BirdCLEFplus_2025

Solution for https://www.kaggle.com/competitions/birdclef-2025

Installation

This repository uses uv package manager to install the required packages. For specific installation, please look docs. On Linux you can install and synchronize it by running:

curl -LsSf https://astral.sh/uv/install.sh | sh
uv sync
source .venv/bin/activate

Usage

Place your Kaggle API token in the ~/.kaggle directory. You can find instructions on how to do it here.

mv kaggle.json ~/.kaggle/kaggle.json
chmod 600 ~/.kaggle/kaggle.json

Place your WandB API token in the .env file so that it logins automatically:

WANDB_API_KEY=your_wandb_api_key

Dataset download and processing

kaggle competitions download -p input -c birdclef-2025
unzip input/birdclef-2025.zip -d input/birdclef-2025/
kaggle datasets download kdmitrie/bc25-separation-voice-from-data-by-silero-vad -p input/voice_data --unzip

If you wish to do manual data precomputation, you can run the following command:

uv run python -m src.audio_processing

Training

uv run python -m src.train -c configs/your_config.yaml

Journal

2025-05-18

Lookead at different librosa.feature.melspectrogram params:

• Changed N_FFT to 2048 from 1024
• Changed HOP_LENGTH to 512 from 256
• Changed FMIN from 20 to 50
• FMAX seems to be better at 15000, not 14000 These changes were made due to the seemingly clearer and full image. However, we have to investigate it with training and validation.

2025-05-19

• Tested mel spectrograms precomputation — indeed, it increased the training speed.

2025-05-21

• Conducted experiments: 001, 002, 003, 004.

2025-05-22

• Conducted experiments: 005, 006.

2025-05-25

• N_MELS has to be 256!!! Because it translates into the height of the meplspec image.
• Changed short audio processing from copying to constant padding

2025-05-30

• Tested FocalLossBCE.

Experiments

• 001-1.yaml — 001-8.yaml are related to the melspec settings. The best ones are:

N_FFT: 2048
HOP_LENGTH: 1024 (but 512 on Public LB)
N_MELS: 128
FMIN: 50 (but 20 on Public LB)
FMAX: 14000 (but 16000 on Public LB)
MINMAX_NORM: true

• 002-1.yaml — 002-7.yaml are related to the optimizer scheduler settings and batch size.
  • Larger BS (32 -> 128) can yield better results — need to investigate further
  • weight_decay: 1.0e-2 is better than weight_decay: 1.0e-5
  • OneCycleLR is really bad
  • ReduceLROnPlateau can also get good results
• 003-13.yaml, 003-17.yaml, 003-18.yaml are changing resolution of 001-3, 001-7, 001-8 experiments from 256x256 to 224x224
  • Based on both Local AUC and Public AUC, 256x256 is a better choice
• 004-13.yaml, 004-17.yaml, 004-18.yaml are changing number of epochs from 10 to 15 and min_lr: 1.0e-6 to min_lr: 1.0e-7.
  • Based on both Local AUC and Public AUC, it is not clear whether these changes actually improve generalization ability, but it's clear that they overfit much worse
• 005-1.yaml — 005-10.yaml are changing lr from 1.0e-3 to 1.0e-2 with the step of 0.1.
  • lr: 3.0e-3 is the best one
• 006-1.yaml — 006-4.yaml are changing in_channels from 1 to 3 and pretrained from True to False
  • in_channels: 3 with ImageNet normaliation qorks fine
• 011-1.yaml — 011-3.yaml are changing voice processing. 1 is not filtering out human voice, 2 is filtering out human voice and leave only the longest segment without voice, 3 is filtering out human voice and leaving concatenated segments without voice.
  • 1 is the best one based on Local AUC.
• 012-* are changing the base model from EfficientNet0 to 1, 2, 3 and 4 with 1 and 3 input channels. Other training parameters are the same.
  • Only EfficientNet3 with 3 channels is better than EfficientNet0 with 1 channel.
• 013-a* are changing the aug_prob and mixup_alpha parameters.
  • Turns out that different values of mixup_alpha do not change anything. Only 0 value can change, but I have never tested it.
  • Also, aug_prob of 0.5 is the best.
• 013-d* are changing the drop_rate and drop_path_rate parameters.
  • drop_rate: 0.5 and drop_path_rate: 0.2 are the best ones. The next is drop_rate: 0.2 and drop_path_rate: 0.35.
• 014-* are changing the precompute_data parameter and adds all 5 folds to the training.
  • precompute_data param does not change anything, but the speed (x30) of training.
  • All 5 folds have 5 different results, from 0.94544 to 0.95592.
• 015-* are changing the BCEWithLogitsLoss to FocalLossBCE with different params.
  • BCEWithLogitsLoss seems to be the best single.
  • FocalLossBCE's reduction sum is worse than mean.
  • FocalLossBCE's gamma 3 is worse than 2.
• 016-* are changing the HOP_LENGTH from 512 to 16, also change N_MELS from 256 to 512 and MINMAX_NORM from True to False.

