Project layout — this bundle contains five stage directories from the AI-Designer pipeline: research/ (literature survey), architect/ (blueprint + ModelConfig), coder/ (PyTorch implementation), validator/ (tests + benchmarks), and documenter/ (this README plus docs/ and CHANGELOG.md). An optional paper/ directory holds the NeurIPS-format writeup when the paper-generation step was triggered.

The original research request that produced this bundle is preserved verbatim in prompt.md — if any URLs in the prompt were fetched server-side for additional context, their cleaned contents are appended there too.

---

SpectroConvNeXt

A ConvNeXt V2-derived CNN family with spectrogram-adapted rectangular kernels for environmental sound classification on ESC-50, spanning five model variants from 5M to 30M parameters.

SpectroConvNeXt addresses the gap in systematic modern CNN scaling on ESC-50, where prior methods are all <5M parameters and use ad-hoc architectures. It applies four key adaptations to ConvNeXt V2 for audio spectrograms: a frequency-preserving two-layer stem, rectangular depthwise kernels in Stage 1 (matching the 1.49:1 spectrogram aspect ratio), Global Response Normalization in every block, and DropPath regularisation scaled with model size.

All accuracy-related claims are TODO: unverified — see docs/BENCHMARKS.md. A training pipeline is required to evaluate the central hypotheses (≥93% at 10M, ≥94% at 30M).

Highlights

• Five cleanly scaled variants (5M–30M) — Atto-S (~4.8M), Femto-S (~9.5M), Pico-S (~15.2M), Nano-S (~20.5M), Tiny-S (~29.8M); see ARCHITECTURE.md#7-variant-scaling-table
• Spectrogram-native adaptations — frequency-preserving two-layer stem, rectangular (7,5) depthwise kernel in Stage 1, asymmetric stride; see ARCHITECTURE.md#4-architecture-overview
• GRN in every block — Global Response Normalization (ConvNeXt V2) for inter-channel competition without extra parameters; see ARCHITECTURE.md#54-global-response-normalization-grn
• Parameter counts verified — all five variants within 15% of targets (grounded by validator); see BENCHMARKS.md#synthetic-benchmarks
• Fully convolutional, any input resolution — no position encoding; operates at any frequency/time resolution (validated at F=64,128,256 and T=43,86,172)

Quick start

pip install torch torchaudio
cd coder/
python smoke_test.py        # all 5 variants + ablations
pytest test_model.py -v     # 30+ tests: shape, gradient, numerics, audio pipeline
python run_ablations.py     # parameter-count check for all 6 ablations

Repository layout

coder/                      # Implementation
    model.py                # Full model: SpectroConvNeXt + config dataclasses
    backbone.py             # Backbone: stem + 4 stages + downsamplers
    layers.py               # Building blocks: GRN, StochasticDepth, SpectroConvNeXtBlock, stems
    head.py                 # Classification head: GAP → LayerNorm → Linear
    frontend.py             # AudioFrontend: waveform → log-mel spectrogram
    smoke_test.py           # Smoke test: variants, ablations, gradient flow
validator/                  # Tests and benchmarks
    test_model.py           # Unit tests (shape, gradient, CV, audio, numerics)
    profile_model.py        # torch.profiler throughput and memory profiling
    run_ablations.py        # Ablation runner (6 ablations, single-config-field changes)
    research_eval/          # Research-quality evaluation artifacts
        scorecard.json      # Numerical scores and gap analysis
        claim_grounding.md  # Trace each claim to source/verification
        experiment_coverage.md  # Coverage matrix of all required experiments
architect/                  # Architecture design
    spectroconvnext_architecture.md  # Full blueprint with pseudocode and ASCII diagram
    spectroconvnext_config.py        # Standalone ModelConfig dataclass
research/                   # Research synthesis
    spectroconvnext_research_synthesis.md  # Landscape, gaps, hypothesis, lifecycle contract

Documentation

• docs/ARCHITECTURE.md — design motivation, inductive biases, tensor shape evolution, design decisions
• docs/TRAINING.md — environment setup, recommended hyperparameters, troubleshooting
• docs/BENCHMARKS.md — test results, ablation catalogue, profiling, research-quality evaluation
• docs/API.md — module-level API reference with shape contracts

Citation

@misc{spectroconvnext-2026,
  title  = {SpectroConvNeXt: A Modern CNN Family for Audio Spectrogram Classification},
  author = {TODO},
  year   = {2026},
  note   = {Generated via ml-designer pipeline}
}