Skip to content

xtmtd/OTU-Former

BranchesTags

Repository files navigation

OTU-Former

中文文档 | English

An image-based morphological OTU (Operational Taxonomic Unit) delineation toolkit. Provides a unified otuformer CLI with commands for self-supervised pretraining, ArcFace metric learning finetuning, embedding extraction (including quality metrics/UMAP when labels are provided), UPGMA hierarchical clustering, expert-corrected annotation, community diversity analysis, CAM visualization, and ONNX model export.

Overview

All functionality is accessed through a single entry point:

otuformer <command> [options]
Command Description
doctor Diagnose environment and dependency status
pretrain Self-supervised contrastive pretraining (DINO/iBOT style)
finetune ArcFace metric learning finetuning
extract Extract image embeddings (ONNX-accelerated)
cluster UPGMA hierarchical clustering into morphological OTUs
annotate Apply expert corrections to produce refined OTU annotations
diversity Compute community alpha diversity indices (Shannon, Simpson, Chao1, Faith's PD, etc.)
cam Generate GradCAM and other heatmaps
export Export model to ONNX format

Features

  • Unified CLI: Single otuformer entry point — covers the complete pipeline from pretraining to diversity analysis
  • Self-Supervised Pretraining: DINO/iBOT-style teacher-student ViT contrastive learning with global/local crops and masked-token consistency
  • ArcFace Metric Learning Finetuning: Optimize model with labelled data to produce discriminative embeddings for OTU clustering
  • Multi-Mode Embedding Extraction: CLS token, patch-topk, and attention-pool modes; optional ONNX acceleration
  • Embedding Quality Evaluation: NMI, ARI, Recall@K, kNN accuracy, mAP@R, Silhouette Score, Linear Probing Accuracy, and more
  • UPGMA Hierarchical Clustering: Build distance matrix and phylogenetic tree, auto-partition OTUs at multiple distance cutoffs
  • PCA Whitening & Local Scaling: Optional PCA whitening and Mutual-Proximity-style local scaling for improved clustering
  • Bootstrap Support Estimation: Subsample/bootstrap modes for branch support estimation
  • Expert-Corrected Annotation: Accept manual correction CSVs to produce refined OTU assignments with intra-class distance summaries
  • Community Diversity Analysis: Richness, Chao1, ACE, Shannon, Simpson, Hill numbers (q0/q1/q2), Pielou evenness, Faith's PD (MPD), weighted PD (MPD_w), and other alpha diversity metrics; phylogenetic diversity is computed from an OTU-centroid neighbor-joining tree when embeddings are provided
  • Model Interpretability: Six CAM algorithms — Grad-CAM, Grad-CAM++, LayerCAM, Score-CAM, Eigen-CAM, Ablation-CAM
  • ONNX Export & Acceleration: Export to ONNX format for 2-5x faster CPU inference
  • UMAP Visualization: Embedding space dimensionality reduction for intuitive morphological distribution
  • Parallel Processing: Multi-threaded image processing with configurable CPU threads
  • Comprehensive Logging: All commands automatically save run logs to the logs/ subdirectory of the output directory
  • Graceful Shutdown: Safe Ctrl+C interruption with partial result preservation

Requirements

  • Python 3.11+
  • Operating Systems: macOS, Linux, Windows
  • Devices: CPU (required); CUDA GPU / Apple Silicon MPS (optional)

Installation

Recommended: use an isolated Python environment to avoid dependency conflicts with your system/site-packages.

Before installation, clone the repository and enter the project directory:

git clone https://github.com/xtmtd/OTU-Former.git
cd OTU-Former

Isolated Environment

Choose one of the following:

Option 1: conda

conda create -n otuformer python=3.11 -y
conda activate otuformer
pip install -e .

Option 2: uv + venv

uv venv .venv
source .venv/bin/activate
uv pip install -e .

Option 3: stdlib venv + pip

python -m venv .venv
source .venv/bin/activate
pip install -e .

Usage

Recommended workflow command order:

  1. doctor — Check environment
  2. pretrain — Self-supervised pretraining
  3. finetune — ArcFace finetuning
  4. extract — Extract embeddings
  5. cluster — UPGMA clustering into OTUs
  6. annotate — Apply expert corrections
  7. diversity — Diversity analysis

Doctor Command

Diagnose environment and dependency readiness.

otuformer doctor

The report includes:

  • Python version and available devices (cpu, cuda, mps)
  • Key package versions and status (ok/missing/outdated)

Output:

  • Diagnostic report printed to terminal

Pretrain Command

Self-supervised contrastive pretraining using DINO/iBOT-style teacher-student ViT, no labels required.

