# 開始進行testing
#test_loss,top_1_accuracy, top_5_accuracy, top_3_accuracy = model.evaluate(X_test, y_test)
#test_loss,top_1_accuracy, top_5_accuracy = model.evaluate(X_test, y_test)
test_loss,top_1_accuracy, top_2_accuracy, top_3_accuracy, top_4_accuracy, top_5_accuracy = model.evaluate(X_test, y_test)
print(model.metrics_names)
print(f'test_loss={test_loss}')
print(f'top_1_accuracy={top_1_accuracy}')
print(f'top_2_accuracy={top_2_accuracy}')
print(f'top_3_accuracy={top_3_accuracy}')
print(f'top_4_accuracy={top_4_accuracy}')
print(f'top_5_accuracy={top_5_accuracy}')
结果很好：
1 个答案:
答案 0 :(得分：1)
您使用的是前1位的准确性指标和前5位的分类准确性指标。将第一个切换为'categorical_accuracy'应该可以解决问题
相关问题
最新问题
- 我写了这段代码，但我无法理解我的错误
- 我无法从一个代码实例的列表中删除 None 值，但我可以在另一个实例中。为什么它适用于一个细分市场而不适用于另一个细分市场？
- 是否有可能使 loadstring 不可能等于打印？卢阿
- java中的random.expovariate()
- Appscript 通过会议在 Google 日历中发送电子邮件和创建活动
- 为什么我的 Onclick 箭头功能在 React 中不起作用？
- 在此代码中是否有使用“this”的替代方法？
- 在 SQL Server 和 PostgreSQL 上查询，我如何从第一个表获得第二个表的可视化
- 每千个数字得到
- 更新了城市边界 KML 文件的来源？