如何在使用 tf.data 处理图像时提取 label

Question

I am using map function to apply preprocessing on dataset in order to read and extract labels from file paths using tf.data but it returns the same label for all images the file path is on the following shape where A is the label /content/drive /MyDrive/prom02/dataset/train/A0_jpg.rf.292a080422ba984985192f413101af41.jpg

images_ds = tf.data.Dataset.list_files('/content/drive/MyDrive/prom02/dataset/train/*', shuffle=True)

images_ds = images_ds.shuffle(200)

train_size = int(image_count*0.8)
train_ds = images_ds.take(train_size)
test_ds = images_ds.skip(train_size)
len(train_ds),len(test_ds)

def hot_encode(label):
    import string
    alphabet = list(string.ascii_uppercase)
    i=0
    while i<26:
      if label==alphabet[i]:
        label=i
        break
      i+=1
    label=tf.one_hot(label,26)
    return label
def get_label(file_path):
    import os
    label=(str(file_path).split('/')[-1][0])
    label= hot_encode(label)
    return label
def scale(image,label):
    return image/255, label
def process_image(file_path):
    label = tf.cast(get_label(file_path),tf.float32)
    img = tf.io.read_file(file_path) # load the raw data from the file as a string
    img = tf.image.decode_jpeg(img)
    img = tf.image.resize(img, [320, 320])
    return tf.cast(img,tf.float32), label

train_ds = train_ds.map(process_image).map(scale).batch(32).cache().prefetch(tf.data.AUTOTUNE)
test_ds = test_ds.map(process_image).map(scale).batch(32).prefetch(tf.data.AUTOTUNE)
for img,label in train_ds.take(1):
  print(label.numpy())

结果始终是字母 T 的编码张量为 label 而当我使用

for img in images_ds.take(2):
  print(get_label(img.numpy()))

它返回真正的 label

Answer 1

我建议进行一些修改，为了使您的代码干净，不要在 map function 中提取标签，创建路径和标签列表，试试这个：

from glob import glob
import string

alphabet = list(string.ascii_uppercase)
idx = range(26)
#create dict alphabet -> idx
dict_alpha = dict(tuple(zip(alphabet, idx)))

paths = glob('/content/drive/MyDrive/prom02/dataset/train/*.jpg')
labels = []
for path in paths:
  alpha = path.split("/")[-1][0]
  labels.append(dict_alpha[alpha])

#split paths, labels using train_test_split sklearn
train_ds = tf.data.Dataset.from_tensor_slices((paths, labels))
train_ds = train_ds.shuffle(len(paths))

def process_image(file_path, label):
  img = tf.io.read_file(file_path)
  img = tf.image.decode_jpeg(img)
  img = tf.cast(img,tf.float32)
  img = tf.image.resize(img, [320, 320], method = 'nearest')
  return img, tf.one_hot(label, 26)
        
    
batch = 1 #for test   

train_ds = train_ds.map(process_image).batch(batch).cache().prefetch(tf.data.AUTOTUNE)
    
for img, label in train_ds.take(1):
  print(label)

观察：

• 您可以在 process_image 中用 img = tf.cast(img, tf.float32) 替换比例映射
• tf.image.resize(img, [320, 320])默认情况下，调整大小的method is 'bilinear' ，这将改变像素值，对于图像我更喜欢tf.image.resize(img, [320, 320], method='nearest')对像素值没有影响，它基于重复。