TensorFlow 图像二元分类器在训练后无法有效工作
TensorFlow image binary classifier not working efficiently after training
我正在对是否是蜜蜂的图像进行二元分类器试验。我收集了一个包含 6 个类别的 12,000 张图像的数据集，其中一个是蜜蜂。所以我有一列 is_bee，其值为 "0" 和 "1"匹配 “不是蜜蜂” 和 “它是蜜蜂”。我正在训练分类器，当我尝试将经过训练的模型应用到任何图像上时，即使是其中一张经过训练的图像，它（几乎）以 73.11% 的置信度给我唯一的值“0”。我的代码如下：
import pandas as pd
import numpy as np
import sys
import os
import random
from pathlib import Path
import imageio
import skimage
import skimage.io
import skimage.transform
import matplotlib.pyplot as plt
import seaborn as sns
import scipy
from sklearn.model_selection import train_test_split
from sklearn import metrics
from keras import optimizers, losses
from keras.models import Sequential
from keras.layers import Dense, Conv2D, Flatten, MaxPool2D, Dropout, BatchNormalization,LeakyReLU
from keras.preprocessing.image import ImageDataGenerator
from keras.callbacks import ModelCheckpoint, Callback, EarlyStopping, ReduceLROnPlateau
from tensorflow.keras.utils import to_categorical
import tensorflow
np.random.seed(42)
tensorflow.random.set_seed(42)
# Global variables
imgs_folder='images/'
img_width=100
img_height=100
img_channels=3
def read_img(file):
img = skimage.io.imread(file)
img = skimage.transform.resize(img, (img_width, img_height), mode='reflect')
return img[:,:,:img_channels]
images=pd.read_csv('/content/gdrive/MyDrive/Colab Notebooks/beeID/images.csv',
index_col=False, sep=";",
dtype={'type':'category', 'is_bee':'category'})
images['file'] = imgs_folder+images['file']
# Cannot impute nans, drop them
images.dropna(inplace=True)
# Some image files don't exist. Leave only bees with available images.
img_exists = images['file'].apply(lambda f: os.path.exists(f))
images = images[img_exists]
images['type'] = images['type'].astype('category')
images['is_bee'] = images['is_bee'].astype('category')
def split_balance(df, field_name):
train_df, test_df = train_test_split(df, random_state=23)
train_df, val_df = train_test_split(train_df, test_size=0.1, random_state=23)
ncat_bal = int(len(train_df)/train_df[field_name].cat.categories.size)
train_df_bal = train_df.groupby(field_name, as_index=False).apply(lambda g: g.sample(ncat_bal, replace=True)).reset_index(drop=True)
return(train_df_bal, val_df, test_df)
def prepare2train(train_p2t_df, val_p2t_df, test_p2t_df, field_name):
# Train data
train_X = np.stack(train_p2t_df['file'].apply(read_img))
#train_y = to_categorical(train_bees[field_name].values)
train_y = pd.get_dummies(train_p2t_df[field_name], drop_first=False)
# Validation during training data to calc val_loss metric
val_X = np.stack(val_p2t_df['file'].apply(read_img))
#val_y = to_categorical(val_bees[field_name].values)
val_y = pd.get_dummies(val_p2t_df[field_name], drop_first=False)
# Test data
test_X = np.stack(test_p2t_df['file'].apply(read_img))
#test_y = to_categorical(test_bees[field_name].values)
test_y = pd.get_dummies(test_p2t_df[field_name], drop_first=False)
# Data augmentation
generator = ImageDataGenerator(
featurewise_center=False, # set input mean to 0 over the dataset
samplewise_center=False, # set each sample mean to 0
featurewise_std_normalization=False, # divide inputs by std of the dataset
samplewise_std_normalization=False, # divide each input by its std
zca_whitening=False, # apply ZCA whitening
rotation_range=180, # randomly rotate images in the range (degrees, 0 to 180)
zoom_range=0.1, # Randomly zoom image
width_shift_range=0.2, # randomly shift images horizontally (fraction of total width)
height_shift_range=0.2, # randomly shift images vertically (fraction of total height)
horizontal_flip=True, # randomly flip images
vertical_flip=True)
generator.fit(train_X)
return (generator, train_X, val_X, test_X, train_y, val_y, test_y)
def eval_model(training, model, test_X, test_y, field_name):
## Trained model analysis and evaluation
f, ax = plt.subplots(2,1, figsize=(5,5))
ax[0].plot(training.history['loss'], label="Loss")
ax[0].plot(training.history['val_loss'], label="Validation loss")
