MyGardenBird

Dataset creation and validation pipeline for Malaysia bird audio classification.

Dataset: The MyGardenBird dataset (6000 annotated 3-second segments, 10 species) is available on Zenodo:

Overview

End-to-end pipeline from Xeno-Canto downloads to optimized train/val/test splits:

1. Download recordings from Xeno-Canto (FLAC format)
2. Annotate bird vocalizations with interactive GUI
3. Extract 3-second segments
4. Quality control and filtering
5. Split with MIP optimization (prevents data leakage)

Pipeline Stages

Stage	Script	Description
1	Stage1_xc_fetch_metadata.py	Fetch recording metadata
2	Stage2_xc_dload_all_from_species_list.py	Download recordings
3	Stage3_xc_dload_delta_by_id.py	Download specific IDs
4	Stage4_audit_downloads.py	Audit downloaded FLACs against metadata
5	Stage5_find_segments_interactive.py	Interactive annotation GUI
6	Stage6_extract_annotated_segments.py	Extract WAV segments
7	Stage7_clip_qc_manifest.py	QC and generate dataset manifest CSV
8a	Stage8a_splitter_mip.py	MIP-based splitting (recommended)
8b	Stage8b_splitter_genetic_algorithm.py	GA-based splitting
8c	Stage8c_splitter_simulated_annealing.py	SA-based splitting
9	Stage9_train_mygardenbird_multifeature.py	Train 3 CNN models

MIP Splitter (Stage 8a)

Generates CSV-based splits with configurable ratios:

python Stage8a_splitter_mip.py /path/to/dataset \
    --train_ratio 0.80 --val_ratio 0.10 --test_ratio 0.10 \
    --output /path/to/splits.csv

Output format:

# split_ratio=80:10:10 seed=42 objective=0 solver=mip_cbc
filename,split
xc1002657_2860.wav,test
xc1003831_2642.wav,train
...

Key features:

• Source-based separation (same recording never in multiple splits)
• Perfect class balance (objective=0 means exact ratios achieved)
• Reproducible via seed parameter
• CSV output for use with any training framework

Splitter Performance Comparison

Benchmark on 6000-sample dataset (10 classes, 1074 sources) with 3 split ratios:

Algorithm	Avg Time	75:10:15	80:10:10	70:15:15	Solution Quality
MIP	1.1s	1.17s	1.24s	1.03s	Optimal (objective=0)
GA	7.5s	3.72s	3.04s	15.59s	Optimal (objective=0)
SA	~19 min	19.2 min	19.6 min	17.6 min	Optimal (objective=0)

Recommendation: Use MIP (Stage8a) for fastest results with guaranteed optimality.

Pre-generated Splits

Ready-to-use splits for 6000-sample dataset (seed=42, all objective=0):

File	Algorithm	Train	Val	Test
seabird_splits_mip_75_10_15.csv	MIP	75%	10%	15%
seabird_splits_mip_80_10_10.csv	MIP	80%	10%	10%
seabird_splits_mip_70_15_15.csv	MIP	70%	15%	15%
seabird_splits_ga_75_10_15.csv	Genetic Algorithm	75%	10%	15%
seabird_splits_ga_80_10_10.csv	Genetic Algorithm	80%	10%	10%
seabird_splits_ga_70_15_15.csv	Genetic Algorithm	70%	15%	15%
seabird_splits_sa_75_10_15.csv	Simulated Annealing	75%	10%	15%
seabird_splits_sa_80_10_10.csv	Simulated Annealing	80%	10%	10%
seabird_splits_sa_70_15_15.csv	Simulated Annealing	70%	15%	15%

Training Results

Benchmark results using Stage9_train_seabird_multifeature.py with 4 CNN architectures, 3 feature types, and 3 random seeds (42, 100, 786). All models use ImageNet pretrained weights with 75:10:15 train/val/test split.

Summary (Test Accuracy %)

Model	Mel	STFT	MFCC	Best
EfficientNetB0	93.4 ± 2.6	91.0 ± 1.5	89.4 ± 1.3	93.4%
ResNet50	88.8 ± 2.9	91.0 ± 1.1	86.0 ± 1.6	91.0%
MobileNetV3S	90.0 ± 0.9	90.1 ± 0.8	83.0 ± 1.0	90.1%
VGG16	88.2 ± 0.8	86.7 ± 1.8	81.9 ± 3.4	88.2%

