TensorFlow |创建图像数据集，按文件名标记

Question

我正在尝试创建 Tensorflow 数据集来训练我的模型。我有一个充满标记照片的文件夹，标记是文件名的一部分。

您是否有合理的方法来加载数据集进行训练而不将其拆分到不同的目录？

例子： 对于文件：

• ./dataset/path/img0_cat.bmp
• ./dataset/path/img1_dog.bmp
• ./dataset/path/img2_horse.bmp
• ./dataset/path/img3_cat.bmp
• ./dataset/path/img4_dog.bmp
• ./数据集/路径/img5_horse.bmp
• ./dataset/path/img6_dog.bmp
• ./dataset/path/img7_cat.bmp
• ./dataset/path/img8_horse.bmp
• ./dataset/path/img9_cat.bmp
• ./dataset/path/img10_dog.bmp

预期输出： tf.Dataset 标记为 (cat,狗、马）

Answer 1

感谢大家的回复

我解决这个问题的方法如下：

classNames = ['dog', 'cat', 'horse']

def getLabel(file_path):
    # Convert the path to a list of path components
    fileName = tf.strings.split(file_path, os.path.sep)[-1]
    # get label name from filename
    className = tf.strings.split(fileName, '_')[1]
    className = tf.strings.split(className, '.')[0]
    # get one_hot vector boolean
    one_hot = className == classNames
    # cast vector type to integer 
    return tf.cast(one_hot, dtype=tf.int8, name=None)

def getImage(file_path):
    # Load the raw data from the file as a string
    img = tf.io.read_file(file_path)
    # Convert the compressed string to a 3D uint8 tensor
    img = tf.io.decode_bmp(img, channels=3)
    # cast tf.Tensor type to uint8 
    return tf.cast(img, dtype=tf.uint8, name=None)

def process_path(file_path):
    label = getLabel(file_path)
    img = getImage(file_path)
    return img, label

path = './dataset/path/*.bmp'
ds = tf.data.Dataset.list_files(path)
ds = ds.map(process_path)

在此过程结束时，您将获得一个 Tensorflow 可训练数据集（批量大小需要更多配置，请参阅参考资料），标签为 one-hot 矢量。

运行时：

for image, label in ds.take(5):
     imageShape = image.numpy().shape
     label = label.numpy()
     labelName = class_names[np.argmax(label)]
     print('Image Shape: {}, Label: {}, LabelName: {}'.format(imageShape, label, labelName))

你得到：

Image Shape: (180, 180, 3), Label: [1 0 0], LabelName: dog
Image Shape: (180, 180, 3), Label: [0 1 0], LabelName: cat
Image Shape: (180, 180, 3), Label: [0 0 1], LabelName: horse
Image Shape: (180, 180, 3), Label: [1 0 0], LabelName: dog
Image Shape: (180, 180, 3), Label: [0 1 0], LabelName: cat

参考：
https://www.tensorflow.org/tutorials/load_data/images

Answer 2

您可以尝试根据您在训练集中使用的任何ID来分配ID，并根据您使用的任何ID来收集路径。

如果您使用的是tensorflow，则 dataset 文档具有信息的方法。加载数据。具体来说，

dataset_dog = tf.data.dataset.list_files（“ ./ dataset/path/path/*dog.bmp）

Answer 3

这是整个辣酱玉米饼馅。我更喜欢将 pandas 数据集与 ImageDataGenerator.flow_from_dataframe 一起使用，因为它很灵活。我创建了一个目录，其中包含 10 张起重机图像和 10 张信天翁图像。文件名的形式为
0_crane.jpf、1_crane.jpg 等... 10_albatross、11_albatross ..... 下面的代码处理此目录。创建一个数据帧 df，然后将其拆分为 train_df、valid_df 和 test_df。然后为train_gen、test_gen和valid_gen创建3个图像数据生成器。我使用了我常用的标准模型并训练了该模型。然后以 100% 的准确度评估测试步骤。代码如下

import numpy as np
import pandas as pd
import os
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
from tensorflow.keras.preprocessing.image import ImageDataGenerator
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.layers import Dense, Activation,Dropout,Conv2D, MaxPooling2D,BatchNormalization
from tensorflow.keras.optimizers import Adam, Adamax
from tensorflow.keras.metrics import categorical_crossentropy
from tensorflow.keras import regularizers
from tensorflow.keras.models import Model