Results

Experment name, fold	Local AUC	Public AUC	Details
001-1, 0	0.94536	0.747	-
001-2, 0	0.94777	-	-
001-3, 0	0.95217	0.751	-
001-4, 0	0.94892	-	-
001-5, 0	0.94621	-	-
001-6, 0	0.94896	-	-
001-7, 0	0.95190	0.779	-
001-8, 0	0.95055	-	-
002-1, 0	0.94536	-	-
002-2, 0	0.94487	-	-
002-3, 0	0.94789	-	-
002-4, 0	0.94842	-	-
002-5, 0	0.94152	-	-
002-6, 0	0.94143	-	-
002-7, 0	0.94771	-	-
003-13, 0	0.95079	-	-
003-17, 0	0.95095	0.774	-
003-18, 0	0.94993	-	-
004-13, 0	0.95005	-	-
004-17, 0	0.95103	0.765	-
004-18, 0	0.95146	-	-
005-1, 0	0.94534	-	-
005-2, 0	0.94883	-	-
005-3, 0	0.95130	-	-
005-4, 0	0.94950	-	-
005-5, 0	0.94731	-	-
005-6, 0	0.93399	-	-
005-7, 0	0.93703	-	-
005-8, 0	0.92751	-	-
005-9, 0	0.91582	-	-
005-10, 0	0.91446	-	-
006-1, 0	0.95190	-	-
006-2, 0	0.92699	-	-
006-3, 0	0.92410	-	-
006-4, 0	0.95110	-	-
011-1, 0	0.95403	0.732	-
011-2, 0	0.95004	0.762	-
011-3, 0	0.94818	0.763	-
012-10, 0	0.95403	-	-
012-11, 0	0.85695	-	-
012-12, 0	0.94916	-	-
012-13, 0	0.95021	-	-
012-14, 0	0.94993	-	-
012-30, 0	0.94919	-	-
012-31, 0	0.90809	-	-
012-32, 0	0.95438	0.746	-
012-33, 0	0.95227	-	-
012-34, 0	0.95086	0.742	-
013-a0, 0	0.95162	-	-
013-a5, 0	0.94826	-	-
013-a10, 0	0.95403	-	-
013-a15, 0	0.95085	-	-
013-a20, 0	0.95222	-	-
013-d0, 0	0.94865	-	-
013-d1, 0	0.94600	-	-
013-d2, 0	0.95319	-	-
013-d3, 0	0.95376	-	-
013-d4, 0	0.95186	-	-
013-d5, 0	0.94825	-	-
013-d6, 0	0.94824	-	-
013-d7, 0	0.95255	-	-
013-d8, 0	0.95196	-	-
013-d9, 0	0.95023	-	-
013-d10, 0	0.94682	-	-
013-d11, 0	0.95443	0.759	-
013-d12, 0	0.95529	0.764	-
013-d13, 0	0.95204	-	-
013-d14, 0	0.94606	-	-
013-d15, 0	0.94986	-	-
013-d16, 0	0.95272	-	-
013-d17, 0	0.95022	-	-
013-d18, 0	0.95159	-	-
013-d19, 0	0.94326	-	-
013-d20, 0	0.95247	-	-
013-d21, 0	0.95592	0.773	-
013-d22, 0	0.95044	-	-
013-d23, 0	0.95028	-	-
015-1, 0	0.95592	-	-
015-2, 0	0.94985	-	-
015-3, 0	0.9462	-	-
015-4, 0	error	-	-
015-5, 0	0.94688	-	-
015-6, 0	0.95063	-	-
015-7, 0	0.94945	-	-
015-8, 0	0.95122	-	-
015-9, 0		-	-

Hypotheses

• Remove human voice link1 link2
• Remove mel_spec_norm
• Test padding audio if it is less than 5s instead of copying it link
• Maybe use TTA link
• Test HOP_LENGTH up to 16
• Test FMIN up to 20
• Test FMAX up to 16000
• Test N_MELS up to 128
• Test
• Test model's drop_rate to something other than 0.2
• Test model's drop_path_rate to something other than 0.2
• Test a different model's classifier link

        self.classifier = nn.Sequential(
            nn.Linear(backbone_out, 512),
            nn.BatchNorm1d(512),
            nn.LeakyReLU(0.1),
            nn.Dropout(0.15),
            nn.Linear(512, 256),
            nn.BatchNorm1d(256),
            nn.LeakyReLU(0.1),
            nn.Dropout(0.1),
            nn.Linear(256, num_classes)
        )

• Test audio denoising link
• Test FocalLossBCE link
• Make prediction based on all 5s segments of the audio link
• Add albumentations
• Test extracting not the center 5 seconds, but the first 5 seconds
• Test 3 channels
• Test ImageNet normalization for 3 channels if the weights are pretrained T.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225]),
• Test melspec more thoroughly (N_MELS, HOP_LENGTH)
• [ ]