Question

我有以下脚本，正在尝试使用VGG16模型预测图像的结果（即转移学习）：

from keras.applications import VGG16
from keras.preprocessing.image import ImageDataGenerator
from keras.preprocessing.image import load_img
from keras.preprocessing.image import img_to_array
from keras.applications.vgg16 import preprocess_input
from keras.applications.vgg16 import decode_predictions
from keras.layers import Input, Flatten, Dense
from keras.models import Model
from keras import models
from keras import layers
from keras import optimizers
import ssl
import os
import cv2
import numpy as np
import matplotlib
# Force matplotlib to not use any Xwindows backend
matplotlib.use('Agg')
import matplotlib.pyplot as plt

# path to the training, validation, and testing directories

train_directory = '/train'
validation_directory = '/valid'
test_directory = '/test'
results_directory = '/results'
number_of_training_samples = 1746
number_of_validation_samples = 108
number_of_test_samples = 510
batch_size = 20

ssl._create_default_https_context = ssl._create_unverified_context

# get back the convolutional part of a VGG network trained on ImageNet
conv_base = VGG16(weights='imagenet',include_top=False,input_shape=(512,512,3))
conv_base.summary() 

# preprocess the data

# rescale images by the factor 1/255
train_data = ImageDataGenerator(rescale=1.0/255)
validation_data = ImageDataGenerator(rescale=1.0/255)
test_data = ImageDataGenerator(rescale=1.0/255)

train_features = np.zeros(shape=(number_of_training_samples,16,16,512))
train_labels = np.zeros(shape=(number_of_training_samples))

train_generator = train_data.flow_from_directory(
    train_directory,
    target_size=(512,512),
    batch_size=batch_size,
    class_mode='binary',
    shuffle=True)

i = 0
for inputs_batch, labels_batch in train_generator:
    features_batch = conv_base.predict(inputs_batch)
    train_features[i*batch_size:(i+1)*batch_size] = features_batch
    train_labels[i*batch_size:(i+1)*batch_size] = labels_batch
    i += 1
    if i * batch_size >= number_of_training_samples:
        break

train_features = np.reshape(train_features, (number_of_training_samples,16*16*512))

validation_features = np.zeros(shape=(number_of_validation_samples,16,16,512))
validation_labels = np.zeros(shape=(number_of_validation_samples))

validation_generator = validation_data.flow_from_directory(
    validation_directory,
    target_size=(512,512),
    batch_size=batch_size,
    class_mode='binary',
    shuffle=False)

i = 0
for inputs_batch, labels_batch in validation_generator:
    features_batch = conv_base.predict(inputs_batch)
    validation_features[i*batch_size:(i+1)*batch_size] = features_batch
    validation_labels[i*batch_size:(i+1)*batch_size] = labels_batch
    i += 1
    if i * batch_size >= number_of_validation_samples:
        break

validation_features = np.reshape(validation_features, (number_of_validation_samples,16*16*512))

test_generator = test_data.flow_from_directory(
    test_directory,
    target_size=(512,512),
    batch_size=batch_size,
    class_mode='binary',
    shuffle=False)

# define the Convolutional Neural Network (CNN) model
model = models.Sequential() 
model.add(layers.Dense(1024,activation='relu',input_shape=(1,16,16,512)))
model.add(layers.Dense(1,activation='sigmoid'))

# compile the model

model.compile(loss='binary_crossentropy',
    optimizer=optimizers.Adam(lr=0.01),
    metrics=['acc'])

# fit the model to the data
history = model.fit(train_features,
    train_labels,
    epochs=1,
    batch_size=batch_size,
    validation_data=(validation_features,validation_labels))

# save the model
model.save('benign_and_melanoma_from_scratch.h5')

# generate accuracy and loss curves for the training process (history of accuracy and loss)
acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']

number_of_epochs = range(1,len(acc)+1)

plt.plot(number_of_epochs, acc, 'r', label='Training accuracy')
plt.plot(number_of_epochs, val_acc, 'g', label='Validation accuracy')
plt.title('Training and validation accuracy')
plt.legend()
plt.savefig('accuracy.png')

plt.close()

plt.plot(number_of_epochs, loss, 'r', label='Training loss')
plt.plot(number_of_epochs, val_loss, 'g', label='Validation loss')
plt.title('Training and validation loss')
plt.legend()
plt.savefig('loss.png')

# evaluate the model

# predict classes
for root, dirs, files in os.walk(test_directory):
    for file in files:
        img = cv2.imread(root + '/' + file)
        img = cv2.resize(img,(512,512),interpolation=cv2.INTER_AREA)
        img = np.expand_dims(img, axis=0)
        img = img/255.0
        feature_value = conv_base.predict(img)
        feature_value= np.reshape(feature_value,(1,512,512,3))
        img_class = model.predict_classes(feature_value)
        prediction = img_class[0]

但是，出现以下错误：

ValueError: Error when checking input: expected dense_1_input to have 5 dimensions, but got array with shape (1746, 131072)

在线：

validation_data=(validation_features,validation_labels))

关于如何解决此问题的任何想法？

谢谢。

Answer 1

我从您的代码中发现了几件事：

train_generator的目标尺寸(512, 512)与输入形状(16*16*512)不同。但是由于您不适合将其放入模型中，因此这无关紧要。这就提出了一个问题，为什么要将它放在代码中。
由于您已经将输入training_features重塑为(number_of_training_samples,16*16*512)（等于(1746, 131072)），所以您实际上并不需要（也不应该）指定{{1} }放在input_shape=(1,16,16,512)层中。

检查输入时出错：预期density_1_input具有5个维，但数组的形状为（1746，131072）

1 个答案: