關(guān)于Keras數(shù)據(jù)增強(qiáng)ImageDataGenerator多輸入多輸出的案例

import os 
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" 
os.environ["CUDA_VISIBLE_DEVICES"]=""
import sys
import gc
import time
import cv2
import random
import numpy as np
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
from tqdm import tqdm

from random_eraser import get_random_eraser
from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img

datagen = ImageDataGenerator(
  rotation_range=20,   #旋轉(zhuǎn)
  width_shift_range=0.1,  #水平位置平移
#   height_shift_range=0.2,  #上下位置平移
  shear_range=0.5,    #錯(cuò)切變換，讓所有點(diǎn)的x坐標(biāo)(或者y坐標(biāo))保持不變，而對(duì)應(yīng)的y坐標(biāo)(或者x坐標(biāo))則按比例發(fā)生平移
  zoom_range=[0.9,0.9],  # 單方向縮放，當(dāng)一個(gè)數(shù)值時(shí)兩個(gè)方向等比例縮放，參數(shù)為list時(shí)長(zhǎng)寬不同程度縮放。參數(shù)大于0小于1時(shí)，執(zhí)行的是放大操作，當(dāng)參數(shù)大于1時(shí)，執(zhí)行的是縮小操作。
  channel_shift_range = 40, #偏移通道數(shù)值，改變圖片顏色，越大顏色越深
  horizontal_flip=True,  #水平翻轉(zhuǎn)，垂直翻轉(zhuǎn)vertical_flip
  fill_mode='nearest',   #操作導(dǎo)致圖像缺失時(shí)填充方式?！癱onstant”、“nearest”（默認(rèn)）、“reflect”和“wrap”
  preprocessing_function = get_random_eraser(p=0.7,v_l=0,v_h=255,s_l=0.01,s_h=0.03,r_1=1,r_2=1.5,pixel_level=True)
  )

# train_generator = datagen.flow_from_directory(
#       'base/Images/',
#       save_to_dir = 'base/fake/',
#       batch_size=1
#       )
# for i in range(5):
#  train_generator.next()

# ！
# df_train = pd.read_csv('base/Annotations/label.csv', header=None)
# df_train.columns = ['image_id', 'class', 'label']
# classes = ['collar_design_labels', 'neckline_design_labels', 'skirt_length_labels', 
#   'sleeve_length_labels', 'neck_design_labels', 'coat_length_labels', 'lapel_design_labels', 
#   'pant_length_labels']
# ！

# classes = ['collar_design_labels']

# ！
# for i in range(len(classes)):
#  gc.enable()

# #  單個(gè)分類
#  cur_class = classes[i]
#  df_load = df_train[(df_train['class'] == cur_class)].copy()
#  df_load.reset_index(inplace=True)
#  del df_load['index']

# #  print(cur_class)

# #  加載數(shù)據(jù)和label
#  n = len(df_load)
# #  n_class = len(df_load['label'][0])
# #  width = 256

# #  X = np.zeros((n,width, width, 3), dtype=np.uint8)
# #  y = np.zeros((n, n_class), dtype=np.uint8)

#  print(f'starting load trainset {cur_class} {n}')
#  sys.stdout.flush()
#  for i in tqdm(range(n)):
# #   tmp_label = df_load['label'][i]
#   img = load_img('base/{0}'.format(df_load['image_id'][i]))
#   x = img_to_array(img)
#   x = x.reshape((1,) + x.shape)
#   m=0
#   for batch in datagen.flow(x,batch_size=1):
# #    plt.imshow(array_to_img(batch[0]))
# #    print(batch)
#    array_to_img(batch[0]).save(f'base/fake/{format(df_load["image_id"][i])}-{m}.jpg')
#    m+=1
#    if m>3:
#     break
#  gc.collect()
# ！  

img = load_img('base/Images/collar_design_labels/2f639f11de22076ead5fe1258eae024d.jpg')
plt.figure()
plt.imshow(img)
x = img_to_array(img)

x = x.reshape((1,) + x.shape)

i = 0
for batch in datagen.flow(x,batch_size=5):
 plt.figure()
 plt.imshow(array_to_img(batch[0]))
#  print(len(batch))
 i += 1
 if i >0:
  break

#多輸入，設(shè)置隨機(jī)種子
# Define the image transformations here
gen = ImageDataGenerator(horizontal_flip = True,
       vertical_flip = True,
       width_shift_range = 0.1,
       height_shift_range = 0.1,
       zoom_range = 0.1,
       rotation_range = 40)

