This repository is Keras's implementation of RICAP(Random Image Cropping And Patching), Data Augmentation algorithm which achieved SoTA with CIFAR-10 at error rate 2.19%. We applied to animeface-character-dataset.
It is easy to inherit from ImageDataGenerator of Keras like this:
from keras.preprocessing.image import ImageDataGenerator
from ricap import ricap
# RICAP Generator
class RICAPGenerator(ImageDataGenerator):
def __init__(self, ricap_beta=0.3, *args, **kwargs):
super().__init__(*args, **kwargs)
self.ricap_beta = ricap_beta
def flow_from_directory(self, *args, **kwargs):
for batch_X, batch_y in super().flow_from_directory(*args, **kwargs):
ricap_X, ricap_y = ricap(batch_X, batch_y, self.ricap_beta)
yield ricap_X, ricap_y
Use same random numbers with beta distribution in batch vs. Use different random numbers for each sample
There are 2 versions here.
- Use same random numbers in batch (default, same as the paper)
- Use different random numbers for each sample (made it on our own)
If you want to try differnt random numbers version, please specify use_same_random_value_on_batch=False
Conditions | Test Accuracy |
---|---|
baseline | 91.5039% |
batchwise RICAP (original) beta=0.1 | 91.7236% |
batchwise RICAP (original) beta=0.3 | 92.3340% |
batchwise RICAP (original) beta=0.5 | 91.6748% |
samplewise RICAP (changed) beta=0.1 | 92.0898% |
samplewise RICAP (changed) beta=0.3 | 91.9922% |
samplewise RICAP (changed) beta=0.5 | 91.6504% |
Accuracy is better for batchwise(the original), but samplewise is smoother in error.
Also, we tried a method to schedule and change the beta value (a method close to the pseudo label of semi supervised learning).
example:
- beta=0.01 if epoch<=20
- increase beta value linearly if 20<epoch<=90
- beta=1 if epoch>90 (i.e. uniform distribution)
On beta scheduling, it seems that samplewise RICAP is better.
Ryo Takahashi, Takashi Matsubara, Kuniaki Uehara. Data Augmentation using Random Image Cropping and Patching for Deep CNNs. ACML2018. 2018