Model Comparison

This example instantiates every registered model and benchmarks parameter count, feature dimensions, and inference speed on synthetic input.

Source: examples/model_comparison.py

Usage

python examples/model_comparison.py
python examples/model_comparison.py --input-channels 12 --output-size 5 --signal-length 5000
python examples/model_comparison.py --device mps

Arguments

Flag	Default	Description
--input-channels	1	Number of ECG leads
--output-size	4	Number of output classes
--signal-length	3000	Length of input signal in samples
--device	cpu	Device to benchmark on
--num-runs	50	Number of inference runs for timing

How It Works

The script iterates over all registered models:

import torch

from deepecgkit.registry import get_model, get_model_names

for model_name in get_model_names():
    model_class = get_model(model_name)
    model = model_class(input_channels=1, output_size=4)
    model.eval()

    x = torch.randn(1, 1, 3000)
    try:
        with torch.no_grad():
            output = model(x)
    except Exception as error:
        print(f"{model_name}: FAILED - {error}")
        continue

    total_params = sum(p.numel() for p in model.parameters())
    print(f"{model_name}: {total_params:,} params, output {tuple(output.shape)}")

A model that cannot take the requested input is reported and skipped: afmodel is built for 30-second recordings, so it needs --signal-length 9000 at 300 Hz.

For each model it reports:

• Total parameters and trainable parameters
• Feature dimension (model.feature_dim)
• Output shape
• Inference time (averaged over multiple runs with warmup)

This helps you choose the right model for your latency and memory constraints. See the Model Zoo for guidance on model selection.