keras 不考虑 batch_input 参数答案

【问题标题】：keras does not consider batch_input parameterkeras 不考虑 batch_input 参数
【发布时间】：2016-09-14 14:31:05
【问题描述】：

我正在使用 keras 训练神经网络，但它似乎无法正确解释 batch_size 参数。

看下面的代码（应用很傻，我关心的是输出）。

import numpy as np 
from keras.models import Sequential
from keras.layers import Activation, Dense, Reshape
import keras 

class LossHistory(keras.callbacks.Callback):
    def on_train_begin(self, logs={}):
        self.losses = []

    def on_batch_end(self, batch, logs={}):
        self.losses.append(logs.get('loss'))

history = LossHistory()


X = np.random.normal(0, 1, (1000, 2))
Y = np.random.normal(0, 1, (1000, 3))

model = Sequential()
model.add(Dense(20, input_shape = (2,), name='input layer dude'))
model.add(Activation('relu'))
model.add(Dense(12))
model.add(Activation('relu'))
model.add(Dense(8))
model.add(Activation('linear'))
model.add(Dense(3))
model.add(Activation('linear'))
model.add(Reshape(target_shape=(3,), name='output layer dude'))
model.compile(optimizer='adam', loss='mse', )

当我通过以下方式调用此模型时：

model.fit(X, Y, batch_size=10, nb_epoch=10, callbacks=[history])

输出似乎表明它不是每批做 10 个项目，而是 1000 个（这是总样本数）。

Epoch 1/10
1000/1000 [==============================] - 0s - loss: 898.6197      
Epoch 2/10
1000/1000 [==============================] - 0s - loss: 31.5123     
Epoch 3/10
1000/1000 [==============================] - 0s - loss: 16.7140     
Epoch 4/10
1000/1000 [==============================] - 0s - loss: 11.4034     
Epoch 5/10
1000/1000 [==============================] - 0s - loss: 8.9275     
Epoch 6/10
1000/1000 [==============================] - 0s - loss: 7.4699     
Epoch 7/10
1000/1000 [==============================] - 0s - loss: 6.5648     
Epoch 8/10
1000/1000 [==============================] - 0s - loss: 5.9576     
Epoch 9/10
1000/1000 [==============================] - 0s - loss: 5.5064     
Epoch 10/10
1000/1000 [==============================] - 0s - loss: 5.1514

有什么线索吗？

【问题讨论】：

标签： keras

【解决方案1】：

他实际上正在考虑它。一个时期是对整个数据集的迭代，因此是 1000/1000。

我将批量大小更改为 128 以更具可读性，并添加了一个回调以在每批之后打印损失，我得到的是这样的（我还增加了数据量以提高可读性）：

Using Theano backend.
Using gpu device 1: GeForce GTX 770 (CNMeM is disabled, cuDNN 5105)
Epoch 1/10
 mbloss 1.00058555603 lr 0.0010000000475
  128/10000 [..............................] - ETA: 3s - loss: 1.0006 mbloss 1.00051558018 lr 0.0010000000475
  256/10000 [..............................] - ETA: 4s - loss: 1.0006 mbloss 1.00094401836 lr 0.0010000000475
  384/10000 [>.............................] - ETA: 4s - loss: 1.0007 mbloss 1.00001847744 lr 0.0010000000475
  512/10000 [>.............................] - ETA: 3s - loss: 1.0005 mbloss 1.00019526482 lr 0.0010000000475
  640/10000 [>.............................] - ETA: 3s - loss: 1.0005 mbloss 0.999684214592 lr 0.0010000000475
  768/10000 [=>............................] - ETA: 3s - loss: 1.0003 mbloss 0.999649345875 lr 0.0010000000475
  896/10000 [=>............................] - ETA: 3s - loss: 1.0002 mbloss 1.00126934052 lr 0.0010000000475
 1024/10000 [==>...........................] - ETA: 3s - loss: 1.0004 mbloss 1.00039303303 lr 0.0010000000475
 1152/10000 [==>...........................] - ETA: 3s - loss: 1.0004 mbloss 1.00083625317 lr 0.0010000000475
 1280/10000 [==>...........................] - ETA: 3s - loss: 1.0004 mbloss 1.00036990643 lr 0.0010000000475
 1408/10000 [===>..........................] - ETA: 2s - loss: 1.0004 mbloss 0.999625504017 lr 0.0010000000475
 1536/10000 [===>..........................] - ETA: 2s - loss: 1.0003 mbloss 1.0005017519 lr 0.0010000000475
 1664/10000 [===>..........................] - ETA: 2s - loss: 1.0004 mbloss 0.999049901962 lr 0.0010000000475
 1792/10000 [====>.........................] - ETA: 2s - loss: 1.0003 mbloss 0.999758243561 lr 0.0010000000475
 1920/10000 [====>.........................] - ETA: 2s - loss: 1.0002 mbloss 0.99894207716 lr 0.0010000000475
 2048/10000 [=====>........................] - ETA: 2s - loss: 1.0001 mbloss 1.00113630295 lr 0.0010000000475
 2176/10000 [=====>........................] - ETA: 2s - loss: 1.0002 mbloss 0.999107062817 lr 0.0010000000475

如果你需要它，在批处理结束时打印一些东西的回调：

class MBLossPrint(Callback):
    def on_batch_end(self, batch, logs={}):
        print ' mbloss', logs['loss'], 'lr', self.model.optimizer.lr.get_value()

希望这会有所帮助:)

【讨论】：