TorchVF

TorchVF is a unifying Python library for using vector fields for efficient proposal-free instance segmentation. Vector field based methods are lightweight, fast to train, and can accurately segment objects with arbitrary morphology and population density. Read more about vector field based methods for instance segmentation in my article. TorchVF provides a vector field agnostic API for ground truth vector field computation, interpolation of discretely sampled vector fields, numeric integration solvers, clustering functions, and various other utilities.

This repository also provides all configs, code, and tools necessary to reproduce the results presented in my article on vector field based methods.

Installation

TorchVF can be install via pip:

pip install torchvf

For the most up-to-date version, you could install directly from GitHub (this is not recommended):

pip install git+https://github.com/ryanirl/torchvf.git

Quick Start

For deriving the instance segmentation from the semantic segmentation and vector field, the TorchVF API is centered around 4 functions:

• interp_vf()
• init_values_semantic()
• ivp_solver()
• cluster()

To demonstrate how these functions work, consider we are given a semantic segmentation semantic and vector field vf. TorchVF can be used to compute the instance segmentation of an image via the following code:

from torchvf.numerics import interp_vf, ivp_solver, init_values_semantic
from torchvf.utils import cluster

# Step 1: Convert our discretely sampled vector field into continuous vector
# field through bilinear interpolation. 
vf = interp_vf(vf, mode = "bilinear")

# Step 2. Convert our semantic segmentation `semantic` into a set of
# initial-values to be integrated through our vector field `vf`.
init_values = init_values_semantic(semantic, device = "cuda:0")

# Step 3. Integrate our initial-values `init_values` through our vector field
# `vf` for 25 steps with a step size of 0.1 using Euler's method for numeric 
# integration. 
solutions = ivp_solver(
    vf, 
    init_values, 
    dx = 0.1,
    n_steps = 25,
    solver = "euler"
)[-1] # Get the final solution. 

# Clustering can only be done on the CPU. 
solutions = solutions.cpu()
semantic = semantic.cpu()

# Step 4. Cluster the integrated semantic points `solutions` to obtain the
# instance segmentation. 
instance_segmentation = cluster(
    solutions, 
    semantic[0], 
    eps = 2.25,
    min_samples = 15,
    snap_noise = False
)

Supported Features

Interpolators:

Interpolator	Implemented
Nearest Neighbor	✅
Nearest Neighbor Batched	⬜
Bilinear	✅
Bilinear Batched	✅

Numeric Integration Solvers:

Interpolator	Implemented
Euler's Method	✅
Midpoint Method	✅
Runge Kutta (4th Order)	✅
Adaptive Dormand Prince	⬜

Clustering Schemes:

Interpolator	Implemented
DBSCAN (Scikit-learn)	✅
DCSCAN (PyTorch)	⬜
...?	⬜

Vector Field Computation:

Interpolator	Implemented
Truncated SDF + Kernel	✅
Affinity Derived	✅
Omnipose	⬜
Centroid Based	⬜

Other Utilities:

• Tiler wrapper for models.
• Semantic -> euclidean conversion.
• The IVP vector field loss function.
• Tversky and Dice semantic loss functions.
• Training and evalution scripts.
• Various pretrained models on the BPCIS dataset.
• Modeling for the presented H1 and H2 models.
• mAP IoU, F1, IoU metrics.

Dependencies

The ultimate goal of TorchVF is to be solely dependent on PyTorch. Although at the moment, the signed distance function computation relies on Seung Lab's euclidean distance transform library and the DBSCAN clustering implementation relies on Scikit-learn. Furthermore, NumPy appears in various places (mAP IoU metric, clustering, ...).

Reproducability

This is a reproducability guide for people who want to reproduce the results presented in my article on vector field based methods. First, install the torchvf library and clone the repository to get access to the scripts:

pip install torchvf 

git clone https://github.com/ryanirl/torchvf.git

Installing the Weights

I provide weights for the H1 and H2 models trained on each subset of the BPCIS dataset. These weights, along with configs and logging information for both training and evaluation, can be downloaded here (157.5 MB zipped | 185.5 MB unzipped).

Once you download the weights:

• Unzip the file.
• Replace the torchvf/weights file with the downloaded and unzipped torchvf_weights file.
• Rename torchvf/torchvf_weights to torchvf/weights.

Installing the BPCIS Dataset

Download the BPCIS dataset here. Then setup the file system this way:

├── torchvf/
├── data/
│   └── bpcis/
│       ├── bact_fluor_train/
│       ├── bact_fluor_test/
│       ├── bact_phase_train/
│       ├── bact_phase_test/
│       ├── worm_train/
│       └── worm_test/
├── weights/
└── ***

If you have cloned the library, downloaded the weights, and downloaded the BPCIS dataset you should be able to do python3 scripts/eval.py --config_dir ./weights/bact_fluor/h1/eval_config.py. This will run evaluation on the bacterial fluorescence subset using the evaluation config file provided with the downloaded weights.

Citation

@article{TorchVF,
   author = {Ryan Peters},
   title = {TorchVF: Vector Fields for Instance Segmentation},
   year = 2022
}

License

Distributed under the Apache-2.0 license. See LICENSE for more information.