# Basic usage
otuformer pretrain \
    --input-images-dir ./images \
    --out-dir runs/pretrain

# Use CSV to specify training subset
otuformer pretrain \
    --train-data images.csv \
    --input-images-dir ./images \
    --out-dir runs/pretrain

# Custom model and training parameters
otuformer pretrain \
    --input-images-dir ./images \
    --model-name vit_small_patch16_224 \
    --max-epochs 100 \
    --lr 1e-3 \
    --batch-size 64 \
    --out-dir runs/pretrain
Parameter Description Default
--train-data Training data CSV (with image column, paths relative to --input-images-dir); if omitted, all images under --input-images-dir are used recursively None
--input-images-dir Root image directory Required
--out-dir Output directory runs/pretrain
--model-name timm backbone name vit_tiny_patch16_224
--out-dim SSL projector output dimension 256
--max-epochs Pretraining epochs 50
--lr Base learning rate 5e-4
--weight-decay AdamW weight decay 0.05
--warmup-epochs Warmup epochs 3
--global-crop-size Global crop resolution 224
--local-crop-size Local crop resolution 96
--local-crops Number of local crops 6
--mask-ratio Masked token ratio 0.5
--lambda-local Local crop loss weight 1.5
--lambda-mask Masked token loss weight 1.0
--teacher-momentum Initial EMA momentum 0.995
--teacher-momentum-end Final EMA momentum 0.999
--student-temp Student temperature 0.1
--teacher-temp-start Initial teacher temperature 0.04
--teacher-temp-end Final teacher temperature 0.07
--disable-cross-view-loss Disable cross-view global loss No
--resume Checkpoint path for resuming interrupted training None
--log-every-n-steps Log metrics every N steps 50
--save-every-epochs Save checkpoint every N epochs 10
--keep-last-checkpoints Keep only last N checkpoints 10
--visualize-data CSV for periodic embedding metrics/UMAP; if omitted, --train-data is reused None
--extract-size Image size for embedding extraction (<=0 means auto) 0
--metrics-sample-size Max samples for periodic metrics 10000
--umap-n-neighbors UMAP n_neighbors 15
--umap-min-dist UMAP min_dist 0.1
--umap-metric UMAP distance metric cosine
--visualize-class-number Max classes in UMAP plot 20
--disable-embedding-metrics Disable periodic embedding metrics + UMAP No
--batch-size Batch size 32
--num-workers DataLoader workers 4
--cpus PyTorch/MKL CPU threads 12
--device auto/cpu/cuda/mps auto
--seed Random seed 42

Output:

  • logs/pretrain.log — Run log
  • checkpoints/ — Model checkpoints (SSL_latest.pth, SSL_best.pth)
  • metrics.json / metrics.csv — Training metrics
  • umap.pdf — UMAP visualization (when not disabled)

Finetune Command

ArcFace metric learning finetuning with labelled data.

# Basic usage
otuformer finetune \
    --checkpoint runs/pretrain/SSL_latest.pth \
    --train-data labels.csv \
    --input-images-dir ./images \
    --out-dir runs/finetune

# Resume training
otuformer finetune \
    --resume runs/finetune/finetune_latest.pth \
    --train-data labels.csv \
    --input-images-dir ./images \
    --out-dir runs/finetune

