Enhancing Feature Diversity Boosts Channel-Adaptive Vision Transformers

Authors: Chau Pham, Bryan Plummer

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on both satellite and cell microscopy datasets, CHAMMI, JUMP-CP, and So2Sat, report Di Cha Vi T yields a 1.5 5.0% gain over the state-of-the-art.
Researcher Affiliation Academia Chau Pham Boston University Boston, MA chaupham@bu.edu Bryan A. Plummer Boston University Boston, MA bplum@bu.edu
Pseudocode Yes Algorithm 1: Diverse Channel Sampling (DCS)
Open Source Code Yes Our code is publicly available at https://github.com/chaudatascience/diverse_channel_vit.
Open Datasets Yes Experiments on both satellite and cell microscopy datasets, CHAMMI, JUMP-CP, and So2Sat, report Di Cha Vi T yields a 1.5 5.0% gain over the state-of-the-art.
Dataset Splits Yes JUMP-CP [12] comprises images and profiles of cells that were individually perturbed using chemical and genetic methods. Our experiments focus on the compound perturbation plate BR00116991, which contains 127K training images, 45K validation images, and 45K test images.
Hardware Specification Yes In this study, experiments were conducted on So2Sat and CHAMMI using a single NVIDIA RTX (48GB RAM) and three Intel(R) Xeon(R) Gold 6226R CPUs @ 2.90GHz. For experiments on JUMP-CP, two NVIDIA RTX A6000 GPUs and six Intel(R) Xeon(R) Gold 6226R CPUs @ 2.90GHz were utilized.
Software Dependencies No The paper mentions software components like DINOv2 and AdamW optimizer, but does not specify their version numbers.
Experiment Setup Yes We train each model for 60 epochs with a learning rate of 0.00004, and a batch size of 64. For JUMP-CP and So2Sat, the learning rate is warmed up for the initial 10 epochs, peaking at 0.0005 after which it will gradually decay to 10 6 following a cosine scheduler. We also apply a weight decay of 0.04... We train each model for 100 epochs, with a batch size of 64 on JUMP-CP, and 128 on So2Sat.