python - 如何修复 : ValueError: Error when checking input: expected conv2d_130_input to have shape (1, 512, 512) 但得到形状为 (79, 512, 512) 的数组

Question

我是使用 CNN 的初学者。

所以，我正在构建一个预测脑肿瘤类型的 2D 卷积神经网络，并对 NumPy 数组有一个疑问。我的模型的输入形状是 (1, 512, 512) as (channels, img_height, img_width)。第四个维度是 num_images，它似乎是由 TensorFlow 自动定义的。这只是一个快速背景。我有 3064 个“.mat”扩展文件，其中包含脑肿瘤的 MRI 扫描。一切都已设置完毕。我将“.mat”文件转换为 numpy 矩阵，并将整个矩阵列表附加到单个 numpy 数组中，以作为 CNN 的输入传递。我还有相应的标签(将输入传递到模型时索引链接到图像)作为 numpy 数组。图像和标签中的所有数字都是浮点类型。

同样，我的输入形状是 (1, 512, 512)。但是，在拟合模型时出现以下错误:

ValueError:检查输入时出错:预期 conv2d_130_input 的形状为 (1, 512, 512)，但得到的数组形状为 (79, 512, 512)

因此，我对 NumPy 数组进行切片以创建 train_images、train_labels、test_images、test_labels。我已经验证了每个训练集和测试集的长度与标签匹配。它们也是数组，我检查了多次。这是一个值(value)错误。那么，我该如何解决这个问题呢？

我什至不知道输入形状变成了哪里(79,512,512)。我有一个循环将 f"{n}.mat"图像转换为矩阵。我使用 100 张图像进行测试，并进行 80 个训练和 20 个测试。我认为错误就在这里，输入形状是( channel ，img-hght，img-wdth)，但是剩下要训练的图像数量被放置在 channel 的值中。因此，输入被放置为 (num_images, img-hght, img-wdth)。这是错误的，应该改变，但我不知道该怎么做。或者，我可能是错的，我所说的可能没有意义。我提供所有代码，并在 Colab 上运行。如果您下载代码并想要运行它，请确保更改图像路径以帮助我。非常感谢!

数据集:https://figshare.com/articles/brain_tumor_dataset/1512427/5

#Importing the necessary libraries through PIP to the Virtual Environment
try:
  !python -m pip install --upgrade pip #Quickly update PIP to latest version
  !python -m pip install pymatreader
  !python -m pip install pyswarm #An interesting library for testing purposes
  print("""
The following libraries are available and have been successfully fetched:
  >>> PyMatReader
  >>> Particle Swarm""")
except Exception:
  print("""
The following libraries have unavailable and have not been fetched:
  >>> PyMatReader
  >>> Particle Swarm""")
  pass

#Importing the necessary libraries to the Virtual Environment
from __future__ import absolute_import, division, print_function, unicode_literals
import random as rnd
from random import shuffle
import numpy as np
import sys
import scipy as sp
from scipy.ndimage import gaussian_filter
import pymatreader as pym
import pandas as pd
import seaborn as sns
import matplotlib as mpl
import matplotlib.image as mplimg
import matplotlib.pyplot as plt
import PIL
from PIL import Image
import imageio
import sklearn as sk
from sklearn.model_selection import train_test_split
from sklearn.feature_extraction import image
import sklearn.metrics as skm

print("""
The following libraries have been successfully imported:
  >>> Future
  >>> Random (with shuffle)
  >>> NumPy
  >>> System
  >>> SciPy (with gaussian filter)
  >>> PyMatReader
  >>> Pandas
  >>> Seaborn
  >>> Matplotlib (with PyPlot & Image)
  >>> PIL (with Image)
  >>> Imageio
  >>> Sci-Kit Learn (with metrics & train_test_split)
  >>> Sci-kit Learn Feature Extraction (with Image)
""")

try:
  %tensorflow_version 2.x
  import keras
  import tensorflow as tf
  print("TensorFlow version 2.x is available and has been successfully imported.")
except Exception:
  %tensorflow_version 1.x
  import keras
  import tensorflow as tf
  print("TensorFlow version 2.x is unavailable. TensorFlow version 1.x has been imported instead.")
  pass

from tensorflow.keras import datasets, layers, models
import keras.preprocessing
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Conv2D, Flatten, Dropout, MaxPooling2D
from keras.optimizers import Adam
import pyswarm
from pyswarm import pso

autoTune = tf.data.experimental.AUTOTUNE

print("""
The following deep learning optimizers have been successfully imported:
  >>> Adam
  >>> Particle Swarm (with pso)
""")

print("All libraries have been successfully imported.")

