CoAtNet: Marrying Convolution and Attention for All Data Sizes

Authors: Zihang Dai, Hanxiao Liu, Quoc V Le, Mingxing Tan

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that our Co At Nets achieve state-of-the-art performance under different resource constraints across various datasets
Researcher Affiliation Industry Google Research, Brain Team {zihangd,hanxiaol,qvl,tanmingxing}@google.com
Pseudocode No The paper describes methods and architectures but does not include explicit pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement about releasing source code or a link to a code repository for the methodology described.
Open Datasets Yes we utilize three datasets of increasingly larger sizes, namely Image Net-1K (1.28M images), Image Net-21K (12.7M images) and JFT (300M images).
Dataset Splits No The paper uses Image Net-1K and JFT datasets and discusses training and evaluation, but it does not explicitly state the specific training/validation/test split percentages, sample counts, or refer to predefined splits with citations in a way that provides concrete split information for reproducibility.
Hardware Specification Yes On our accelerator of choice (TPU), such operation turns out to be extremely slow [34]" and "TPUv3-core-days denotes the pretraining time
Software Dependencies No The paper discusses various models and techniques but does not provide specific software versions (e.g., deep learning framework versions, Python versions, or library versions) used for implementation.
Experiment Setup Yes For all Conv and MBConv blocks, we always use the kernel size 3. For all Transformer blocks, we set the size of each attention head to 32, following [22]. The expansion rate for the inverted bottleneck is always 4 and the expansion (shrink) rate for the SE is always 0.25." and "we first pre-train our models on each of the three datasets at resolution 224 for 300, 90 and 14 epochs respectively. Then, we finetune the pre-trained models on Image Net-1K at the desired resolutions for 30 epochs