MobileViT-FPN-PixelShuffle

Research codebase for ultrasound FMC-style reconstruction experiments: MobileViT-style backbones, FPN / UNet-style decoders, pixel shuffle heads, optional Mixture-of-Experts (MoE) and DANN variants. Documentation and primary code comments are in English.

Related conference paper

This repository supports work presented in connection with the Acoustical Society of America — Honolulu 2025:

Enabling low-cost full matrix capture acquisition using MobileViT with binary ultrasonic data.

Authors: Rafael Niddam (École de technologie supérieure, 1100 Notre-Dame St. W., A-1340, Montréal, QC H3C 1K3, Canada, Guillaume Painchaud-April (Evident Sci., QC, Canada), Alain Le Duff (Evident Sci., QC, Canada), and Pierre Belanger (École de technologie supérieure, Montréal, QC, Canada).

Contact: Rafael Niddam on LinkedIn

Features

• Models: MobileViT v3 family (FPN, UNet, SegFormer, SPT, LSA, pixel variants), Lite MobileViT blocks, MobileNet v2 FPN/UNet, MoE and DANN wrappers (see models/__init__.py registry). The final FMC pixel-shuffle segmentation model is MobileViTv3_v1_dynamicFPNpixel2 implemented in models/segmentation/mobilevit_v3_pixel2.py.
• Training: training/train.py (argparse, --list-models), plus training/train_moe.py and training/train_dann.py for specialized setups.
• Evaluation & analysis: use evaluation/evaluate_loop2.py as the main FMC validation script; evaluation/evaluate_moe2.py for MoE. A trailing 2 in a script name marks the final version in that line (see evaluation/__init__.py).
• Utilities: losses, NCC, plotting helpers in utils/.

Repository layout

MobileViT-FPN-PixelShuffle/
├── configs/          # Hyperparameters and training defaults
├── evaluation/       # Evaluation and post-hoc analysis
├── models/           # Backbones, decoders, segmentation heads, MoE
├── training/         # Training entry points
├── utils/            # Metrics, losses, FMC visualization helpers
├── requirements.txt
├── LICENSE
└── README.md

Setup

Requirements

• Python 3.10+
• CUDA GPU recommended for training (CPU may work for tiny tests)

Install

python -m venv .venv

Activate:

• Windows (PowerShell): .\.venv\Scripts\Activate.ps1
• Linux / macOS: source .venv/bin/activate

pip install -r requirements.txt

Usage

List registered model names

python training/train.py --list-models

Train (example; adjust --model and --batch_size as needed)

python training/train.py --model MobileViTv3_v1_dynamicFPNpixel2 --batch_size 16

Evaluate (final workflow: loop presets + project paths)

python evaluation/evaluate_loop2.py --preset 0
# Optional: checkpoint and folder under data/ (e.g. Test_dataset) or absolute path
python evaluation/evaluate_loop2.py --preset 0 --checkpoint path/to/Model.pth --test-dir Test_dataset

Preset 0 reads .mat files from data/Test_dataset/. Other presets use data/<subdir>/ (see test_subdir in evaluate_loop2.py). Weights: weights/, or --checkpoint. Older scripts: evaluation/evaluate.py, evaluation/evaluate_loop.py.

Additional trainers (older, script-style imports—may require aligning sys.path or project root with your layout):

• python training/train_moe.py
• python training/train_dann.py

Configuration & data paths

• Defaults live in configs/ (e.g. config_mobileunet.py).
• Training: data/train_dataset / data/valid_dataset (see training/train.py). Evaluation: evaluate_loop2.py — preset 0 uses data/Test_dataset/; other presets use data/<subdir>/ + weights/ (or --checkpoint / --test-dir).

License

This project is released under the MIT License; see LICENSE.