# Custom parameters
otuformer finetune \
    --checkpoint runs/pretrain/SSL_latest.pth \
    --train-data labels.csv \
    --input-images-dir ./images \
    --model-name vit_small_patch16_224 \
    --finetune-epochs 50 \
    --finetune-lr 3e-4 \
    --freeze-ratio 0.5 \
    --loss arcface \
    --out-dir runs/finetune
Parameter Description Default
--checkpoint Pretrained checkpoint path (typically runs/pretrain/SSL_latest.pth) Required
--resume Finetune checkpoint path to resume from None
--train-data Training data CSV (with image and label columns) Required
--input-images-dir Root image directory Required
--out-dir Output directory runs/finetune
--model-name timm backbone name vit_tiny_patch16_224
--metric-embed-dim ArcFace projector embedding dimension 256
--finetune-epochs Finetuning epochs 20
--finetune-lr Finetuning learning rate 1e-4
--freeze-ratio Fraction of backbone blocks to freeze (0.0=none, 1.0=all) 0.7
--loss Metric-learning loss name arcface
--batch-size Batch size 32
--num-workers DataLoader workers 4
--cpus PyTorch/MKL CPU threads 12
--device auto/cpu/cuda/mps auto
--seed Random seed 42
--log-every-n-steps Log metrics every N steps 50
--save-every-epochs Save checkpoint every N epochs 10
--keep-last-checkpoints Keep only last N checkpoints 10
--visualize-data CSV for periodic embedding metrics/UMAP; if omitted, --train-data is reused None
--extract-size Image size for embedding extraction (<=0 means auto) 0
--metrics-sample-size Max samples for periodic metrics 10000
--umap-n-neighbors UMAP n_neighbors 15
--umap-min-dist UMAP min_dist 0.1
--umap-metric UMAP distance metric cosine
--visualize-class-number Max classes in UMAP plot 20
--disable-embedding-metrics Disable periodic embedding metrics + UMAP No

Output:

  • logs/finetune.log — Run log
  • checkpoints/ — Model checkpoints (finetune_latest.pth, finetune_best.pth)
  • metrics.json / metrics.csv — Training metrics
  • umap.pdf — UMAP visualization (when not disabled)

Extract Command

Extract embeddings from images. Supports ONNX model for accelerated CPU inference.

# Using PyTorch checkpoint (CLS token mode)
otuformer extract \
    --checkpoint runs/finetune/finetune_latest.pth \
    --input-images-dir ./images \
    --out-dir runs/extract

# Using ONNX model (2-5x faster on CPU)
otuformer extract \
    --checkpoint runs/finetune/finetune_latest.pth \
    --input-images-dir ./images \
    --onnx-path runs/export/encoder.onnx \
    --out-dir runs/extract

# With label CSV in attention-pool mode
otuformer extract \
    --checkpoint runs/finetune/finetune_latest.pth \
    --input-images-dir ./images \
    --label-csv labels.csv \
    --token-mode attention-pool \
    --out-dir runs/extract
Parameter Description Default
--checkpoint Pretrain or finetune checkpoint path; required unless --onnx-path is provided None
--input-images-dir Input image directory Required
--out-dir Output directory runs/extract
--model-name timm backbone name vit_tiny_patch16_224
--extract-size Resize/crop size for extraction 224
--use-projector-output Use projector output instead of CLS token No
--use-student Load student weights instead of teacher (EMA) No
--token-mode cls/patch-topk/attention-pool cls
--topk-patches Top-K patch tokens for patch-topk mode 20
--attention-pooling-type lightweight/multihead/gated lightweight
--attention-pooling-epochs Epochs to finetune attention query 20
--label-csv Label CSV for embedding quality evaluation and UMAP generation None
--metrics-sample-size Max samples for metrics evaluation 10000
--umap-n-neighbors UMAP n_neighbors 15
--umap-min-dist UMAP min_dist 0.1
--umap-metric UMAP distance metric cosine
--visualize-class-number Max classes in UMAP plot 20
--disable-umap Skip UMAP generation No
--batch-size Batch size 32
--num-workers DataLoader workers 4
--device auto/cpu/cuda/mps auto
--onnx-path ONNX model path (uses ONNX Runtime when provided) None
--seed Random seed 42

Output:

  • embeddings.csv — Embedding vectors (id, dim_0, dim_1, ...)
  • metrics.csv — Quality metrics (when --label-csv is provided)
  • umap.pdf — UMAP visualization (when --label-csv is provided and not disabled)

Cluster Command

Cluster embeddings into morphological OTUs via UPGMA hierarchical clustering.

# Basic usage
otuformer cluster \
    --embeddings runs/extract/embeddings.csv \
    --out-dir runs/cluster

# With PCA whitening, local scaling, and bootstrap support
otuformer cluster \
    --embeddings runs/extract/embeddings.csv \
    --pca-whitening true \
    --local-scaling true \
    --num-replicates 100 \
    --out-dir runs/cluster

# Custom distance metric and cutoff range
otuformer cluster \
    --embeddings runs/extract/embeddings.csv \
    --distance euclidean \
    --cutoff-min 0.1 \
    --cutoff-max 0.8 \
    --cutoff-step 0.02 \
    --out-dir runs/cluster