#Understanding the Image Data using Seaborn and Matplotlib
classNames = {1 : "Meningioma", 2 : "Glioma", 3 : "Pituitary Tumor", 4 : "Unkown", 5 : "Unkown"}
outputSize = len(classNames)

chooseImgNum = 2978
example = pym.read_mat(f'/content/gdrive/My Drive/My Files/Neuroimaging/Neuroimaging Datasets/MATLAB Files/{chooseImgNum}.mat')
cjdata = example['cjdata']
pid = cjdata['PID']
img = cjdata['image']
label = cjdata['label']

tumorBorder = cjdata['tumorBorder']
tumorMask = cjdata['tumorMask']
print("Tumor Border is: \n", tumorBorder, "\n")
print("Tumor Mask is: \n", tumorMask, "\n")

def printImage():
  plt.figure(figsize=(5, 5))
  plt.imshow(img, cmap=None)

def matrixConv(): #Data Visualization only
  matrix = np.asmatrix(tumorBorder)
  plt.figure(figsize=(5, 5))
  return matrix

def applyGrayscale():
  plt.figure(figsize=(5, 5))
  plt.imshow(img, cmap='gray')

print("""
      Below is the original image followed by a grayscale application:
____________________________________________________________________________
""")

printImage()
applyGrayscale()

#Preprocessing Brain Images from Dataset
range1 = np.arange(0, 100)
imgMatrices = []
imgNum = 1
i = 1

while imgNum in range1:
  imgNum = pym.read_mat(f'/content/gdrive/My Drive/My Files/Neuroimaging/Neuroimaging Datasets/MATLAB Files/{imgNum}.mat')
  cjdata = imgNum['cjdata']
  imgMatrix = cjdata['image']
  # plt.figure(figsize=(5, 5))
  # plt.imshow(image_matrix, cmap='gray')
  imgMatrixNP = np.asmatrix(imgMatrix)
  imgArrayNP = np.asarray(imgMatrixNP)
  imgMatrices.append(imgArrayNP)
  imgNum = i
  i = i + 1

print("The length of the image input list is:", len(imgMatrices))

imgMatricesNP = np.asarray(imgMatrices)
print("The length of the converted image input array is:", len(imgMatricesNP), "\n")

print("The image input array:")
imgMatricesNP #Prints the raw array

#Supervised Learning: Understanding Cancer Type labels
np.set_printoptions(threshold=3)
#np.set_printoptions(threshold=sys.maxsize) #To check the content of the entire array

rawMatData = pym.read_mat('/content/gdrive/My Drive/My Files/Neuroimaging/Neuroimaging Datasets/cvind.mat')
print("Labels file in \".mat\" format converted to dictionary format:", rawMatData)

matDataList = list(rawMatData.values())
print("Labels converted to list format:", matDataList)

matDataArray = np.asarray(matDataList)
print("Labels converted to array format:", matDataArray, "\n")
shapedMatDataArray = matDataArray.reshape(-1, 3064, 1)
print("Reshaped labels in array format:\n", shapedMatDataArray, "\n")

matData = pd.DataFrame(matDataArray)
print("Labels converted to a Pandas DataFrame:")
matData #Prints out the DataFrame

#Viewing labels based on image number
def imgLabelCheck(n):
  callback = matData.at[0, n-1]
  print(f"Image Number {n} has the following Cancer Type: {classNames[callback]}.")
  return

pickImg = 1 #Choose an image number to look for its Cancer Type
imgLabelCheck(pickImg)

#Preparing the Datasets: Looping Train Set & Test Set
print("___________________________________________________________________________________\n")

train_images = np.array([imgMatricesNP[0:79]])
print("Training images range is:\n", train_images, "\n")

uppTrBn = len(train_images)
loqTrRng = 0
uppTrRng = 79
train_labels = np.asarray(matData.loc[:, loqTrRng:uppTrRng], dtype=float, order='A')
print("Training labels range is:", train_labels)

print("___________________________________________________________________________________\n")

test_images = np.array([imgMatricesNP[80:100]])
print("Testing images range is: \n", test_images, "\n")

uppTsBn = len(test_images)
loqTsRng = 80
uppTsRng = 100
test_labels = np.asarray(matData.loc[:, loqTsRng:uppTsRng], dtype=float, order='A')
print("Testing labels range is:", test_labels)

print("___________________________________________________________________________________")
#train_labels #Verify if the ranges are in fact NumPy arrays
#test_labels

