keras 簡單 多分類例子

問題：

數軸上，

[    0,100]是第1類

[100,500]是第2類

[500,1000]是第3類

現在，要搭建一個NN，來達到分類的效果。

 

引用

import keras
from keras.models import Sequential
from keras.layers import Dense, Dropout

import numpy as np

數據生成（爲了快捷，這裏直接這麼寫了，因爲python用得不熟···）

def get_rand(low=0, high=1, size=(1,1)):
    if type(size) == type(1):
        return low + (high-low)*np.random.rand(size)
    if len(size) == 2:
        return low + (high-low)*np.random.rand(size[0], size[1])
    if len(size) == 3:
        return low + (high-low)*np.random.rand(size[0], size[1], size[2])

data_len = 1000

data_x = get_rand(low=1, high=1000, size=(data_len,1))
#np.random.randint(3, size=[10, 1])
data_y = np.arange(data_len)*0

for i in range(data_len):
    if data_x[i] < 100 :
        data_y[i] = 0
    elif data_x[i] < 500 :
        data_y[i] = 1
    else :
        data_y[i] = 2

train_x = np.array(data_x)
train_y = keras.utils.to_categorical(data_y, num_classes=3)

 

模型搭建和訓練1：

使用非生成器的fit，直接使用上面的train數據

batch_size=100

model = Sequential()

model.add(Dense(100, input_dim=1, activation='sigmoid'))
model.add(Dropout(0.2))
model.add(Dense(3, activation='sigmoid'))
#model.add(Dense(3, activation='softmax')) #二分問題，最後一層激活函數爲sigmoid，損失函數loss爲binary_crossentropy

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

model.fit(x=train_x, y=train_y, epochs=50, batch_size=batch_size )

 

模型搭建和訓練2：

使用生成器的fit_generator，需要做一個生成函數

def generate_datas(batch_size=100):
    batches = data_len//batch_size #這裏有可能導致溢出，暫時先這麼寫吧，還沒分析清楚 
    while(True): #不停地生成batchs，應該是類似於epochs，每循環一個while，是一個epoch
        for i in range(batches):
            tmp_x = train_x[i*batch_size : (i+1)*batch_size]
            tmp_y = train_y[i*batch_size : (i+1)*batch_size]
            #print('\ni=', i, '\tx=', tmp_x.shape, '\ty=', tmp_y.shape)
            yield (tmp_x, tmp_y)   

batch_size = 100

model = Sequential()

model.add(Dense(100, input_dim=1, activation='sigmoid'))
model.add(Dropout(0.2))
#model.add(Dense(1, activation='sigmoid'))
model.add(Dense(3, activation='softmax'))

model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
model.fit_generator(generator=generate_datas(batch_size=batch_size), epochs=50, steps_per_epoch=data_len//batch_size )

#還要思考一下，steps_per_epoch這個的含義是什麼。從代碼來看，這個就是一個epoch的batchs數量。

 

運行結果：

沒有使用生成器的：

Epoch 40/50
1000/1000 [==============================] - 0s 203us/step - loss: 0.2436 - acc: 0.9270
Epoch 41/50
1000/1000 [==============================] - 0s 78us/step - loss: 0.2358 - acc: 0.9220
Epoch 42/50
1000/1000 [==============================] - 0s 47us/step - loss: 0.2239 - acc: 0.9370
Epoch 43/50
1000/1000 [==============================] - 0s 94us/step - loss: 0.2213 - acc: 0.9380
Epoch 44/50
1000/1000 [==============================] - 0s 109us/step - loss: 0.2158 - acc: 0.9320
Epoch 45/50
1000/1000 [==============================] - 0s 203us/step - loss: 0.2113 - acc: 0.9410
Epoch 46/50
1000/1000 [==============================] - 0s 78us/step - loss: 0.2117 - acc: 0.9390
Epoch 47/50
1000/1000 [==============================] - 0s 78us/step - loss: 0.2226 - acc: 0.9160
Epoch 48/50
1000/1000 [==============================] - 0s 47us/step - loss: 0.2191 - acc: 0.9170
Epoch 49/50
1000/1000 [==============================] - 0s 125us/step - loss: 0.1989 - acc: 0.9450
Epoch 50/50
1000/1000 [==============================] - 0s 63us/step - loss: 0.2013 - acc: 0.9370

 

使用生成器的：

Epoch 40/50
10/10 [==============================] - 0s 42ms/step - loss: 0.2319 - acc: 0.9350
Epoch 41/50
10/10 [==============================] - 0s 6ms/step - loss: 0.2224 - acc: 0.9410
Epoch 42/50
10/10 [==============================] - 0s 5ms/step - loss: 0.2272 - acc: 0.9380
Epoch 43/50
10/10 [==============================] - 0s 5ms/step - loss: 0.2243 - acc: 0.9380
Epoch 44/50
10/10 [==============================] - 0s 14ms/step - loss: 0.2259 - acc: 0.9350
Epoch 45/50
10/10 [==============================] - 0s 14ms/step - loss: 0.2237 - acc: 0.9260
Epoch 46/50
10/10 [==============================] - 0s 5ms/step - loss: 0.2220 - acc: 0.9320
Epoch 47/50
10/10 [==============================] - 0s 5ms/step - loss: 0.2134 - acc: 0.9450
Epoch 48/50
10/10 [==============================] - 0s 9ms/step - loss: 0.2220 - acc: 0.9280
Epoch 49/50
10/10 [==============================] - 0s 5ms/step - loss: 0.2199 - acc: 0.9300
Epoch 50/50
10/10 [==============================] - 0s 3ms/step - loss: 0.1961 - acc: 0.9420

可見兩者收斂時間是一致的。 使用生成器來fit，是爲了避免內存不足等問題。