ax[0].set_title('%s: loss' % field_name)
ax[0].set_xlabel('Epoch')
ax[0].set_ylabel('Loss')
ax[0].legend()
# Accuracy
ax[1].plot(training.history['accuracy'], label="Accuracy")
ax[1].plot(training.history['val_accuracy'], label="Validation accuracy")
ax[1].set_title('%s: accuracy' % field_name)
ax[1].set_xlabel('Epoch')
ax[1].set_ylabel('Accuracy')
ax[1].legend()
plt.tight_layout()
plt.show()
# Accuracy by subspecies
test_pred = model.predict(test_X)
acc_by_subspecies = np.logical_and((test_pred > 0.5), test_y).sum()/test_y.sum()
acc_by_subspecies.plot(kind='bar', title='Accuracy by %s' % field_name)
plt.ylabel('Accuracy')
plt.show()
# Print metrics
print("Classification report")
test_pred = np.argmax(test_pred, axis=1)
test_truth = np.argmax(test_y.values, axis=1)
# Loss function and accuracy
test_res = model.evaluate(test_X, test_y.values, verbose=1)
# Split/balance and plot the result
train_bees_bal, val_bees, test_bees = split_balance(bees, 'health')
# Split/balance and plot the result
train_images_bal, val_images, test_images = split_balance(images, 'type')
# Will use balanced dataset as main
train_images = train_images_bal
generator_images, train_images_X, val_images_X, test_images_X, train_images_y, val_images_y, test_images_y = prepare2train(train_images, val_images, test_images, 'is_bee')
# We'll stop training if no improvement after some epochs
earlystopper_images = EarlyStopping(monitor='val_accuracy', patience=20, verbose=1)
# Save the best model during the training
checkpointer_images = ModelCheckpoint('model_images.h5'
,monitor='val_accuracy'
,verbose=1
,save_best_only=True
,save_weights_only=True)
# Build CNN model
model_images=Sequential()
model_images.add(Conv2D(5, kernel_size=3, input_shape=(img_width, img_height,3), activation='relu', padding='same'))
model_images.add(MaxPool2D(2))
model_images.add(Conv2D(10, kernel_size=3, activation='relu', padding='same'))
model_images.add(Dropout(0.25))
model_images.add(Flatten())
model_images.add(Dropout(0.5))
model_images.add(Dense(train_images_y.columns.size, activation='sigmoid', name='preds'))
# show model summary
model_images.summary()
model_images.compile(loss=losses.binary_crossentropy,optimizer='adam',metrics=['accuracy'])
# Train
training_images = model_images.fit_generator(generator_images.flow(train_images_X,train_images_y, batch_size=60)
,epochs=100
,validation_data=(val_images_X, val_images_y)
,steps_per_epoch=10
,callbacks=[earlystopper_images, checkpointer_images])
# Get the best saved weights
model_images.load_weights('model_images.h5')
eval_model(training_images, model_images, test_images_X, test_images_y, 'is_bee')
scores = model_images.evaluate(test_images_X, test_images_y, verbose=0)
print("Accuracy: %.2f%%" % (scores[1]*100))
import numpy as np
from google.colab import files
from keras.preprocessing import image
for fn in os.listdir('/content/bee_imgs'):
# predicting images
path = '/content/bee_imgs/' + fn
img = image.load_img(path, target_size=(100, 100))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
checkImages = np.vstack([x])
classes_images = model_images.predict(checkImages)
score = tensorflow.nn.softmax(classes_images[0])
print(
"This image {} most likely belongs to {} with a {:.2f} percent confidence."
.format(fn, test_images_y.columns[np.argmax(score)], 100 * np.max(score))
)
以上对已经训练过的图像进行测试的结果（但当我尝试使用未训练过的图像时也会出现）是这样的：
This image 016_252.png most likely belongs to 0 with a 73.11 percent confidence.
This image 031_117.png most likely belongs to 0 with a 73.11 percent confidence.
This image 019_1026.png most likely belongs to 0 with a 73.11 percent confidence.
This image 008_243.png most likely belongs to 1 with a 73.11 percent confidence.
...
This image 039_016.png most likely belongs to 0 with a 73.11 percent confidence.
This image 022_364.png most likely belongs to 0 with a 73.11 percent confidence.
This image 022_274.png most likely belongs to 0 with a 73.11 percent confidence.
