uday-bhatia/ecg-digitization-experiments

ECG Digitization Experiments

This repository contains archived PyTorch .pth checkpoints from ECG digitization research experiments (Kaggle competition).

About

These are model weights for ECG image-to-signal digitization - extracting ECG waveform values from paper ECG images.

Competition: ECG Image Digitization Challenge

Experiments Overview

Experiment	Architecture	Description	Best SNR
v7	EfficientNet-B4	Early baseline	~15 dB
v9	EfficientNet-B4	Improved training	~16 dB
v10	EfficientNet-B4	Multi-scale features	~17 dB
v10_1	EfficientNet-B4	Refinements	~17 dB
v11	EfficientNet-B4	1.5x scale	~17 dB
v14	ConvNeXt + SimDR	SimDR heatmap approach	~18 dB
v15	Per-Lead CNN	Per-lead extraction	~18 dB
v16	Per-Lead + BiLSTM	Temporal coherence	~19 dB
v18	Refiner Network	Post-processing refiner	~19 dB
v19	Augraphy Augmentation	Paper degradation aug	~20 dB
v20	Integral Regression	Integral loss	~20 dB
v21	GRU Refiner	GRU-based refinement	~20 dB
v22	ConvNeXt-Base + U-Net	Multi-scale fusion + Height Attention + BiLSTM	~22 dB
v23	V22 + DSNT	DSNT Sub-Pixel Head (Rank 3 technique)	~22.35 dB
mixed	Various	Early mixed training	~16 dB
mixed_v4	Various	Mixed v4	~17 dB
mixed_v5	Various	Mixed v5	~17 dB

V23 Architecture (Best Model)

① ConvNeXt-Base Encoder → L3 + L4 multi-scale features
② Feature Fusion (fine details + semantics)
③ U-Net Decoder with skip connections
④ Height Attention (learns vertical regions)
⑤ BiLSTM (temporal coherence)
⑥ Conv1D + Linear → Sigmoid (V22 head)
⑦ DSNT Sub-Pixel Head (Rank 3 winner technique)

Directory Structure

├── README.md
├── v7/                    # EfficientNet-B4 baseline
├── v9/                    # Improved training
├── v10/                   # Multi-scale features
├── v10_1/                 # Refinements
├── v11/                   # 1.5x scale
├── v14/                   # ConvNeXt + SimDR
├── v15/                   # Per-lead CNN
├── v16/                   # Per-lead + BiLSTM
├── v18/                   # Refiner network
├── v19/                   # Augraphy augmentation
├── v20/                   # Integral regression
├── v21/                   # GRU refiner
├── v22/                   # ConvNeXt-Base + U-Net
├── v23/                   # V22 + DSNT (BEST)
├── mixed/                 # Mixed training v1
├── mixed_v4/              # Mixed training v4
├── mixed_v5/              # Mixed training v5
└── code/
    ├── scripts/           # All training scripts
    └── notebooks/         # Kaggle inference notebooks

Checkpoint Format

Each .pth file contains:

• model: Model state dict
• opt: Optimizer state dict
• epoch: Training epoch
• snr: Validation SNR (dB)
• mae: Mean Absolute Error (pixels)

Usage

import torch

# Load checkpoint
checkpoint = torch.load("v23/v23_epoch038.pth", map_location="cpu")

# Load model state
model.load_state_dict(checkpoint['model'])

# Check metrics
print(f"Epoch: {checkpoint['epoch']}")
print(f"SNR: {checkpoint['snr']:.2f} dB")

Key Files

• v23/v23_epoch038.pth - Best V23 checkpoint (22.35 dB)
• v22/v22_best_snr.pth - Best V22 checkpoint
• code/scripts/train_v23.py - V23 training script
• code/scripts/inference_v23.py - V23 inference script
• code/notebooks/kaggle_v23_inference.ipynb - Kaggle submission notebook

License

Research use only. Please cite if you use these weights.

Framework

• PyTorch 2.x
• timm (ConvNeXt-Base encoder)
• albumentations (augmentation)