geirhos_conflict_stimuli

  • विवरण :

"इमेजनेट-प्रशिक्षित CNN बनावट के प्रति पक्षपाती हैं; आकार पूर्वाग्रह बढ़ाने से सटीकता और मजबूती में सुधार होता है।"

ध्यान दें कि, हालांकि डेटासेट स्रोत में आकार और बनावट से मेल खाने वाली छवियां होती हैं और हम उन्हें यहां शामिल करते हैं, मूल पेपर में अधिकांश मूल्यांकनों के लिए उन्हें अनदेखा कर दिया जाता है।

विभाजित करना उदाहरण
'test' 1,280
  • फ़ीचर संरचना :
FeaturesDict({
    'file_name': Text(shape=(), dtype=string),
    'image': Image(shape=(None, None, 3), dtype=uint8),
    'shape_imagenet_labels': Sequence(ClassLabel(shape=(), dtype=int64, num_classes=1000)),
    'shape_label': ClassLabel(shape=(), dtype=int64, num_classes=16),
    'texture_imagenet_labels': Sequence(ClassLabel(shape=(), dtype=int64, num_classes=1000)),
    'texture_label': ClassLabel(shape=(), dtype=int64, num_classes=16),
})
  • फ़ीचर दस्तावेज़ीकरण :
विशेषता कक्षा आकार डीटाइप विवरण
विशेषताएं डिक्ट
फ़ाइल का नाम मूलपाठ डोरी
छवि छवि (कोई नहीं, कोई नहीं, 3) uint8
आकार_इमेजनेट_लेबल अनुक्रम (कक्षा लेबल) (कोई भी नहीं,) int64
आकार_लेबल क्लासलेबल int64
बनावट_इमेजनेट_लेबल अनुक्रम (कक्षा लेबल) (कोई भी नहीं,) int64
बनावट_लेबल क्लासलेबल int64

VISUALIZATION

  • उद्धरण :
@inproceedings{
  geirhos2018imagenettrained,
  title={ImageNet-trained {CNN}s are biased towards texture; increasing shape
         bias improves accuracy and robustness.},
  author={Robert Geirhos and Patricia Rubisch and Claudio Michaelis and
          Matthias Bethge and Felix A. Wichmann and Wieland Brendel},
  booktitle={International Conference on Learning Representations},
  year={2019},
  url={https://openreview.net/forum?id=Bygh9j09KX},
}