Team Name: Wellermen
Team Members:
- Pribék Barnabás - ETWSGU
- Csernák Gergő - TVEXOV
- Molnár Bence - JINVFB
This project uses the BirdCLEF 2024 competition dataset for bird species identification from audio recordings. The solution is built using Keras and TensorFlow, implementing the following main steps:
- Data Source: BirdCLEF 2024 competition dataset (182 bird species, ~24,459 audio recordings)
- Data Processing: Converting OGG format audio files to mel-spectrograms
- Model: EfficientNetV2-B2 based image classifier (pretrained on ImageNet)
- Augmentation: MixUp, time-masking, and frequency-masking techniques
The dataset can be downloaded from the BirdCLEF 2024 Kaggle competition. This project works with data stored in Google Drive for easier access.
The primary notebook containing the complete data pipeline:
1. Data Source and Download:
- Google Drive integration for accessing the BirdCLEF 2024 dataset
- Dataset verification and validation
- Metadata loading (
train_metadata.csv)
2. Data Exploration and Visualization:
- Audio file loading and playback
- Waveform visualization
- Mel-spectrogram visualization
- Species distribution statistics
3. Data Preparation:
- Train/Validation/Test Split: 80/20 ratio split
- Audio Preprocessing:
- Cropping/padding to uniform 15-second length
- 32,000 Hz sampling rate
- Mel-spectrogram generation (128x384 pixels)
- Augmentation (training data only):
- MixUp (α=0.4)
- Time-masking (6-12%)
- Frequency-masking (6-10%)
- TensorFlow Dataset Pipeline: Efficient, batch-based data loading
Final Output:
train_ds: Training dataset (19,567 samples, augmented)valid_ds: Validation dataset (4,892 samples, not augmented)- Format: (128, 384, 3) spectrograms + one-hot encoded labels (182 classes)
- Google Colab account
- Google Drive: Dataset uploaded at the following path:
My Drive/Deep_Learning_NHF/birdclef-2024/ ├── train_audio/ │ ├── asbfly/ │ ├── ashdro1/ │ └── ... └── train_metadata.csv
-
Open the notebook in Google Colab:
-
Execute cells in order:
- Import Libraries: Install dependencies
- Kaggle Setup and Data Download: Mount Google Drive
- Configuration: Set hyperparameters
- Load Class Names: Initialize classes
- Meta Data: Load and filter data
- Data Split: Create train/validation split
- Data Loaders: Build TensorFlow Dataset pipeline
- Visualization: Display and verify batches
- Model: Builds, compiles, trains, evaluates and tests the model using all species data and all the audioscapes for evauation
- Configure for Reduced Species: Configures the model to use only the first 10 species data for training
- Reduced Model: Builds, compiles, trains and tests the model using the first 10 species and all the audioscapes for evaluation
- Modify Test Filepaths to Limit Count: Evaluates using only the first 10 audioscape files
-
Drive Authorization: On first run, authorize Google Drive access
-
Verification: Sample images and spectrograms will be displayed at the end of the notebook
img_size = [128, 384] # Spectrogram dimensions
batch_size = 64 # Batch size
duration = 15 # Audio length (seconds)
sample_rate = 32000 # Sampling rate
epochs = 10 # Training epochs
preset = 'efficientnetv2_b2_imagenet' # Model- TensorFlow 2.19.0
- Keras 3.10.0
- KerasCV 0.9.0
- Librosa (audio processing)
- Pandas, NumPy (data handling)
- Matplotlib (visualization)
- Total samples loaded: 24,459
- Number of bird species: 182
- Training samples: 19,567
- Validation samples: 4,892
- Spectrogram dimensions: 128 (frequency) × 384 (time) × 3 (channels)
- The original Kaggle notebook has been modified with Google Drive support
- Compatibility issues fixed for the latest library versions
- Reproducibility:
seed = 42
This project follows the BirdCLEF 2024 competition rules and dataset license.