Diagonalwise Refactorization: An Efficient Training Method for Depthwise Convolutions
Abstract
Domain fit: AI-adjacent · Paper appears method- or tooling-adjacent to AI workflows with partial ecosystem coverage.
Depthwise convolutions provide significant performance benefits owing to the reduction in both parameters and mult-adds. However, training depthwise convolution layers with GPUs is slow in current deep learning frameworks because their implementations cannot fully utilize the GPU capacity. To address this problem, in this paper we present an efficient method (called diagonalwise refactorization) for accelerating the training of depthwise convolution layers. Our key idea is to rearrange the weight vectors of a depthwise convolution into a large diagonal weight matrix so as to convert the depthwise convolution into one single standard convolution, which is well supported by the cuDNN library that is highly-optimized for GPU computations. We have implemented our training method in five popular deep learning frameworks. Evaluation results show that our proposed method gains 15.4× training speedup on Darknet, 8.4× on Caffe, 5.4× on PyTorch, 3.5× on MXNet, and 1.4× on TensorFlow, compared to their original implementations of depthwise convolutions.
Results and benchmarks
Depthwise convolutions provide significant performance benefits owing to the reduction in both parameters and mult-adds.
Benchmark evidence is limited
Evidence graph: 2 refs, 1 links.
Utility signals: depth 65/100, grounding 58/100, status medium.
Implementation
No direct implementation yet
Maintained implementation evidence is not confirmed for this paper yet.
Use the implementation status and reproduction sections for the current action plan.
No verified maintained repo yet
There is no verified maintained implementation yet. Use this baseline plan to decide whether to prototype now or defer.
- No direct maintained implementation was found. Use the paper PDF and citation graph to design a baseline reproduction.
- Start from related paper: Another view on parallel speedup.
- Track assumptions and missing details in an experiment log before coding.
Time to first repro: a few days
Recommendation evidence is currently too limited for a maintained-repo choice. Use Implementation Status and Reproduction Path for a practical baseline plan.
- Estimate is based on paper-only reproduction flow
Reproduction readiness
No repo
No verified implementation available
- No maintained repository has been identified for this paper. Check adjacent implementations or HF artifacts below.
Hardware requirements
- Expect multi-day setup/compute for meaningful reproduction based on current guidance.
Validation caveat
Hugging Face artifacts
No trustworthy direct or curated related Hugging Face artifacts were found yet. Use targeted searches to quickly locate candidate models, datasets, and demos.
Models
Tip: start with models, then check datasets and spaces if you need evaluation data or demos.
Research context
29
Citations
64
References
Tasks
Computer science, Convolution (computer science), Speedup, Diagonal, Computation, Deep learning, Implementation, Key (lock)
Methods
Algorithm
Domains
Artificial intelligence, Computer Vision and Pattern Recognition
Related papers
- Another view on parallel speedupSearch on Paper2Code
1990 · Semantic similarity
- Shared virtual memory and generalized speedupSearch on Paper2Code
2002 · Semantic similarity
- Performance considerations of shared virtual memory machinesSearch on Paper2Code
1995 · Semantic similarity
- Toward a better parallel performance metricSearch on Paper2Code
1991 · Semantic similarity
- Speedup for Multi-Level Parallel ComputingSearch on Paper2Code
2012 · Semantic similarity
- Program Speedup in a Heterogeneous Computing NetworkSearch on Paper2Code
1994 · Semantic similarity
Open this paper in HFEPX to review benchmark signals, evaluation modes, and human-feedback protocol context.
Open in HFEPXJump to Paper2Code search queries derived from this paper's research context.