Installation

Make sure cuda-10-1 and cudnn are installed. They are needed to create the detectron2 data sets.

pip3 install -U torch==1.7.0 torchvision==0.8.1 detectron2 scikit-image u-msgpack-python -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/index.html
pip3 install .../synthetic-yeast-cells/

Usage

Minimal example

notebook

A minimal run to create synthetic brightfield images of yeast cells is by only specifying how many samples should be generated. The other parameters are considered at their defaults.

from syntheticyeastcells import create_samples

images, labels = create_samples(100)

Example with all parameters

notebook

Alternatively you can specify these parameters to tweak the properties of the images. This example shows the default values

images, labels = create_samples(
  # number of samples to create:
  n_images=len(axes),

  # number of cells per image
  n_cells_per_image=100,

  # image dimensions (height, width)
  size=(512, 512),

  # area at the boundary to not have cells
  min_distance_boundary=50,

  # range for the first radius of the ellipse as
  # an approximate of a cell, from which is sampled
  # uniformly
  r0_range=(2, 14),

  # range for the second radius of the ellipse as
  # an approximate of a cell as a factor of the first
  # radius, from which is sampled uniformly.
  r1_factor_range=(0.7, 1.3),

  # the background consists of gaussian noise, blurred by this
  # standard deviation.
  spatial_blur_std=1.5,

  # expected background greyscale, assuming a [0, 1] range
  background_intensity=0.4,

  # standard deviation of the noise outside of cells
  background_contrast=0.00188,

  # standard deviation of the noise inside cells
  core_contrast=0.0282,

  # probability the outside of a cell is white, and the
  # inside is white, instead of the other way around.
  p_white_outside=0.5
)

Example outputs for all parameters

Output

images is a $n_images \times height \times width \times 3$ greyscale stack of images encoded as uint8, and labels is a $n_images \times height \times width$ stack of labels encoded as uint32. Each cell is encoded as a different integer and the background is encoded as 0.

from matplotlib import pyplot

axes = pyplot.subplots(3, 3, figsize=(30, 30))[1].ravel()

# Change the labels to rgba colors
colors = pyplot.get_cmap('hsv')(numpy.linspace(0, 1, labels.max() + 1))
colors[0] = [0, 0, 0, 0] # background is fully transparent.
colors[1:, 3] = 0.2
colors = (colors[labels] * 255).astype(numpy.uint8)

for axis, image, label in zip(axes, images, colors):
  axis.imshow(image)
  axis.imshow(label)

Create a dataset suitable for detectron2

notebook

from syntheticyeastcells.detectron2 import create_dataset

synthetic_data_folder = (
  '/slow-data/p253591/yeast cells/data/synthetic/v7')

create_dataset(
  # Where to store the data set
  synthetic_data_folder,

  # sizesof several subsets
  sets={'test': 1000, 'val': 1000, 'train': 20000},

  # same parameters as before
  n_cells_per_image=100,
  size=(512, 512),
  min_distance_boundary=50,
  r0_range=(2, 14),
  r1_factor_range=(0.7, 1.3),
  spatial_blur_std=1.5,
  background_intensity=0.4,
  background_contrast=0.00188,
  core_contrast=0.0752,
  p_white_outside=0.5,

  # number of processes to run in parallel, number of samples created in one batch
  njobs=40, batch_size=1,

  # show a progressbar
  progressbar=True)

with open(f'{synthetic_data_folder}/labels.json', 'w') as f:
  json.dump(labels, f)

This results in these files:

ls '{synthetic_data_folder}/'*    
/slow-data/p253591/yeast cells/data/synthetic/v8/labels.json

/slow-data/p253591/yeast cells/data/synthetic/v8/test:
image-0.jpg  image-1.jpg  image-2.jpg

/slow-data/p253591/yeast cells/data/synthetic/v8/train:
image-0.jpg  image-2.jpg  image-4.jpg  image-6.jpg  image-8.jpg
image-1.jpg  image-3.jpg  image-5.jpg  image-7.jpg  image-9.jpg

/slow-data/p253591/yeast cells/data/synthetic/v8/val:
image-0.jpg  image-1.jpg  image-2.jpg


cat '{synthetic_data_folder}/labels.json'
{"test": [{"height": 512, "width": 512, "annotations": [{"bbox": [421.5, 139.5, 434.5, 154.5], "bbox_mode": 0, "segmentation": [[ ... ]], "category_id": 0, "iscrowd": 0}, ...., "valid": ...., "train": ...}