[英]Cast ImageDataGenerator Data Output
I'm writing a network for Image Segmentation.我正在为图像分割编写一个网络。 I have my ImageDataGenerator for my masks (which are RGB images with only 0 and 255 as values, black and white) which is:我有我的 ImageDataGenerator 用于我的蒙版(它们是只有 0 和 255 作为值的 RGB 图像,黑色和白色),它是:
train_mask_data_gen = ImageDataGenerator(rotation_range=10,
width_shift_range=10,
height_shift_range=10,
zoom_range=0.3,
horizontal_flip=True,
vertical_flip=True,
fill_mode='nearest',#interpolation used for augmenting the image
cval=0,
rescale=1./255)
And flow_from_directory:和 flow_from_directory:
train_mask_gen = train_mask_data_gen.flow_from_directory(os.path.join(training_dir, 'masks'),
target_size=(img_h, img_w),
batch_size=bs,
class_mode=None, # Because we have no class subfolders in this case
shuffle=True,
interpolation='nearest',#interpolation used for resizing
#color_mode='grayscale',
seed=SEED)
The code works fine, the only problem is that, when i'm applying data augmentation to the masks, i won't have binary images anymore, but i get some values between 0 and 1 (normalized).代码工作正常,唯一的问题是,当我对掩码应用数据增强时,我将不再有二进制图像,但我得到一些 0 和 1 之间的值(标准化)。 For example, if i print my output matrix (the image) i get something like this:例如,如果我打印我的输出矩阵(图像),我会得到如下信息:
[[0. 0. 0. ]
[0. 0. 0. ]
[0. 0. 0. ]
...
[1. 1. 1. ]
[1. 1. 1. ]
[1. 1. 1. ]]
...
[[0. 0. 0. ]
[0.3457849 0.3457849 0.3457849 ]
[1. 1. 1. ]
...
[0. 0. 0. ]
[0. 0. 0. ]
[0. 0. 0. ]]
Which contains also those "extra" values due to augmentation.其中还包含由于增强而产生的那些“额外”值。 If i don't apply any augmentation i get binary images as i wanted.如果我不应用任何增强,我会得到我想要的二进制图像。
How can i embedd the casting to integer?我怎样才能将铸造嵌入到整数? (in order to get values which are only 0 or 1) I tried to use the field dtype=int
in the ImageDataGenerator
, but it doesn't do anything, i keep getting the same results. (为了获得只有 0 或 1 的值)我尝试在ImageDataGenerator
使用字段dtype=int
,但它没有做任何事情,我一直得到相同的结果。
setting the dtype argument to 'uint8' worked for me:将 dtype 参数设置为 'uint8' 对我有用:
Original:原来的:
datagen = ImageDataGenerator(dtype = 'float32')
val_set = datagen.flow_from_directory(data_dir, batch_size=1, target_size = (257,144))
Output:输出:
[[[ 52. 58. 61.]
[ 53. 53. 61.]
[ 54. 57. 66.]
...
[ 5. 12. 0.]
[ 19. 26. 12.]
[ 1. 15. 0.]]]
New:新的:
datagen = ImageDataGenerator(dtype = 'uint8')
val_set = datagen.flow_from_directory(data_dir, batch_size=1, target_size = (257,144))
output:输出:
[[[ 52 58 61]
[ 53 53 61]
[ 54 57 66]
...
[ 5 12 0]
[ 19 26 12]
[ 1 15 0]]]
The Keras docs do suggest that setting Dtype is the correct thing to do, so it may be a bug... One thing you could do is wrap the Keras generator yourself and cast it correctly: Keras 文档确实建议设置 Dtype 是正确的做法,所以它可能是一个错误......你可以做的一件事是自己包装 Keras 生成器并正确投射:
# quick stand in for a Keras image generator...
def img_gen():
for i in range(3):
yield np.random.rand(1, 2, 3) + 0.5
def int_gen(gen):
for i in gen:
yield i.astype(np.uint8)
for i in img_gen():
print(i)
for i in int_gen(img_gen()):
print(i)
output:输出:
...
[[[0.53385283 1.47129752 0.98338025]
[0.56875012 1.19955292 0.90370756]]]
[[[1.03524687 0.66555768 1.08211682]
[1.23256381 0.84470396 0.53269755]]]
[[[0.76095154 1.15223349 0.86353093]
[0.63276903 0.74591046 0.50097586]]]
[[[1 1 0]
[0 0 1]]]
[[[1 1 0]
[1 1 1]]]
[[[1 1 0]
[1 1 0]]]
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.