因此，尽管该模型在预测图像是否为蜜蜂方面的准确率已达到 96%，但当我将其应用于蜜蜂的任何图像时，它很少将其识别为蜜蜂。是我的模型构建有问题还是我尝试应用它的时候有问题？
删除 softmax 函数和 np.argmax 后，您应该只使用在预测训练期间使用的相同 read_img 函数，它应该没问题。
我正在对是否是蜜蜂的图像进行二元分类器试验。我收集了一个包含 6 个类别的 12,000 张图像的数据集，其中一个是蜜蜂。所以我有一列 is_bee，其值为 "0" 和 "1"匹配 “不是蜜蜂” 和 “它是蜜蜂”。我正在训练分类器，当我尝试将经过训练的模型应用到任何图像上时，即使是其中一张经过训练的图像，它（几乎）以 73.11% 的置信度给我唯一的值“0”。我的代码如下：
import pandas as pd
import numpy as np
import sys
import os
import random
from pathlib import Path
import imageio
import skimage
import skimage.io
import skimage.transform
import matplotlib.pyplot as plt
import seaborn as sns
import scipy
from sklearn.model_selection import train_test_split
from sklearn import metrics
from keras import optimizers, losses
from keras.models import Sequential
from keras.layers import Dense, Conv2D, Flatten, MaxPool2D, Dropout, BatchNormalization,LeakyReLU
from keras.preprocessing.image import ImageDataGenerator
from keras.callbacks import ModelCheckpoint, Callback, EarlyStopping, ReduceLROnPlateau
from tensorflow.keras.utils import to_categorical
import tensorflow
np.random.seed(42)
tensorflow.random.set_seed(42)
# Global variables
imgs_folder='images/'
img_width=100
img_height=100
img_channels=3
def read_img(file):
img = skimage.io.imread(file)
img = skimage.transform.resize(img, (img_width, img_height), mode='reflect')
return img[:,:,:img_channels]
images=pd.read_csv('/content/gdrive/MyDrive/Colab Notebooks/beeID/images.csv',
index_col=False, sep=";",
dtype={'type':'category', 'is_bee':'category'})
images['file'] = imgs_folder+images['file']
# Cannot impute nans, drop them
images.dropna(inplace=True)
# Some image files don't exist. Leave only bees with available images.
img_exists = images['file'].apply(lambda f: os.path.exists(f))
images = images[img_exists]
images['type'] = images['type'].astype('category')
images['is_bee'] = images['is_bee'].astype('category')
def split_balance(df, field_name):
train_df, test_df = train_test_split(df, random_state=23)
train_df, val_df = train_test_split(train_df, test_size=0.1, random_state=23)
ncat_bal = int(len(train_df)/train_df[field_name].cat.categories.size)
train_df_bal = train_df.groupby(field_name, as_index=False).apply(lambda g: g.sample(ncat_bal, replace=True)).reset_index(drop=True)
return(train_df_bal, val_df, test_df)
def prepare2train(train_p2t_df, val_p2t_df, test_p2t_df, field_name):
# Train data
train_X = np.stack(train_p2t_df['file'].apply(read_img))
#train_y = to_categorical(train_bees[field_name].values)
train_y = pd.get_dummies(train_p2t_df[field_name], drop_first=False)
# Validation during training data to calc val_loss metric
val_X = np.stack(val_p2t_df['file'].apply(read_img))
#val_y = to_categorical(val_bees[field_name].values)
val_y = pd.get_dummies(val_p2t_df[field_name], drop_first=False)
# Test data
test_X = np.stack(test_p2t_df['file'].apply(read_img))
#test_y = to_categorical(test_bees[field_name].values)
test_y = pd.get_dummies(test_p2t_df[field_name], drop_first=False)
# Data augmentation
generator = ImageDataGenerator(
featurewise_center=False, # set input mean to 0 over the dataset
samplewise_center=False, # set each sample mean to 0
featurewise_std_normalization=False, # divide inputs by std of the dataset
samplewise_std_normalization=False, # divide each input by its std
zca_whitening=False, # apply ZCA whitening
rotation_range=180, # randomly rotate images in the range (degrees, 0 to 180)
zoom_range=0.1, # Randomly zoom image
width_shift_range=0.2, # randomly shift images horizontally (fraction of total width)
height_shift_range=0.2, # randomly shift images vertically (fraction of total height)
horizontal_flip=True, # randomly flip images
vertical_flip=True)
generator.fit(train_X)
return (generator, train_X, val_X, test_X, train_y, val_y, test_y)
def eval_model(training, model, test_X, test_y, field_name):
## Trained model analysis and evaluation
f, ax = plt.subplots(2,1, figsize=(5,5))
ax[0].plot(training.history['loss'], label="Loss")
ax[0].plot(training.history['val_loss'], label="Validation loss")
ax[0].set_title('%s: loss' % field_name)