Feature MFLOP Accounting

Input: 16kHz × 3s audio (48,000 samples) → 224×224 spectrogram

Parameters: N_FFT=2048, hop=214, 224 frames, 128 mel bins

Operation	STFT	Mel	MFCC
Framing + Windowing (224 × 2048)	0.5	0.5	0.5
FFT (224 frames × 5N log N)	25.2	25.2	25.2
Magnitude (224 × 1025 × 3 ops)	0.7	0.7	0.7
Mel Filterbank (sparse, ~20 weights/bin)	-	5.1	5.1
Log Compression (28K × 10 ops)	-	0.3	0.3
DCT-II (128 × 80 × 224 × 2)	-	-	4.6
Total MFLOPs	26	32	36

Total MFLOPs (CNN + Feature)

Model	Best Feature	CNN MFLOPs	Feature MFLOPs	Total MFLOPs	Accuracy (%)
EfficientNetB0	Mel	390	32	422	93.4
MobileNetV3S*	Mel	60	32	92	90.1
ResNet50	STFT	4100	26	4126	91.0
VGG16	Mel	15500	32	15532	88.2

Key Findings

• Best model: EfficientNetB0 + Mel spectrogram (93.4% accuracy)
• Best feature: Mel spectrogram consistently outperforms STFT and MFCC
• Most efficient: MobileNetV3S + Mel (90.1% accuracy at 92 MFLOPs = 4.6× less compute than EfficientNetB0)
• Most stable: VGG16 + Mel (lowest variance across seeds)
• MFCC: Unsuitable for bird sounds (although it works well with speech processing)

CNN Ranking

• EfficientNetB0 + Mel achieves highest accuracy (93.4%) at moderate compute.
• MobileNetV3S + Mel is optimal for edge deployment: 90.1% accuracy at just 92 MFLOPs.
• ResNet50 is Pareto-dominated: much higher compute for marginal accuracy gain.
• VGG16 is clearly inefficient.

Detailed Results

Model	Feature	Seed 42	Seed 100	Seed 786	Mean	Std
EfficientNetB0	Mel	95.89	93.56	90.67	93.37	2.62
EfficientNetB0	STFT	92.33	91.33	89.33	91.00	1.53
EfficientNetB0	MFCC	90.33	90.00	87.89	89.41	1.33
ResNet50	Mel	85.89	88.89	91.67	88.81	2.89
ResNet50	STFT	91.78	89.67	91.44	90.96	1.13
ResNet50	MFCC	84.33	87.56	86.22	86.04	1.62
VGG16	Mel	87.89	87.67	89.11	88.22	0.78
VGG16	STFT	87.89	87.56	84.56	86.67	1.84
VGG16	MFCC	79.56	80.33	85.78	81.89	3.39
MobileNetV3S	Mel	90.33	90.67	88.89	89.96	0.95
MobileNetV3S	STFT	90.22	90.89	89.33	90.15	0.78
MobileNetV3S	MFCC	82.56	82.22	84.11	82.96	1.01

MobileNetV3S Training Improvements

MobileNetV3S requires different hyperparameter settings. The original MobileNetV3S results (63.5%) showed significant underfitting. An improved training strategy achieved 90.1% accuracy on average:

Change	Other CNNs	MobileNetV3S
Warmup epochs	5	10
Fine-tuning scope	Top 20% layers	All layers
Fine-tune learning rate	5e-5	1e-4
Weight decay	1e-4	1e-5
Dropout (classifier)	0.5/0.4	0.3/0.2
Hidden units	256	512
Early stopping patience	7	15

This demonstrates that MobileNetV3S is highly suitable for bird audio classification when properly trained, achieving near-EfficientNetB0 accuracy at 4.6× lower compute.

Training Command

python Stage9_train_seabird_multifeature.py \
    --splits_csv ./seabird_splits_mip_75_10_15.csv \
    --model efficientnetb0 \
    --feature mel \
    --use_pretrained \
    --seed 42

For audio-focused CNN training, see also mun3im/mynanet.

Setup

1. Configure paths

Edit the top of config.py to match your environment:

PROJECT_ROOT = "/Volumes/Evo"   # change to your mount point or local path
DATASET_NAME = "SEABIRD"        # top-level folder created under PROJECT_ROOT

All stages derive their input/output paths from these two constants.

2. Place target_species.csv in the dataset directory

Before running any stage, copy target_species.csv into your dataset root:

<PROJECT_ROOT>/<DATASET_NAME>/target_species.csv

For example, with the defaults above:

/Volumes/Evo/SEABIRD/target_species.csv

The CSV controls which species are active in the pipeline. Set the active column to yes for the species you want to include and no to exclude them. The file must be in the dataset directory, not alongside the scripts.

Installation

pip install numpy scipy librosa soundfile requests tqdm matplotlib sounddevice pulp

License

MIT