Inceptionv3 分类错误

Posted 2023-03-13

【中文标题】Inceptionv3 分类错误

【英文标题】：Wrong classification with Inceptionv3

【发布时间】：2020-11-15 20:00:30

【问题描述】：

我使用了一个带有先前训练过的权重的初始模型，并使用 TENSORFLOW IN PRACTICE 导师 Laurence Moroney 提供的数据集添加了几个密集层来对马与人类进行分类。 我认为该模型已经完美训练，但它做出的预测始终是一匹马。

上图清楚地表明准确度相当不错。

我用来实时上传图片到colab并进行预测的代码如下：

    import numpy as np
    from google.colab import files
    from keras.preprocessing import image
    
    uploaded = files.upload()
    
    for fn in uploaded.keys():
     
      # predicting images
      path = '/content/' + fn
      img = image.load_img(path, target_size=(150, 150))
      x = image.img_to_array(img)
      x = np.expand_dims(x, axis=0)
    
      images = np.vstack([x])
      classes = model.predict(images, batch_size=10)
      print(classes[0])
      if classes[0]>0.5:
        print(fn + " is a horse")
      else:
        print(fn + " is a human")

用于训练模型的代码如下：

!wget --no-check-certificate \
  https://storage.googleapis.com/mledu-datasets/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5 \
  -O /tmp/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5

local_weights = '/tmp/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5'

from tensorflow.keras.applications.inception_v3 import InceptionV3

pre_trained_model = InceptionV3(include_top = False,
                                weights = None,
                                input_shape = (150, 150, 3))
pre_trained_model.load_weights(local_weights)

for layer in pre_trained_model.layers:
  layer.trainable = False

last_layer = pre_trained_model.get_layer('mixed7')
print(last_layer.output_shape)
last_output = last_layer.output

import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras import Model
from tensorflow.keras.optimizers import RMSprop

x = layers.Flatten()(last_output)
x = layers.Dense(1024, activation = 'relu')(x)
x = layers.Dropout(0.2)(x)
x = layers.Dense(1, activation = 'sigmoid')(x)

model = Model(pre_trained_model.input, x)
model.compile(optimizer = RMSprop(lr=0.0001),
              loss = 'binary_crossentropy',
              metrics = ['accuracy'])

!wget --no-check-certificate https://storage.googleapis.com/laurencemoroney-blog.appspot.com/horse-or-human.zip -O /tmp/horse-or-human.zip

# Get the Horse or Human Validation dataset
!wget --no-check-certificate https://storage.googleapis.com/laurencemoroney-blog.appspot.com/validation-horse-or-human.zip -O /tmp/validation-horse-or-human.zip 
  

import zipfile
import os

local_zip = '/tmp/horse-or-human.zip'
ref_zip = zipfile.ZipFile(local_zip, 'r')
ref_zip.extractall('/tmp/train')
ref_zip.close()

local_zip = '/tmp/validation-horse-or-human.zip'
ref_zip = zipfile.ZipFile(local_zip, 'r')
ref_zip.extractall('/tmp/validate')
ref_zip.close()

train_dir = '/tmp/train'
validate_dir = '/tmp/validate'

from tensorflow.keras.preprocessing.image import ImageDataGenerator

train_datagen = ImageDataGenerator(rescale = 1./255,
                                   width_shift_range = 0.2,
                                   height_shift_range = 0.2,
                                   rotation_range = 0.2,
                                   shear_range = 0.2,
                                   horizontal_flip = True,
                                   zoom_range = 0.2)
training_generator = train_datagen.flow_from_directory(train_dir,
                                                       target_size = (150, 150),
                                                       class_mode = 'binary',
                                                       batch_size = 64)

test_datagen = ImageDataGenerator(rescale = 1./255)
validation_generator = test_datagen.flow_from_directory(validate_dir,
                                                        target_size = (150, 150),
                                                        class_mode = 'binary',
                                                        batch_size = 64)

history = model.fit_generator(training_generator,
                    validation_data = validation_generator,
                    epochs = 20,
                    verbose = 2,
                    )

【问题讨论】：

能否请您发布您的完整代码。

我已经更新了问题，请查看上方！

【参考方案1】：

主要问题：原因是缺少重新缩放。当使用图像数据生成器时，给出了重新缩放因子。但是，当手动上传图像时，缺少重新缩放的代码。你必须添加一个x = x/255.0。

您必须使用classes<0.5。因此，如果概率 > 0.5，那么我们将其归类为人，而不是您指定的马。您可以使用training_generator.class_indices 进行检查，这将为您提供：

考虑以下两个例子（我从数据集中拿一张马的图片和一张人的图片来说明它）：

错误的结果：

# predicting images
path = '/tmp/train/humans/human01-00.png'
img = image.load_img(path, target_size=(150, 150))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
# = x/255.0

images = np.vstack([x])
classes = model.predict(images, batch_size=10)
print(classes)
if classes<0.5:
  print(path + " is a horse")
else:
  print(path + " is a human")
path = "/tmp/train/horses/horse01-0.png"
img = image.load_img(path, target_size=(150, 150))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
#x = x/255.0

images = np.vstack([x])
classes = model.predict(images, batch_size=10)
print(classes)
if classes<0.5:
  print(path + " is a horse")
else:
  print(path + " is a human")

输出：

您可以看到两者都被错误分类。然而，被错误地归类为人类。你错了，你的模型总是预测人类，因为 1 是人类，0 是马。

正确结果：

# predicting images
path = '/tmp/train/humans/human01-00.png'
img = image.load_img(path, target_size=(150, 150))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = x/255.0

images = np.vstack([x])
classes = model.predict(images, batch_size=10)
print(classes)
if classes<0.5:
  print(path + " is a horse")
else:
  print(path + " is a human")
path = "/tmp/train/horses/horse01-0.png"
img = image.load_img(path, target_size=(150, 150))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = x/255.0

images = np.vstack([x])
classes = model.predict(images, batch_size=10)
print(classes)
if classes<0.5:
  print(path + " is a horse")
else:
  print(path + " is a human")

输出：

所以你可以看到它被正确分类了。

因此，请记住您的“>”是错误的。如果它被分类为 1 (>0.5)，它被分类为人类。否则，如果小于 0.5 则作为马。

（您可能被课程中的官方剪纸示例误导了，代码中也缺少该示例。这是一个报告的问题。）

【讨论】：

非常感谢！我还没有尝试过剪刀石头布的例子，但我会记住这一点！