Skip to main content

Large Scale 3d Convolution Net Inference

Project description

chunkflow

Build Status PyPI version Coverage Status

Chunk operations for large scale 3D image dataset processing

Introduction

3D image dataset could be too large to be processed in a single computer, and distributed processing was required. In most cases, the image dataset could be choped to chunks and distributed to computers for processing. This package provide a framework to perform distributed chunk processing.

Features

  • Decoupled frontend and backend. The computational heavy backend could be any computer with internet connection and Amazon Web Services (AWS) authentication.
  • Composable Commandline interface. The chunk operators could be freely composed in commandline for flexible usage. This is also super useful for tests and experiments.

Usage

Installation

This package was registered in PyPi, just run a simple command to install:

pip install chunkflow

Get Help

chunkflow --help

get help for commands: chunkflow command --help

Examples

The commands could be composed and used flexiblly. The first command should be a generator though.

chunkflow create-chunk view
chunkflow create-chunk 

A Typical pipeline to run ConvNet inference is something like:

chunkflow --verbose fetch-task --queue-name="$QUEUE_NAME" --visibility-timeout=$VISIBILITY_TIMEOUT cutout --volume-path="$IMAGE_LAYER_PATH" --expand-margin-size 4 64 64 inference --convnet-model=your-model-name --convnet-weight-path=path/of/net/weight --patch-size 20 256 256 --patch-overlap 4 64 64 --output-key your-output-key --framework='identity' --batch-size 2 crop-margin save --volume-path="$OUTPUT_LAYER_PATH" --upload-log --nproc 4 --create-thumbnail delete-task-in-queue

Some Typical Operators

  • Convolutional Network Inference. Currently, we support PyTorch and pznet
  • Task Generator. Fetch task from AWS SQS.
  • Cutout service. Cutout chunk from datasets formatted as neuroglancer precomputed using cloudvolume
  • Save. Save chunk to neuroglancer precomputed.
  • Save Images. Save chunk as a serials of PNG images in local directory.
  • Real File. Read image from hdf5 and tiff files.
  • View. View chunk using cloudvolume viewer.
  • Mask. Mask out the chunk using a precomputed dataset.
  • Cloud Watch. Realtime speedometer using AWS CloudWatch.

Use specific GPU device

We can simply set an environment variable to use specific GPU device.

CUDA_VISIBLE_DEVICES=2 chunkflow

Produce tasks to AWS SQS queue

in bin,

python produce_tasks.py --help

Terminology

  • patch: ndarray as input to ConvNet. Normally it is pretty small due to the limited memory capacity of GPU.
  • chunk: ndarray with global offset and arbitrary shape.
  • block: the array with a shape and global offset aligned with storage backend. The block could be saved directly to storage backend. The alignment with storage files ensures that there is no writting conflict when saved parallelly.

Development

Create a new release in PyPi

python setup.py bdist_wheel --universal
twine upload dist/my-new-wheel

Add a new operator

To be added.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

chunkflow-0.2.8.tar.gz (29.6 kB view hashes)

Uploaded Source

Built Distribution

chunkflow-0.2.8-py2.py3-none-any.whl (14.2 kB view hashes)

Uploaded Python 2 Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page