2. Detect

PartiNet uses a modified version of an adaptive YOLO architecture called DynamicDet (Lin et al., 2023) to identify particles in cryo-EM micrographs. This stage provides highly accurate particle detection with customizable confidence and overlap thresholds.

Quick Start

Local Installation

partinet detect \
    --weight /path/to/downloaded/model_weights.pt \
    --source /data/partinet_picking/denoised \
    --device 0,1,2,3 \
    --project /data/partinet_picking

Apptainer/Singularity

apptainer exec --nv --no-home \
    -B /data oras://ghcr.io/wehi-researchcomputing/partinet:main-singularity partinet detect \
    --weight /path/to/downloaded/model_weights.pt \
    --source /data/partinet_picking/denoised \
    --device 0,1,2,3 \
    --project /data/partinet_picking

Docker

docker run --gpus all -v /data:/data \
    ghcr.io/wehi-researchcomputing/partinet:main partinet detect \
    --weight /path/to/downloaded/model_weights.pt \
    --source /data/partinet_picking/denoised \
    --device 0,1,2,3 \
    --project /data/partinet_picking

Parameters

Required Parameters

Parameter	Description	Example
--weight	Path to pre-trained model weights file	/data/models/partinet_yolov7_weights.pt
--source	Path to directory containing micrographs to process	/data/partinet_picking/denoised
--project	Path to project directory where outputs will be saved	/data/partinet_picking

Customizable Parameters

Parameter	Type	Description	Default	Notes
--conf-thres	float	Confidence threshold for detection (0.0-1.0)	0.1	Lower = more detections, higher = fewer but more confident
--iou-thres	float	IoU threshold for non-maximum suppression (0.0-1.0)	0.2	Higher = more aggressive overlap removal
--device	string	GPU devices to use (comma-separated)	0,1,2,3	Use 0 for single GPU, 0,1 for two GPUs, etc. Leave empty for CPU only

Parameter Configuration Guide

Confidence Threshold (--conf-thres)

The confidence threshold determines the minimum confidence score required for a detection to be considered valid. Even with denoised micrographs, overall confidence will be low due to low SNR. It is not recommended to increase this above 0.5. We recommend setting this threshold quite low 0.0-0.3 during this stage, and then filtering during STAR file generation.

Recommended starting values:

• New datasets, small proteins, low defocus: 0.0-0.2
• Large proteins/complexes, higher defocus: 0.2-0.5

IoU Threshold (--iou-thres)

The Intersection over Union (IoU) threshold controls how aggressively overlapping detections are removed. The higher the value, the more that overlapping particles for removed

Recommended starting value: 0.2

Model Weights

Pre-trained Weights

PartiNet model weights are available on HuggingFace. Weights can be downloaded through the browser or through CLI via Git LFS

# Verify Git LFS is installed
git lfs --help
mkdir PartiNet_weights
cd PartiNet_weights
git clone git@hf.co:MihinP/PartiNet

You will see two .pt files available: denoised_micrographs.pt and raw_micrographs.pt.

Custom Model Weights

If you have trained your own PartiNet model you may use custom weights:

partinet detect \
    --backbone-detector yolov7-w6 \
    --weight /path/to/your/custom_weights.pt \
    --source /data/partinet_picking/denoised \
    --conf-thres 0.4 \
    --iou-thres 0.2 \
    --device 0,1 \
    --project /data/partinet_picking

Input Requirements

Supported Formats

• PNG files (.png)
• JPEG files (.jpg, .jpeg)

Directory Structure

Your source directory should contain micrographs:

denoised/
├── micrograph_001.png
├── micrograph_002.png
├── micrograph_003.png
└── ...

Output

Directory Structure

After detection, your project directory will contain:

partinet_picking/
├── motion_corrected/          # 📁 Your input micrographs
│   ├── micrograph1.mrc
│   ├── micrograph2.mrc
│   └── ...
├── denoised/                  # 🧹 Created by denoise stage
│   ├── micrograph1.mrc
│   ├── micrograph2.mrc
│   └── ...
├── exp/                       # 🎯 Created by detect stage
│   ├── labels/               # 📋 Detection coordinates
│   │   ├── micrograph1.txt
│   │   ├── micrograph2.txt
│   │   └── ...
│   ├── micrograph1.mrc       # 🖼️ Micrographs with detections drawn
│   ├── micrograph2.mrc
│   └── ...
└── partinet_detect.log

Log File Output

The partinet_detect.log file provides detailed processing information:

2025-09-10 16:08:07 | INFO | Using devices: ['cuda:0', 'cuda:1', 'cuda:2', 'cuda:3']
2025-09-10 16:08:11 | INFO | Model loaded on cuda:0
2025-09-10 16:08:15 | INFO | Model loaded on cuda:1
2025-09-10 16:08:17 | INFO | Model loaded on cuda:2
2025-09-10 16:08:20 | INFO | Model loaded on cuda:3
2025-09-10 16:08:21 | INFO | [cuda:0] Processing 000000633063778279389_FoilHole_30903560_Data_30900510_30900512_20201031_010810_fractions_patch_aligned.png
2025-09-10 16:08:21 | INFO | [cuda:1] Processing 000005251007454303721_FoilHole_30909875_Data_30900510_30900512_20201101_040227_fractions_patch_aligned.png
2025-09-10 16:08:21 | INFO | [cuda:2] Processing 000012049843629116854_FoilHole_30902824_Data_30900510_30900512_20201030_230632_fractions_patch_aligned.png
2025-09-10 16:08:22 | INFO | [cuda:3] Processing 000018235333323367981_FoilHole_30905009_Data_30900510_30900512_20201031_082713_fractions_patch_aligned.png
2025-09-10 16:08:24 | INFO | [cuda:1] 000005251007454303721_FoilHole_30909875_Data_30900510_30900512_20201101_040227_fractions_patch_aligned.png: 94 particles, 3516.2 ms
2025-09-10 16:08:24 | INFO | [cuda:1] Processing 000022392129605454366_FoilHole_30903097_Data_30900534_30900536_20201031_002940_fractions_patch_aligned.png
2025-09-10 16:08:25 | INFO | [cuda:2] 000012049843629116854_FoilHole_30902824_Data_30900510_30900512_20201030_230632_fractions_patch_aligned.png: 272 particles, 3604.4 ms

Output File Formats

Detection Coordinates (labels/*.txt)

Each txt file corresponds to a single micrograph. Each line represents one detection in YOLO format:

class_id x_center y_center width height confidence
0 0.5234 0.3456 0.0234 0.0198 0.8567

Where coordinates are normalized (0.0-1.0) relative to micrograph dimensions.

What's Next

• STAR

Congratulations! You have picked particles from your micrographs, you can now move to the final stage with STAR.

References

Lin, Z., Wang, Y., Zhang, J., & Chu, X. (2023). DynamicDet: A Unified Dynamic Architecture for Object Detection (arXiv:2304.05552). arXiv. https://doi.org/10.48550/arXiv.2304.05552

See Also

• Denoise
• STAR