Analyzing the Training Dynamics of Image Restoration Transformers: A Revisit to Layer Normalization
This work addresses a critical training issue in image restoration Transformers, offering a simple drop-in replacement that stabilizes dynamics and boosts performance, though it is incremental as it modifies an existing component rather than introducing a new paradigm.
The paper tackles the problem of feature magnitude divergence and channel-wise entropy collapse in image restoration Transformers caused by conventional LayerNorm, and introduces i-LN, a tailored normalization method that improves stability and performance across various IR tasks, with experimental verification showing enhanced results.
This work investigates the internal training dynamics of image restoration~(IR) Transformers and uncovers a critical yet overlooked issue: conventional LayerNorm leads feature magnitude divergence, up to a million scale, and collapses channel-wise entropy. We analyze this phenomenon from the perspective of networks attempting to bypass constraints imposed by conventional LayerNorm due to conflicts against requirements in IR tasks. Accordingly, we address two misalignments between LayerNorm and IR tasks, and later show that addressing these mismatches leads to both stabilized training dynamics and improved IR performance. Specifically, conventional LayerNorm works in a per-token manner, disrupting spatial correlations between tokens, essential in IR tasks. Also, it employs an input-independent normalization that restricts the flexibility of feature scales, required to preserve input-specific statistics. Together, these mismatches significantly hinder IR Transformer's ability to accurately preserve low-level features throughout the network. To this end, we introduce Image Restoration Transformer Tailored Layer Normalization~(i-LN), a surprisingly simple drop-in replacement for conventional LayerNorm. We propose to normalize features in a holistic manner across the entire spatio-channel dimension, preserving spatial relationships among individual tokens. Additionally, we introduce an input-adaptive rescaling strategy that maintains the feature range flexibility required by individual inputs. Together, these modifications effectively contribute to preserving low-level feature statistics of inputs throughout IR Transformers. Experimental results verify that this combined strategy enhances both the stability and performance of IR Transformers across various IR tasks.