PCG Transformer Feature Extraction Model
Budget / Salary$30–250
TypeFreelance project
LocationRemote
Posted2 hours ago
I’m advancing my Master’s research on congenital-heart-disease diagnosis from phonocardiogram (PCG) recordings and now need a complete, ready-to-run model. Your core mission is to craft rich feature representations: convert the raw heart-sound waveforms into spectrograms, then feed them through a Transformer backbone enhanced with a Minimax Transformation (MMT) module and Cosine-Similarity-Guided Context Attention.
Key goals
• Produce discriminative spectrogram-based features that alleviate the severe class imbalance of the ZCHSound dataset.
• Integrate MMT and the cosine-guided attention cleanly into PyTorch so each component can be toggled on/off for ablation.
• Train, validate and test on the official ZCHSound splits, reporting multi-class metrics (accuracy, F1, AUC) and benchmarking against at least two published baselines.
Deliverable checklist
1. Fully documented Python/PyTorch codebase (compatible with CUDA) that downloads/reads the dataset, preprocesses PCG audio into spectrograms, builds the Transformer-MMT network, trains, and outputs metrics.
2. Saved best model weights plus an inference script that takes a new .wav file and returns a predicted class with confidence.
3. Concise technical report (Markdown is fine) summarising feature-extraction rationale, architecture diagram, training settings, and comparative results.
Acceptance criteria
– Replicable results on my machine using a single requirements.txt or conda-env.yml.
– F1-score meets or exceeds the strongest baseline cited in the report.
– Code includes clear inline comments for every custom layer, especially the MMT block and cosine attention.
Tools & concepts that should appear naturally in your implementation: Python 3.10+, PyTorch, torchaudio, NumPy, SciPy, matplotlib for spectrogram visualisation, and standard data-imbalance techniques such as focal loss or weighted sampling.
If this challenge aligns with your deep-learning and biomedical-signal experience, I’m eager to review your approach and timeline.
Key goals
• Produce discriminative spectrogram-based features that alleviate the severe class imbalance of the ZCHSound dataset.
• Integrate MMT and the cosine-guided attention cleanly into PyTorch so each component can be toggled on/off for ablation.
• Train, validate and test on the official ZCHSound splits, reporting multi-class metrics (accuracy, F1, AUC) and benchmarking against at least two published baselines.
Deliverable checklist
1. Fully documented Python/PyTorch codebase (compatible with CUDA) that downloads/reads the dataset, preprocesses PCG audio into spectrograms, builds the Transformer-MMT network, trains, and outputs metrics.
2. Saved best model weights plus an inference script that takes a new .wav file and returns a predicted class with confidence.
3. Concise technical report (Markdown is fine) summarising feature-extraction rationale, architecture diagram, training settings, and comparative results.
Acceptance criteria
– Replicable results on my machine using a single requirements.txt or conda-env.yml.
– F1-score meets or exceeds the strongest baseline cited in the report.
– Code includes clear inline comments for every custom layer, especially the MMT block and cosine attention.
Tools & concepts that should appear naturally in your implementation: Python 3.10+, PyTorch, torchaudio, NumPy, SciPy, matplotlib for spectrogram visualisation, and standard data-imbalance techniques such as focal loss or weighted sampling.
If this challenge aligns with your deep-learning and biomedical-signal experience, I’m eager to review your approach and timeline.
Apply on Freelancer →
Project sourced from Freelancer.com. Applications happen directly on the original platform — we never collect your data.