# Here is the function that merges our two generators
# We use the exact same generator with the same random seed for both the y and angle arrays
def gen_flow_for_two_inputs(X1, X2, y):
 genX1 = gen.flow(X1,y, batch_size=batch_size,seed=666)
 genX2 = gen.flow(X1,X2, batch_size=batch_size,seed=666)
 while True:
   X1i = genX1.next()
   X2i = genX2.next()
   #Assert arrays are equal - this was for peace of mind, but slows down training
   #np.testing.assert_array_equal(X1i[0],X2i[0])
   yield [X1i[0], X2i[1]], X1i[1]

#手動(dòng)構(gòu)造，直接輸出多l(xiāng)abel
generator = ImageDataGenerator(rotation_range=5.,
        width_shift_range=0.1, 
        height_shift_range=0.1, 
        horizontal_flip=True, 
        vertical_flip=True)

def generate_data_generator(generator, X, Y1, Y2):
 genX = generator.flow(X, seed=7)
 genY1 = generator.flow(Y1, seed=7)
 while True:
   Xi = genX.next()
   Yi1 = genY1.next()
   Yi2 = function(Y2)
   yield Xi, [Yi1, Yi2]
model.fit_generator(generate_data_generator(generator, X, Y1, Y2),
    epochs=epochs)

def batch_generator(generator,X,Y):
 Xgen = generator.flow(X)
 while True:
  yield Xgen.next(),Y
h = model.fit_generator(batch_generator(datagen, X_all, y_all), 
       steps_per_epoch=len(X_all)//32+1,
       epochs=80,workers=3,
       callbacks=[EarlyStopping(patience=3), checkpointer,ReduceLROnPlateau(monitor='val_loss',factor=0.5,patience=1)], 
       validation_data=(X_val,y_val))

from PIL import Image
import numpy as np
from PIL import Image
from keras.preprocessing import image
import matplotlib.pyplot as plt
import os
import cv2
# from scipy.misc import toimage
import matplotlib
# 生成圖片地址和對(duì)應(yīng)標(biāo)簽
file_dir = '../train/'
image_list = []
label_list = []
cate = [file_dir + x for x in os.listdir(file_dir) if os.path.isdir(file_dir + x)]
for name in cate:
 temp = name.split('/')
 path = '../train_new/' + temp[-1]
 isExists = os.path.exists(path)
 if not isExists:
  os.makedirs(path) # 目錄不存在則創(chuàng)建
 class_path = name + "/"

 for file in os.listdir(class_path):
  print(file)
  img_obj = Image.open(class_path + file) # 讀取圖片
  img_array = np.array(img_obj)
  resized = cv2.resize(img_array, (256, 256)) # 裁剪
  resized = resized.astype('float32')
  resized /= 255.
  # plt.imshow(resized)
  # plt.show()
  save_path = path + '/' + file
  matplotlib.image.imsave(save_path, resized) # 保存

from PIL import Image
import numpy as np
from PIL import Image
from keras.preprocessing import image
import os
import cv2
# 生成圖片地址和對(duì)應(yīng)標(biāo)簽
file_dir = '../train/'

label_list = []
cate = [file_dir + x for x in os.listdir(file_dir) if os.path.isdir(file_dir + x)]
for name in cate:
 image_list = []
 class_path = name + "/"
 for file in os.listdir(class_path):
  image_list.append(class_path + file)
 batch_size = 64
 if len(image_list) < 10000:
  num = int(10000 / len(image_list))
 else:
  num = 0
 # 設(shè)置生成器參數(shù)
 datagen = image.ImageDataGenerator(fill_mode='wrap', # 填充模式
          rotation_range=40, # 指定旋轉(zhuǎn)角度范圍
          width_shift_range=0.2, # 水平位置平移
          height_shift_range=0.2, # 上下位置平移
          horizontal_flip=True, # 隨機(jī)對(duì)圖片執(zhí)行水平翻轉(zhuǎn)操作
          vertical_flip=True, # 對(duì)圖片執(zhí)行上下翻轉(zhuǎn)操作
          shear_range=0.2,
          rescale=1./255, # 縮放
          data_format='channels_last')
 if num > 0:
  temp = name.split('/')
  path = '../train_datage/' + temp[-1]
  isExists = os.path.exists(path)
  if not isExists:
   os.makedirs(path)

  for image_path in image_list:
   i = 1
   img_obj = Image.open(image_path) # 讀取圖片
   img_array = np.array(img_obj)
   x = img_array.reshape((1,) + img_array.shape)  #要求為4維
   name_image = image_path.split('/')
   print(name_image)
   for batch in datagen.flow(x,
        batch_size=1,
        save_to_dir=path,
        save_prefix=name_image[-1][:-4] + '_',
        save_format='jpg'):
    i += 1
    if i > num:
     break