# With label CSV for partition quality evaluation
otuformer cluster \
    --embeddings runs/extract/embeddings.csv \
    --label-csv labels.csv \
    --out-dir runs/cluster
Parameter Description Default
--embeddings Embeddings CSV Required
--out-dir Output directory runs/cluster
--distance Distance metric: cosine/euclidean cosine
--prefix Cluster prefix for partition table labels OTU
--pca-whitening Enable PCA whitening (true/false) false
--pca-components PCA components 256
--local-scaling Enable Mutual-Proximity local scaling (true/false) false
--local-k Fixed k for local scaling (0=auto) 0
--local-k-strategy Auto-k strategy: adaptive/sqrt/log/fixed adaptive
--cutoff-min Minimum distance cutoff 0.05
--cutoff-max Maximum distance cutoff 1.0
--cutoff-step Cutoff step size 0.05
--custom-cutoffs Custom cutoffs (comma-separated, overrides min/max/step) None
--support-mode Support estimation mode: subsample/bootstrap subsample
--num-replicates Support estimation replicates (0=disabled) 0
--subsample-ratio Feature fraction for subsample mode 0.8
--support-display-cutoff Minimum support threshold for tree labels 50.0
--save-bootstrap-trees Save all bootstrap trees (true/false) false
--save-distances Save full distance matrix (true/false) false
--max-distance-pairs Max distance matrix pairs 1000000
--label-csv / --labels Optional label CSV for partition quality evaluation None
--metrics-sample-size Max samples for metrics 10000
--cpus CPU threads 8
--random-state Random seed 42

Output:

  • UPGMA/UPGMA_Cosine.nwk — Newick format phylogenetic tree
  • UPGMA/partitions/partition_scan.csv — Threshold scan results
  • UPGMA/partitions/tables/partition_<cutoff>_assignments.csv — OTU assignments at each cutoff
  • UPGMA/partitions/UPGMA_tree_partitions.pdf — Tree and partition visualization
  • UPGMA/metrics.csv — Partition quality metrics (when --label-csv is provided)
  • UPGMA/metrics_dashboard.pdf — Metrics dashboard (when --label-csv is provided)
  • distance_statistics/ — Distance statistics and distribution plots

Annotate Command

Apply expert corrections to cluster assignments, producing refined OTU annotations.

# Basic usage
otuformer annotate \
    --raw-assignments runs/cluster/UPGMA/partitions/tables/partition_0.30_assignments.csv \
    --corrections corrections.csv \
    --out-dir runs/annotate

# With embeddings (recompute intra-class distances and annotated UPGMA tree)
otuformer annotate \
    --raw-assignments runs/cluster/UPGMA/partitions/tables/partition_0.30_assignments.csv \
    --corrections corrections.csv \
    --embeddings runs/extract/embeddings.csv \
    --show-annotation-bar \
    --out-dir runs/annotate
Parameter Description Default
--raw-assignments Raw partition assignment CSV from cluster output Required
--corrections Corrections CSV (minimum: id/image and cluster columns) Required
--embeddings Optional embeddings CSV for distance recomputation None
--support-display-cutoff Minimum support threshold for annotated UPGMA tree labels 50.0
--figure-width Annotated UPGMA PDF width in inches Auto
--annotate-bar-width Relative width of corrected OTU color bars 0.08
--show-annotation-bar Show corrected OTU annotation bars in annotated UPGMA PDF No
--show-partitioning-bars Show partitioning bars in annotated UPGMA PDF No
--out-dir Output directory runs/annotate

Output:

  • partition_<cutoff>_assignments.csv — Corrected assignments
  • partition_<cutoff>_assignments_changed_only.csv — Changed rows only
  • otu_table.csv — OTU table
  • pairwise_distance_summary_intra-class.csv — Intra-class distance summary (with --embeddings)
  • UPGMA_tree_partitions_annotated.pdf — Annotated UPGMA tree (with --embeddings)
  • annotation_summary.json — Correction summary

Diversity Command

Compute community alpha diversity indices. Supports two input modes: partition assignment CSV or OTU table CSV. For Faith's PD, the recommended path is to provide --embeddings, which builds an OTU-centroid neighbor-joining (NJ) tree internally. The older --tree input remains available as a legacy fallback.

