Created
June 14, 2019 16:09
-
-
Save crearo/5c702bf40d8df1f888108c22e0291ed1 to your computer and use it in GitHub Desktop.
A bug in Keras where it resizes first and then applies the preprocessing function
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
''' | |
Keras's ImageDataGenerator is buggy. It first resizes based on `target_size`, | |
and then applies the preprocessing function you specified. This sucks. | |
Who wants that to happen anyway. | |
''' | |
def crop_image(img): | |
print(img.shape) | |
return img | |
def load_data(): | |
train_datagen = ImageDataGenerator(rotation_range=90, | |
rescale=1. / 255, | |
preprocessing_function=crop_image) | |
test_datagen = ImageDataGenerator(rescale=1. / 255, | |
preprocessing_function=crop_image) | |
train_gen = train_datagen.flow_from_directory( | |
'%s/train' % base, | |
target_size=(256, 256), | |
color_mode='grayscale', | |
batch_size=batch_size, | |
class_mode='categorical') | |
test_gen = test_datagen.flow_from_directory( | |
'%s/test' % base, | |
target_size=(256, 256), | |
color_mode='grayscale', | |
batch_size=batch_size, | |
class_mode='categorical') | |
return train_gen, test_gen | |
def plot_data_gen(train_gen, test_gen): | |
for X, y in train_gen: | |
plt.figure(figsize=(16, 16)) | |
for i in range(25): | |
plt.subplot(5, 5, i + 1) | |
plt.axis('off') | |
plt.title('Label: %d' % np.argmax(y[i])) | |
img = np.uint8(255 * X[i, :, :, 0]) | |
plt.imshow(img, cmap='gray') | |
break | |
plt.show() | |
if __name__ == '__main__': | |
model = load_model() | |
train_gen, test_gen = load_data() | |
plot_data_gen(train_gen, test_gen) |
I'm not talking about the divide by 255.
I'd like Keras to apply the preprocessing_function
before it transforms the image into target_size
.
Right now it does it the other way around.
Is this bug corrected yet? Reading the current documentation of preprocessing_function seems like it is not. (https://www.tensorflow.org/api_docs/python/tf/keras/preprocessing/image/ImageDataGenerator)
Is there a workaround to this bug? I need to apply the preprocessing function first to make my images square after filling them with empty spaces on the edges and then resize it.
Thank you,
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Isn't it the correct way ? resize first and then divide by 255.