sdir = r'c:\temp\single'
filepaths=[]
labels=[]
flist=os.listdir(sdir)
for f in flist:
    fpath=os.path.join(sdir,f)   
    filepaths.append(fpath)
    index1=f.rfind('_')+1
    index2=f.rfind('.')    
    klass=f[index1:index2]     
    labels.append(klass)
Fseries=pd.Series(filepaths, name='filepaths')
Lseries=pd.Series(labels, name='labels')
df=pd.concat([Fseries, Lseries], axis=1)
# now you can use train_test_split to create a train_df, a test_df and a valid_df
trsplit=.8
vsplit=.1
dsplit=vsplit/(1-trsplit)
train_df, dummy_df=train_test_split (df, train_size=trsplit, shuffle=True, random_state=123, stratify=df['labels'])
valid_df, test_df= train_test_split(dummy_df, train_size=dsplit, shuffle=True, random_state=123, stratify =dummy_df['labels'])
print ('train_df length: ', len(train_df), ' test_df length: ', len(test_df) ,'  valid_df length: ', len(valid_df))

# create a train_gen,  a valid_gen and a test_gen
# for trgen you can specify augmentations like horizontal_flip, vertical_flip etc
img_size=(224,224)
batch_size=16
trgen=ImageDataGenerator(horizontal_flip=True,rotation_range=20, width_shift_range=.2,
                                  height_shift_range=.2, zoom_range=.2   )
tvgen=ImageDataGenerator()
train_gen=trgen.flow_from_dataframe(train_df, x_col='filepaths', y_col='labels',target_size=img_size,
                                   class_mode='categorical', color_mode='rgb', shuffle=True, batch_size=batch_size)
valid_gen=tvgen.flow_from_dataframe(valid_df, x_col='filepaths', y_col='labels',target_size=img_size,
                                   class_mode='categorical', color_mode='rgb', shuffle=False, batch_size=2)
test_gen=tvgen.flow_from_dataframe(test_df, x_col='filepaths', y_col='labels',target_size=img_size,
                                   class_mode='categorical', color_mode='rgb', shuffle=False, batch_size=2)
train_files=train_gen.filenames
classes=list(train_gen.class_indices.keys())
class_count=len(classes)
labels=train_gen.labels
images, labels=next(train_gen)
plt.figure(figsize=(20, 12))
for i in range (len(images)):
    plt.subplot(4,4, i+1) 
    index=np.argmax(labels[i])
    class_name=classes[index]
    plt.title(class_name, color='yellow', fontsize=18)   
    plt.axis('off')
    plt.imshow(images[i]/255)
plt.show()
# build a model
img_shape=(img_size[0], img_size[1], 3)
model_name='EfficientNetB5'
base_model=tf.keras.applications.efficientnet.EfficientNetB5(include_top=False, weights="imagenet",input_shape=img_shape, pooling='max') 
# Note you are always told NOT to make the base model trainable initially- that is WRONG you get better results leaving it trainable
base_model.trainable=True
x=base_model.output
x=BatchNormalization(axis=-1, momentum=0.99, epsilon=0.001 )(x)
x = Dense(1024, kernel_regularizer = regularizers.l2(l = 0.016),activity_regularizer=regularizers.l1(0.006),
                bias_regularizer=regularizers.l1(0.006) ,activation='relu')(x)
x=Dropout(rate=.3, seed=123)(x)
x = Dense(128, kernel_regularizer = regularizers.l2(l = 0.016),activity_regularizer=regularizers.l1(0.006),
                bias_regularizer=regularizers.l1(0.006) ,activation='relu')(x)
x=Dropout(rate=.45, seed=123)(x)        
output=Dense(class_count, activation='softmax')(x)
model=Model(inputs=base_model.input, outputs=output)
lr=.001 # start with this learning rate
model.compile(Adamax(learning_rate=lr), loss='categorical_crossentropy', metrics=['accuracy']) 

epochs=5
history=model.fit(x=train_gen,  epochs=epochs, verbose=1,  validation_data=valid_gen,
               validation_steps=None,  shuffle=False,  initial_epoch=0)

loss, acc =model.evaluate(test_gen)
print (' accuracy on test set is: ', acc* 100, '%')