ax[0].set_xlabel('Epoch')
ax[0].set_ylabel('Loss')
ax[0].legend()
# Accuracy
ax[1].plot(training.history['accuracy'], label="Accuracy")
ax[1].plot(training.history['val_accuracy'], label="Validation accuracy")
ax[1].set_title('%s: accuracy' % field_name)
ax[1].set_xlabel('Epoch')
ax[1].set_ylabel('Accuracy')
ax[1].legend()
plt.tight_layout()
plt.show()
# Accuracy by subspecies
test_pred = model.predict(test_X)
acc_by_subspecies = np.logical_and((test_pred > 0.5), test_y).sum()/test_y.sum()
acc_by_subspecies.plot(kind='bar', title='Accuracy by %s' % field_name)
plt.ylabel('Accuracy')
plt.show()
# Print metrics
print("Classification report")
test_pred = np.argmax(test_pred, axis=1)
test_truth = np.argmax(test_y.values, axis=1)
# Loss function and accuracy
test_res = model.evaluate(test_X, test_y.values, verbose=1)
# Split/balance and plot the result
train_bees_bal, val_bees, test_bees = split_balance(bees, 'health')
# Split/balance and plot the result
train_images_bal, val_images, test_images = split_balance(images, 'type')
# Will use balanced dataset as main
train_images = train_images_bal
generator_images, train_images_X, val_images_X, test_images_X, train_images_y, val_images_y, test_images_y = prepare2train(train_images, val_images, test_images, 'is_bee')
# We'll stop training if no improvement after some epochs
earlystopper_images = EarlyStopping(monitor='val_accuracy', patience=20, verbose=1)
# Save the best model during the training
checkpointer_images = ModelCheckpoint('model_images.h5'
,monitor='val_accuracy'
,verbose=1
,save_best_only=True
,save_weights_only=True)
# Build CNN model
model_images=Sequential()
model_images.add(Conv2D(5, kernel_size=3, input_shape=(img_width, img_height,3), activation='relu', padding='same'))
model_images.add(MaxPool2D(2))
model_images.add(Conv2D(10, kernel_size=3, activation='relu', padding='same'))
model_images.add(Dropout(0.25))
model_images.add(Flatten())
model_images.add(Dropout(0.5))
model_images.add(Dense(train_images_y.columns.size, activation='sigmoid', name='preds'))
# show model summary
model_images.summary()
model_images.compile(loss=losses.binary_crossentropy,optimizer='adam',metrics=['accuracy'])
# Train
training_images = model_images.fit_generator(generator_images.flow(train_images_X,train_images_y, batch_size=60)
,epochs=100
,validation_data=(val_images_X, val_images_y)
,steps_per_epoch=10
,callbacks=[earlystopper_images, checkpointer_images])
# Get the best saved weights
model_images.load_weights('model_images.h5')
eval_model(training_images, model_images, test_images_X, test_images_y, 'is_bee')
scores = model_images.evaluate(test_images_X, test_images_y, verbose=0)
print("Accuracy: %.2f%%" % (scores[1]*100))
import numpy as np
from google.colab import files
from keras.preprocessing import image
for fn in os.listdir('/content/bee_imgs'):
# predicting images
path = '/content/bee_imgs/' + fn
img = image.load_img(path, target_size=(100, 100))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
checkImages = np.vstack([x])
classes_images = model_images.predict(checkImages)
score = tensorflow.nn.softmax(classes_images[0])
print(
"This image {} most likely belongs to {} with a {:.2f} percent confidence."
.format(fn, test_images_y.columns[np.argmax(score)], 100 * np.max(score))
)
以上对已经训练过的图像进行测试的结果（但当我尝试使用未训练过的图像时也会出现）是这样的：
This image 016_252.png most likely belongs to 0 with a 73.11 percent confidence.
This image 031_117.png most likely belongs to 0 with a 73.11 percent confidence.
This image 019_1026.png most likely belongs to 0 with a 73.11 percent confidence.
This image 008_243.png most likely belongs to 1 with a 73.11 percent confidence.
...
This image 039_016.png most likely belongs to 0 with a 73.11 percent confidence.
This image 022_364.png most likely belongs to 0 with a 73.11 percent confidence.
This image 022_274.png most likely belongs to 0 with a 73.11 percent confidence.
因此，尽管该模型在预测图像是否为蜜蜂方面的准确率已达到 96%，但当我将其应用于蜜蜂的任何图像时，它很少将其识别为蜜蜂。是我的模型构建有问题还是我尝试应用它的时候有问题？
删除 softmax 函数和 np.argmax 后，您应该只使用在预测训练期间使用的相同 read_img 函数，它应该没问题。