Towards Reproducible Learning-based Compression

TL;DR

The irreproducibility issue is analyzed where deep learning is employed in compression systems while the encoding and decoding may be run on devices from different manufacturers and a safeguarding mechanism is proposed to tackle the challenges.

Abstract

A deep learning system typically suffers from a lack of reproducibility that is partially rooted in hardware or software implementation details. The irreproducibility leads to skepticism in deep learning technologies and it can hinder them from being deployed in many applications. In this work, the irreproducibility issue is analyzed where deep learning is employed in compression systems while the encoding and decoding may be run on devices from different manufacturers. The decoding process can even crash due to a single bit difference, e.g., in a learning-based entropy coder. For a given deep learning-based module with limited resources for protection, we first suggest that reproducibility can only be assured when the mismatches are bounded. Then a safeguarding mechanism is proposed to tackle the challenges. The proposed method may be applied for different levels of protection either at the reconstruction level or at a selected decoding level. Furthermore, the overhead introduced for the protection can be scaled down accordingly when the error bound is being suppressed. Experiments demonstrate the effectiveness of the proposed approach for learning-based compression systems, e.g., in image compression and point cloud compression.

Figures (8)

• Figure 1: (a) We ensure reproducibility for learning-based compression systems to achieve interoperability across platforms. (b) Without protection, a lossless point cloud decoder on an H100 GPU fails to decode a point cloud encoded with an A100 GPU. (c) With our proposal, the decoder decodes correctly. In (d)(e), a similar example of lossy image coding is provided. In both cases, the bitstream overhead is around $1\%$.
• Figure 2: Reproducibility in a hyperprior-based image codec. Point $A$ is to check reconstruction reproducibility. Point $B$ is to check decoding reproducibility.
• Figure 3: Quantization of an error-bounded variable.
• Figure 4: Encoder and decoder flowcharts of our proposal. The steps in black---utilize the direction flag $f_d$ to reproduce $Q^{-1}(Q(v))$ on the decoder. Replacing the circled parts by the associated steps in blue color---the simplified proposal without flag $f_d$.
• Figure 5: Error correction when a variable falls in the risky zone.