You can deploy this model and start earning money today!

architechs/
birdclef2026-convnextv2-tiny

Name: birdclef2026-convnextv2-tiny
Availability: InStock

Views

No views yet

BirdCLEF 2026 — ConvNeXtV2-Tiny with ASL + GEM-Attention Head

Backbone: ConvNeXt-V2-Tiny (ImageNet-22k pretrained, 28.6M params)
Head: GEM pooling (frequency) → Attention-weighted temporal pooling → per-class logits
Loss: Asymmetric Loss (ASL) with γ⁻=4, γ⁺=0, clip=0.05
Input: 128-bin mel spectrogram, 32kHz, 5-second clips, n_fft=1024, hop=320

Freq-MixStyle (DCASE 2024 winner): domain-invariant frequency normalization
Waveform Mixup (p=0.7, α=0.4): SNR-controlled mixing with union-style labels
Colored Noise augmentation: random spectral slope noise injection
Balanced Sampling: inverse-frequency weighting with smoothing + 2× soundscape weight
Secondary Labels: soft targets at 0.5 weight for co-occurring species
Rating Filter: quality filter ≥2.0, but preserves all rare species (<10 recordings)

4-fold MultiLabelStratifiedGroupKFold, grouped by author/recordist. Zero group leakage between train and validation.

Paper	Key Contribution
Bird-MAE (2504.12880)	Augmentation pipeline, ASL for bioacoustics
PSLA (2102.01243)	Multi-head attention pooling, balanced sampling
EfficientAT (2310.15648)	Efficient audio CNN architectures
ASL (2009.14119)	Asymmetric loss for multi-label with extreme neg:pos ratio
DCASE 2024 Winner (2407.12997)	Freq-MixStyle for domain generalization
BirdSet (2403.10380)	Bioacoustic evaluation benchmark

bash

export HF_DATASET_ID="your-org/birdclef2026-mirror"
python train_birdclef.py --fold 0 --backbone convnextv2_tiny.fcmae_ft_in22k_in1k

bash

export KAGGLE_USERNAME="your_username"
export KAGGLE_KEY="your_key"
python train_birdclef.py --fold 0

Backbone	Params	CPU Speed	Notes
`convnextv2_tiny.fcmae_ft_in22k_in1k`	28.6M	~60ms/clip	Best quality, IN22k pretrain
`efficientnet_b2.ra_in1k`	9.1M	~30ms/clip	PSLA baseline, fast
`eca_nfnet_l0.ra2_in1k`	24.1M	~50ms/clip	No BN, diverse from ConvNeXt
`regnety_016.tv2_in1k`	11.2M	~35ms/clip	Lightweight, diverse
`hgnetv2_b0.ssld_stage1_in22k_in1k`	6.0M	~20ms/clip	Fastest, your current baseline

Competition data subject to Kaggle BirdCLEF 2026 rules. Code: MIT