Progressive Rising of Patch Dimension: Useful resource-Environment friendly Curriculum Studying for Dense Prediction Duties
Authors: Stefan M. Fischer, Lina Felsner, Richard Osuala, Johannes Kiechle, Daniel M. Lang, Jan C. Peeken, Julia A. Schnabel
Summary: On this work, we introduce Progressive Rising of Patch Dimension, a resource-efficient implicit curriculum studying strategy for dense prediction duties. Our curriculum strategy is outlined by rising the patch measurement throughout mannequin coaching, which regularly will increase the duty’s problem. We built-in our curriculum into the nnU-Internet framework and evaluated the methodology on all 10 duties of the Medical Segmentation Decathlon. With our strategy, we’re in a position to considerably scale back runtime, computational prices, and CO2 emissions of community coaching in comparison with classical fixed patch measurement coaching. In our experiments, the curriculum strategy resulted in improved convergence. We’re in a position to outperform customary nnU-Internet coaching, which is skilled with fixed patch measurement, by way of Cube Rating on 7 out of 10 MSD duties whereas solely spending roughly 50% of the unique coaching runtime. To one of the best of our data, our Progressive Rising of Patch Dimension is the primary profitable employment of a sample-length curriculum within the type of patch measurement within the discipline of pc imaginative and prescient. Our code is publicly accessible at https://github.com/compai-lab/2024-miccai-fischer
