pablo-reyes8/outgrid-vision-transformer
Hybrid vision backbone: VOLO-style Outlook local mixing + MaxViT grid attention + MBConv. Two variants (front-only vs per-block Outlook), FP16 training CLI with CutMix, and reproducible CIFAR-100 64×64 benchmarks.
What's novel
Hybrid vision backbone: VOLO-style Outlook local mixing + MaxViT grid attention + MBConv. Two variants (front-only vs per-block Outlook), FP16 training CLI with CutMix, and reproducible CIFAR-100 64×64 benchmarks.
Code Analysis
18 files read · 4 roundsA research project implementing OutGridViT, a hybrid CNN-Transformer vision architecture that fuses VOLO-style dynamic local kernels, MBConv, and grid-partitioned attention, with a full training stack, ablation variants, and attention analysis tools (MAD + entropy metrics).
Strengths
Genuinely non-trivial architecture with clean component decomposition (Outlooker, MBConv, GridAttention, MLP) and a well-structured YAML-first experiment API with strict config validation. The training pipeline is production-quality with AMP, gradient clipping, Mixup/CutMix, early stopping, and detailed instrumentation (grad norms, overflow tracking, memory stats).
Weaknesses
Code duplication of helper functions (make_activation, _softmax_local, sample_xy) across multiple files, and the test suite is thin — only 5 test files with basic forward-pass and shape checks, lacking edge-case coverage (e.g., non-divisible spatial dims, extreme configs, gradient flow verification).
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
80
Contributors
1
Files
223
Active weeks
7
Repository
Language
Python
Stars
1
Forks
0
License
MIT