#Defining the Convolutional Neural Network
model = models.Sequential()

model.add(layers.Conv2D(512, (3, 3), activation='relu', data_format="channels_first", input_shape=(1, 512, 512))) #The Input Layer
#model.add(layers.Dropout(0.5, noise_shape=None, seed=None)) #Optional Dropout Layer

model.add(layers.MaxPooling2D((2, 2), padding='same')) #MaxPooling Layer 1
model.add(layers.Conv2D(1024, (3, 3), activation='relu', padding='same')) #Hidden Convolutional Layer 1
model.add(layers.Dropout(0.5, noise_shape=None, seed=None)) #Optional Dropout Layer

model.add(layers.MaxPooling2D((2, 2), padding='same')) #MaxPooling Layer 2
model.add(layers.Conv2D(1024, (3, 3), activation='relu', padding='same')) #Hidden Convolutional Layer 2
#model.add(layers.Dropout(0.5, noise_shape=None, seed=None)) #Optional Dropout Layer

model.add(layers.MaxPooling2D((2, 2), padding='same')) #MaxPooling Layer 3
model.add(layers.Conv2D(1024, (3, 3), activation='relu', padding='same')) #Hidden Convolutional Layer 3
model.add(layers.Dropout(0.5, noise_shape=None, seed=None)) #Optional Dropout Layer

model.add(layers.MaxPooling2D((2, 2), padding='same')) #MaxPooling Layer 4
model.add(layers.Conv2D(1024, (3, 3), activation='relu', padding='same')) #Hidden Convolutional layer 4
#model.add(layers.Dropout(0.5, noise_shape=None, seed=None)) #Optional Dropout Layer

model.add(layers.MaxPooling2D((2, 2), padding='same')) #MaxPooling Layer 5
model.add(layers.Conv2D(1024, (3, 3), activation='relu', padding='same')) #Hidden Convolutional Layer 5
model.add(layers.Dropout(0.5, noise_shape=None, seed=None)) #Optional Dropout Layer

model.add(layers.MaxPooling2D((2, 2), padding='same')) #MaxPooling Layer 6
model.add(layers.Conv2D(1024, (3, 3), activation='relu', padding='same')) #Hidden Convolutional Layer 6
#model.add(layers.Dropout(0.5, noise_shape=None, seed=None)) #Optional Dropout Layer

model.add(layers.Flatten()) #The Flattening Layer

model.add(layers.Dense(512, activation='relu')) #Dense Layer 1
model.add(layers.Dense(256, activation='relu')) #Dense Layer 2
model.add(layers.Dense(128, activation='relu')) #Dense Layer 3
model.add(layers.Dense(64, activation='relu')) #Dense Layer 4
model.add(layers.Dense(32, activation='relu')) #Dense Layer 5
model.add(layers.Dense(16, activation='relu')) #Dense Layer 6

model.add(layers.Dense(outputSize, activation='softmax')) #The Output Layer

model.summary()

#Compiling the Convolutional Neural Network with an Optimizer
#The Adam Optimizer is ideal for biological image classification.
#The Optimizer automatically performs forward and backward propagation.

model.compile(
    optimizer='Adam',
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy'],
    loss_weights=None,
    sample_weight_mode=None,
    weighted_metrics=None,
    target_tensors=None
  )

print("The Neuroimaging Model has been successfully compiled.")

#Training the Convolutional Neural Network
history = model.fit(train_images, train_labels, epochs=10, batch_size=1, verbose=1,
                    validation_data=(test_images, test_labels))

print("\nThe Neuroimaging Model has been successfully trained.")

此页面上的每个代码框代表 Colab 或 Jupyter 笔记本的单个代码单元。再次欢迎并感谢所有帮助! (该模型尚未完全构建，但添加层仅用于实验。

Answer 1

添加行:

train_images = np.reshape(train_images, (-1,1,512,512))

在代码中的以下行之后

train_images = np.array([imgMatricesNP[0:79]])

获取单个图像的 input_shape=(1, 512, 512) 而不是 (79, 512, 512)，因为模型期望的输入形状为(1, 1, 512, 512)(根据尺寸(batch_size、 channel 、高度、宽度))，而您当前的代码提供的输入形状为 (1、79、512、512)。如果您有足够的计算资源，请将batch_size增加到8(比如说)，这样您的总输入形状将为(8, 1, 512, 512)。 p>

此外，对 test_images 执行类似的操作:

test_images = np.reshape(test_images, (-1,1,512,512))

行后:

test_images = np.array([imgMatricesNP[80:100]])

PS:此外，您的意图似乎是从输入 imgMatricesNP 中分割前 80 个图像。但是，使用 imgMatricesNP[0:79]，您只能获得前 79 个图像(因为 Python 中排除了切片的最后一个索引)。因此，更正为:

train_images = np.array([imgMatricesNP[0:80]])

并分配uppTrRng=80。

希望这有帮助! :)