RI2FL

This repository contains PyTorch implementation of RI2FL.

System requirements and installation guide

Installation should be fairly quick (typically less than an hour). On a computer with CUDA-compatible GPUs and Linux operating system (e.g., Ubuntu 16.04), install CUDA/cuDNN and Python 3 (tested on 3.8) with the following command:

➜ pip install -r requirements.txt

Demo and instructions for use

YAML script

RI2FL class parses arguments from the yaml_file.

Exmaple of yaml_file:

# scripts/infer.yaml
path:
    dataset: testset
    model_path: models
    save_path: result

setup:
    fl_list: mem, nuc, oli
    batch_size: 4
    gpus: 0,1,2,3
    cpus: 4
    zoomed_size: [512, 512]
    patch_size: 256
    cropped_depth: 64
    num_drop: 0
    num_tta: 4

Inference

# example.py
import argparse
from ri2fl import Ri2Fl
import torch.distributed as dist


argparser = argparse.ArgumentParser()
argparser.add_argument("yaml")
argparser.add_argument("--local_rank", default=0, type=int)
cmd_args = argparser.parse_args()
ri2fl = Ri2Fl(f"{cmd_args.yaml}.yaml", cmd_args)
ri2fl.predict_all()
dist.destroy_process_group()

Then, run the python script with the following command as bellow.

➜ python -m torch.distributed.launch --nproc_per_node=4 example.py infer

The demo data in the testset folder will output FL tomograms inferred from the input RI tomograms. In order to run RI2FL with your own data, organize your RI tomogram in this format and repeat the procedures above. Simple statistical analyses of the input/output tomograms could reproduce the results in the manuscript. Run time depends on data size and hardware; for a full-sized tomogram, it is expected to take less than a minute with a NVIDIA V100 GPU.

Timelapse Example

Training Data Samples

The samples for training a model are now available (Link).

You can build your own model and train it using samples.

➜ python -m torch.distributed.launch --nproc_per_node=4 train.py --data_root trainset --fl_type nuc --save_path outs/nuc