Adaptive Depth Networks with Skippable Sub-Paths

Authors: Woochul Kang, HYUNGSEOP LEE

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

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 4, we empirically demonstrate that our adaptive depth networks outperform counterpart individual networks, both in CNNs and vision transformers, and achieve actual inference acceleration and energy-saving.
Researcher Affiliation Academia Woochul Kang Department of Embedded Systems Incheon National University Yeonsu-gu, Incheon, South Korea, 22012 wchkang@inu.ac.kr Hyungseop Lee Department of Embedded Systems Incheon National University Yeonsu-gu, Incheon, South Korea, 22012 hhss0927@inu.ac.kr
Pseudocode Yes Algorithm 1 Training algorithm for an adaptive depth network M.
Open Source Code Yes Source codes are available at https://github.com/wchkang/depth
Open Datasets Yes We evaluate our method on ILSVRC2012 dataset [49] that has 1000 classes. The dataset consists of 1.28M training and 50K validation images.
Dataset Splits Yes The dataset consists of 1.28M training and 50K validation images.
Hardware Specification Yes Table 2: Training time (1 epoch), measured on Nvidia RTX 4090 (batch size: 128).
Software Dependencies No The paper mentions training recipes from DeiT [50] and PyTorch [51] but does not provide specific version numbers for software dependencies such as Python, PyTorch, or CUDA.
Experiment Setup Yes For CNN models, we follow most training settings in the original papers [34, 35], except that Res Net models are trained for 150 epochs. Vi T and Swin-T are trained for 300 epochs, following Dei T s training recipe [50, 51].