用于转换和手动加载图像的Keras input_shape [英] Keras input_shape for conv2d and manually loaded images

问题描述

我正在从384x286黑白图像中手动创建数据集.

I am manually creating my dataset from a number of 384x286 b/w images.

我加载这样的图像:

x = []
for f in files:
        img = Image.open(f)
        img.load()
        data = np.asarray(img, dtype="int32")
        x.append(data)
x = np.array(x)

这导致x为数组(num_samples，286、384)

this results in x being an array (num_samples, 286, 384)

print(x.shape) => (100, 286, 384)

阅读keras文档，并检查我的后端，我应该向卷积步骤提供由(rows，cols，channels)组成的input_shape

reading the keras documentation, and checking my backend, i should provide to the convolution step an input_shape composed by ( rows, cols, channels )

由于我不随意知道样本大小，因此我希望将其作为输入大小，类似

since i don't arbitrarily know the sample size, i would have expected to pass as an input size, something similar to

( None, 286, 384, 1 )

该模型的构建如下:

model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=input_shape))
# other steps...

作为input_shape(286、384、1)传递会导致:

passing as input_shape (286, 384, 1) causes:

检查输入时出错:预期conv2d_1_input具有4维，但数组的形状为(85，286，384)

Error when checking input: expected conv2d_1_input to have 4 dimensions, but got array with shape (85, 286, 384)

传递as_input_shape(无，286、384、1)会导致:

passing as_input_shape (None, 286, 384, 1 ) causes:

输入0与conv2d_1层不兼容:预期ndim = 4，找到ndim = 5

Input 0 is incompatible with layer conv2d_1: expected ndim=4, found ndim=5

我在做什么错?我该如何重塑输入数组的形状?

what am i doing wrong ? how do i have to reshape the input array?

推荐答案

将input_shape设置为(286,384,1).现在，模型需要一个4维的输入.这意味着您必须使用.reshape(n_images, 286, 384, 1)重塑图像.现在，您已经添加了一个额外的维度，而无需更改数据，并且模型可以运行了.基本上，您需要将数据重塑为(n_images，x_shape，y_shape，channels).

Set the input_shape to (286,384,1). Now the model expects an input with 4 dimensions. This means that you have to reshape your image with .reshape(n_images, 286, 384, 1). Now you have added an extra dimension without changing the data and your model is ready to run. Basically, you need to reshape your data to (n_images, x_shape, y_shape, channels).

很棒的事情是您还可以使用RGB图像作为输入.只需将channels更改为3.

The cool thing is that you also can use an RGB-image as input. Just change channels to 3.

还要检查以下答案: Keras输入说明:input_shape，units，batch_size，dim，等等

Check also this answer: Keras input explanation: input_shape, units, batch_size, dim, etc

示例

import numpy as np
from keras.models import Sequential
from keras.layers.convolutional import Convolution2D
from keras.layers.core import Flatten, Dense, Activation
from keras.utils import np_utils

#Create model
model = Sequential()
model.add(Convolution2D(32, kernel_size=(3, 3), activation='relu', input_shape=(286,384,1)))
model.add(Flatten())
model.add(Dense(2))
model.add(Activation('softmax'))

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

#Create random data
n_images=100
data = np.random.randint(0,2,n_images*286*384)
labels = np.random.randint(0,2,n_images)
labels = np_utils.to_categorical(list(labels))

#add dimension to images
data = data.reshape(n_images,286,384,1)

#Fit model
model.fit(data, labels, verbose=1)

这篇关于用于转换和手动加载图像的Keras input_shape的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！