用于训练考虑 keras 中最后一层的网络
Posted
技术标签:
【中文标题】用于训练考虑 keras 中最后一层的网络【英文标题】：for training considering network from last layer in keras 【发布时间】：2019-05-21 03:48:01 【问题描述】：这是我的型号代码：
Model=Sequential()
input_img = Input(shape=(180,180,3)) # adapt this if using channels_first` image data format
x = Conv2D(64, (3, 3), padding='valid')(input_img)
x = Conv2D(64, (3, 3), padding='valid',strides=2)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
y = Conv2D(64, (3, 3), padding='valid')(x)
model=Model(input_img,y)
生成器部分如下
train_datagen = ImageDataGenerator(rescale=1./255)
test_datagen = ImageDataGenerator(rescale=1./255)
train_generator=train_datagen.flow_from_directory(
'\Dipti\medical_image_comp',
target_size=(180,180),
batch_size=128,
class_mode=None)
validation_generator = test_datagen.flow_from_directory(
'D:\Dipti\medical_image_comp\scale0',
target_size=(180,180),
batch_size=128,
class_mode=None)
通过以下方式拟合这个简单的网络：
history=model.fit_generator(
train_generator,
epochs=100,
steps_per_epoch=training_samples/batch_size,
validation_data=validation_generator,
validation_steps=testing_samples/batch_size)
The following error occurs:
纪元 1/100
ValueError Traceback (most recent
call last)
<ipython-input-41-bf2c0dd3bbcf> in <module>()
4 epochs=100,
5 validation_data=validation_generator,
- ---> 6 个验证步骤=testing_samples/batch_size)
~\Anaconda3\lib\site-packages\keras\legacy\interfaces.py in
wrapper(*args, **kwargs)
89 warnings.warn('Update your `' + object_name +
90 '` call to the Keras 2 API: ' + signature, stacklevel=2)
- --> 91 返回函数(*args, **kwargs) 92 包装器._original_function = func 93 返回包装器
~\Anaconda3\lib\site-packages\keras\models.py in fit_generator(self, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch) 第1254章 第1255章 -> 1256 初始时期=初始时期） 1257 第1258章
~\Anaconda3\lib\site-packages\keras\legacy\interfaces.py 在 wrapper(*args, **kwargs)
89 warnings.warn('更新您对 Keras 2 API 的' + object_name +
90 ' 调用：' + 签名，stacklevel=2)
---> 91 返回函数（*args，**kwargs）
92 包装器._original_function = func
93 返回包装器
~\Anaconda3\lib\site-packages\keras\engine\training.py in
fit_generator(self, generator, steps_per_epoch, epochs, verbose,
callbacks, validation_data, validation_steps, class_weight,
max_queue_size, workers, use_multiprocessing, shuffle,
initial_epoch)
2160 'a tuple `(x,
y, sample_weight)` '
2161 'or `(x,
y)`. Found: ' +
-> 2162
str(generator_output))
2163 # build batch logs
2164 batch_logs =
ValueError: Output of generator should be a tuple `(x, y,
sample_weight)` or `(x, y)`. Found: [[[[1.
0.91372555 1. ]
[0.8980393 0.78823537 0.87843144]
[0.8705883 0.7607844 0.85098046]
...
[0.8313726 0.7411765 0.8117648 ]
[0.85098046 0.7607844 0.8313726 ]
[0.83921576 0.7490196 0.8196079 ]]
[[0.9333334 0.8352942 0.9215687 ]
[0.8980393 0.8000001 0.8862746 ]
[0.9294118 0.8313726 0.9176471 ]
...
[0.7803922 0.6901961 0.7607844 ]
[0.8196079 0.7294118 0.8000001 ]
[0.8588236 0.7686275 0.83921576]]
[[0.9176471 0.8235295 0.909804 ]
[0.854902 0.7607844 0.8470589 ]
[0.8745099 0.7803922 0.86666673]
...
[0.7686275 0.6784314 0.7490196 ]
[0.79215693 0.7019608 0.7725491 ]
[0.83921576 0.7490196 0.8196079 ]]
...
[[0.81568635 0.6784314 0.7725491 ]
[0.80392164 0.6666667 0.7607844 ]
[0.8196079 0.68235296 0.77647066]
...
[0.8470589 0.6784314 0.78823537]
[0.8352942 0.6666667 0.77647066]
[0.8745099 0.7058824 0.81568635]]
[[0.7686275 0.6313726 0.7254902 ]
[0.7607844 0.62352943 0.7176471 ]
[0.79215693 0.654902 0.7490196 ]
...
[0.8431373 0.6745098 0.7843138 ]
[0.83921576 0.67058825 0.7803922 ]
[0.882353 0.7137255 0.8235295 ]]
[[0.8235295 0.6862745 0.7725491 ]
[0.7725491 0.63529414 0.72156864]
[0.78823537 0.6509804 0.74509805]
...
[0.8588236 0.6901961 0.8000001 ]
[0.86666673 0.69803923 0.8078432 ]
[0.8862746 0.7176471 0.82745105]]]
[[[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
...
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]]
[[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
...
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]]
[[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
...
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]]
...
[[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
...
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]]
[[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
...
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]]
[[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
...
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]
[0.8705883 0.8705883 0.8705883 ]]]
[[[0.92549026 0.82745105 0.90196085]
[0.89019614 0.7843138 0.8588236 ]
[0.9176471 0.8078432 0.8941177 ]
...
[0.7960785 0.47450984 0.6627451 ]
[0.76470596 0.43529415 0.627451 ]
[0.77647066 0.44705886 0.6392157 ]]
[[0.9058824 0.8000001 0.8745099 ]
[0.8941177 0.7803922 0.8588236 ]
[0.86666673 0.7411765 0.8313726 ]
...
[0.80392164 0.48235297 0.67058825]
[0.79215693 0.47058827 0.65882355]
[0.8588236 0.5294118 0.72156864]]
[[0.83921576 0.7254902 0.80392164]
[0.87843144 0.75294125 0.8352942 ]
[0.8235295 0.6901961 0.7843138 ]
...
[0.8078432 0.48627454 0.6745098 ]
[0.80392164 0.48235297 0.67058825]
[0.8862746 0.5647059 0.75294125]]
...
我无法获得这样一个简单的网络。我已经建立了很多具有相同概念的模型，但是这里这个网络无法训练。请建议我用使用 Adam 优化器和 MSE 作为损失函数从 dsidirectory 概念流出。 我希望你明白我的意思
先生，通过这个小网络，我只是想减小图像的大小，在训练这个网络之后，我必须将该网络的输出应用到一个图像编解码器，进一步必须做相反的过程来生成重建的图像。然后出于测试目的，我必须比较原始图像和比较图像。因为这基本上是一个压缩任务，减少图像的大小，所以特别是我的工作不需要像分类和回归那样的标签。我想复制题为“使用卷积神经网络的端到端压缩框架”的论文的结果，这个小网络基本上是我想用它们的参数训练的第一个模块。您也可以检查纸张 我希望你现在了解整个问题
【问题讨论】：
【参考方案1】：可能在您的生成器中，您将标签作为元组的第一个元素返回，输入图像作为第二个元素返回。交换这两个，问题就解决了。
【讨论】：
此生成器仅取自 keras 教程 我没有得到要更改和交换的内容 @DiptiMishra 您能否编辑您的问题并包含生成器的代码？然后我可以告诉你可能出了什么问题。 为生成器部分编辑 @DiptiMishra 我明白了。那是什么fixed_generator？也发布代码。以上是关于用于训练考虑 keras 中最后一层的网络的主要内容，如果未能解决你的问题，请参考以下文章