MFF¶

Improving Pixel-based MIM by Reducing Wasted Modeling Capability

Abstract¶

There has been significant progress in Masked Image Modeling (MIM). Existing MIM methods can be broadly categorized into two groups based on the reconstruction target: pixel-based and tokenizer-based approaches. The former offers a simpler pipeline and lower computational cost, but it is known to be biased toward high-frequency details. In this paper, we provide a set of empirical studies to confirm this limitation of pixel-based MIM and propose a new method that explicitly utilizes low-level features from shallow layers to aid pixel reconstruction. By incorporating this design into our base method, MAE, we reduce the wasted modeling capability of pixel-based MIM, improving its convergence and achieving non-trivial improvements across various downstream tasks. To the best of our knowledge, we are the first to systematically investigate multi-level feature fusion for isotropic architectures like the standard Vision Transformer (ViT). Notably, when applied to a smaller model (e.g., ViT-S), our method yields significant performance gains, such as 1.2% on fine-tuning, 2.8% on linear probing, and 2.6% on semantic segmentation.

Train/Test Command

Prepare your dataset according to the docs.

Train:

python tools/train.py configs/mff/mff_vit-base-p16_8xb512-amp-coslr-300e_in1k.py

Test:

python tools/test.py configs/mff/benchmarks/vit-base-p16_8xb128-coslr-100e_in1k.py None

Models and results¶

Pretrained models¶

Model	Params (M)	Flops (G)	Config	Download
`mff_vit-base-p16_8xb512-amp-coslr-300e_in1k`	-	-	config	model \| log
`mff_vit-base-p16_8xb512-amp-coslr-800e_in1k`	-	-	config	model \| log

Image Classification on ImageNet-1k¶

Model	Pretrain	Params (M)	Flops (G)	Top-1 (%)	Config	Download
`vit-base-p16_mff-300e-pre_8xb128-coslr-100e_in1k`	MFF 300-Epochs	86.57	17.58	83.00	config	model / log
`vit-base-p16_mff-800e-pre_8xb128-coslr-100e_in1k`	MFF 800-Epochs	86.57	17.58	83.70	config	model / log
`vit-base-p16_mff-300e-pre_8xb2048-linear-coslr-90e_in1k`	MFF 300-Epochs	304.33	61.60	64.20	config	log
`vit-base-p16_mff-800e-pre_8xb2048-linear-coslr-90e_in1k`	MFF 800-Epochs	304.33	61.60	68.30	config	model / log

Citation¶

@article{MFF,
  title={Improving Pixel-based MIM by Reducing Wasted Modeling Capability},
  author={Yuan Liu, Songyang Zhang, Jiacheng Chen, Zhaohui Yu, Kai Chen, Dahua Lin},
  journal={arXiv},
  year={2023}
}