# From partition assignments
otuformer diversity \
    --assignments runs/cluster/UPGMA/partitions/tables/partition_0.30_assignments.csv \
    --out-dir runs/diversity

# With Faith's PD from OTU-centroid NJ (recommended)
otuformer diversity \
    --assignments runs/cluster/UPGMA/partitions/tables/partition_0.30_assignments.csv \
    --phylo \
    --embeddings runs/extract/embeddings.csv \
    --save-nj-tree \
    --nj-bootstrap 100 \
    --out-dir runs/diversity

# Legacy Faith's PD path from an existing Newick tree
otuformer diversity \
    --assignments runs/cluster/UPGMA/partitions/tables/partition_0.30_assignments.csv \
    --phylo \
    --tree runs/cluster/UPGMA/UPGMA_Cosine.nwk \
    --out-dir runs/diversity

# From OTU table
otuformer diversity \
    --otu-table-csv otu_table.csv \
    --out-dir runs/diversity
Parameter Description Default
--assignments Partition assignment CSV (id/image, cluster, optional sample) One of two
--otu-table-csv OTU table CSV (sample column + OTU ID headers) One of two
--otu-table-has-header Force first row as header (required if OTU IDs are numeric) No
--out-dir Output directory runs/diversity
--min-abundance Minimum abundance thresholds (comma-separated) 0,2,5
--phylo Enable Faith's PD / PD-related metrics No
--embeddings Embeddings CSV; recommended for PD because OTU-centroid NJ is built internally None
--tree Legacy Newick tree path used only when --embeddings is absent None
--save-nj-tree Save the inferred OTU-centroid NJ tree to NJ_OTU.nwk No
--nj-bootstrap Bootstrap replicates for NJ support; writes NJ_OTU_bootstrap.nwk 0
--save-nj-centroids Save OTU centroid embeddings to NJ_OTU_centroids.csv No

Output metrics (saved to diversity_indices.csv, global and per-sample):

  • Richness: Number of unique OTUs
  • Chao1: Estimated richness (accounts for rare OTUs)
  • ACE: Abundance-based coverage estimator
  • Shannon: Entropy-based diversity (higher = more diverse)
  • Simpson: Probability two individuals differ (higher = more diverse)
  • Hill q0/q1/q2: Hill numbers (richness/evenness/diversity at orders 0,1,2)
  • Pielou_J: Evenness (Shannon / log(richness))
  • Faith's PD (MPD): Morphological phylogenetic diversity computed from the OTU-centroid NJ tree (or legacy --tree input)
  • MPD_w: Abundance-weighted rooted PD (rPD_w)
  • PD_richness_norm: Faith's PD / species richness

Output:

  • diversity_indices.csv — Global diversity indices
  • per-sample/ — Per-sample diversity files (when valid sample column exists)
  • NJ_OTU.nwk — OTU-centroid NJ tree (with --embeddings --save-nj-tree)
  • NJ_OTU_bootstrap.nwk — NJ tree with bootstrap support labels (with --nj-bootstrap > 0)
  • NJ_OTU_centroids.csv — OTU centroid embeddings (with --save-nj-centroids)

CAM Command

Generate CAM heatmaps for model interpretability analysis.

# Basic usage
otuformer cam \
    --checkpoint runs/finetune/finetune_latest.pth \
    --images-dir ./images \
    --out-dir runs/cam

# Use Grad-CAM++ and save raw arrays
otuformer cam \
    --checkpoint runs/finetune/finetune_latest.pth \
    --images-dir ./images \
    --cam-method gradcampp \
    --save-npy \
    --out-dir runs/cam

# Inspect model layer structure first
otuformer cam \
    --checkpoint runs/finetune/finetune_latest.pth \
    --images-dir ./images \
    --dump-model-structure \
    --out-dir runs/cam
Parameter Description Default
--checkpoint OTU-Former checkpoint path (.ckpt or .pth) Required
--images-dir Directory of images to generate CAM for Required
--label-csv Optional CSV with image and label columns; if omitted, all images in --images-dir are used None
--out-dir Output directory runs/cam
--cam-method gradcam/gradcampp/layercam/scorecam/eigencam/ablationcam gradcam
--arch Force architecture: cnn/vit (auto-detected if omitted) Auto
--target-layer-name Specific CAM target layer (auto-selected if omitted) Auto
--image-weight Blend weight of original image in CAM overlay (0-1) 0.5
--fig-format Output format: png/jpg/pdf png
--save-npy Save raw CAM arrays as NumPy files No
--dump-model-structure Write model layer names to model_layers.txt No
--max-images Maximum images to process (None = all) All
--cam-batch-size Batch size for CAM inference 32
--num-workers Dataloader worker processes 4
--device auto/cpu/cuda/mps auto

Output:

  • figures/ — CAM overlay images
  • cam_summary.csv — Metadata summary
  • arrays/ — Raw CAM arrays (with --save-npy)
  • model_layers.txt — Model layer names (with --dump-model-structure)

Export Command

Export a PyTorch checkpoint to ONNX format for deployment and accelerated inference.

# Basic usage
otuformer export \
    --checkpoint runs/finetune/finetune_latest.pth \
    --out-dir runs/export

# Custom image size and opset
otuformer export \
    --checkpoint runs/finetune/finetune_latest.pth \
    --imgsz 224 \
    --opset 17 \
    --out-dir runs/export
Parameter Description Default
--checkpoint Checkpoint path Required
--out-dir Output directory runs/export
--imgsz Input image size for ONNX export (auto-inferred from backbone if not provided) Auto
--opset ONNX opset version 18

Output:

  • encoder.onnx — ONNX encoder model

Common Behaviours

Logging

All commands save log files (e.g. pretrain.log, finetune.log, extract.log, cluster.log, annotate.log, diversity.log, cam.log, export.log) to the logs/ subdirectory of the output directory, containing:

  • Full command line
  • Timestamp
  • All parameter values
  • Runtime output

Graceful Shutdown

Press Ctrl+C — the current image finishes before exiting; partial results are saved.

Device Selection

--device auto chooses automatically:

  1. CUDA (if available)
  2. MPS / Apple Silicon (if available)
  3. CPU (fallback)

Shell Completion

Install shell completion for otuformer:

otuformer --install-completion

Supported shells: bash, zsh, fish

Version

Show installed version:

otuformer --version
otuformer -v

Project Structure

.
├── src/otuformer/            # Core package
│   ├── cli/                  # Command-line interface
│   │   ├── main.py           # Entry point dispatcher
│   │   ├── doctor.py         # otuformer doctor
│   │   ├── pretrain.py       # otuformer pretrain
│   │   ├── finetune.py       # otuformer finetune
│   │   ├── extract.py        # otuformer extract
│   │   ├── cluster.py        # otuformer cluster
│   │   ├── annotate.py       # otuformer annotate
│   │   ├── diversity.py      # otuformer diversity
│   │   ├── cam.py            # otuformer cam
│   │   └── export.py         # otuformer export
│   ├── training/             # Training logic
│   │   ├── model.py          # Model definition
│   │   ├── dataset.py        # Dataset
│   │   ├── loss.py           # Loss functions
│   │   ├── trainer.py        # Trainer
│   │   └── scheduler.py      # Learning rate scheduler
│   ├── embedding/            # Embedding extraction
│   │   ├── extractor.py      # Embedding extractor
│   │   └── evaluator.py      # Embedding evaluator
│   ├── delineation/          # OTU delineation
│   │   ├── distance.py       # Distance computation
│   │   ├── tree.py           # UPGMA tree building
│   │   ├── partition.py      # Partitioning and metrics
│   │   └── annotate.py       # Correction and annotation
│   ├── vision/               # Vision utilities
│   │   ├── cam.py            # CAM visualization
│   │   └── export.py         # ONNX export
│   └── utils/                # Utility functions
│       ├── device.py         # Device detection
│       ├── logging.py        # TeeLogger
│       ├── checkpoint.py     # Checkpoint handling
│       └── io.py             # CSV/JSON I/O
├── tests/                    # Test files
├── examples/                 # Sample data assets
├── docs/                     # Documentation
└── pyproject.toml            # Project configuration

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contact

  • Email: xtmtd.zf@gmail.com

Citation

If you use OTU-Former in your research, please cite:

@software{otuformer2026,
  author = {Zhang, Feng},
  title = {OTU-Former: Image-based Morphological OTU Delineation Toolkit},
  year = {2026},
  url = {https://github.com/xtmtd/OTU-Former}
}

A preprint is available at: https://doi.org/10.64898/2026.04.28.721370

About

An image-based morphological OTU delineation toolkit.

Resources

Readme

License

MIT license
Activity

Stars

2 stars

Watchers

0 watching

Forks

1 fork
Report repository

Releases 2

OTU-Former v0.1.1 Latest
May 17, 2026
+ 1 release

Packages

 
 
 

Contributors

Languages