Training models

Training data

Training data should be organised as follows:

set1/
  phase/
    0001_phase.tif
    0002_phase.tif
    ...
  gfp/
  rfp/
  ...
  label/
    0001_mask.tif
    0002_mask.tif
    ...
  weights/
    0001_weights.tif
    0002_weights.tif
    ...
set2/

Use the following scripts to generate weigthmaps and TFRecord files...

$ python weightmap.py
$ python generate_records.py

Training weights

Per-pixel weighting can be applied during training to emphasise certain regions of images important for accurate segmentation. Below is an example of higher weighting in regions with very few pixels separating objects:

Object labels and pre-calculated weight maps

Typical job files

Train a classifier

module = classifier
func = SERVER_train_classifier
device = GPU
params = {'path':'/media/arl/DataII/Data/competition/training/CNN_Training_Anna/png',
 'training_data': 'cell_cycle_CNN.tfrecord',
 'name':'CNN_competition',
 'num_inputs':2,
 'num_epochs':1000,
 'batch_size':32}

Run a classifier prediction:

module = classifier
func = SERVER_classify_from_HDF
device = GPU
params = {'path': '/media/arl/DataII/Data/competition/colony',
 'image_dict': {'brightfield': 'brightfield/BF_pos11.tif',
                'gfp':'fluorescence/GFP_pos12.tif',
                'rfp':'fluorescence/RFP_pos11.tif'},
 'name':'CNN_competition'}

Train a UNet

module = unet2d
func = SERVER_train_unet2d
device = GPU
params = {'path': '/media/arl/DataII/Data/competition/training',
 'training_data': 'train_competition_UNet_w0-30.00_sigma-3.00.tfrecord',
 'name':'UNet2D_test',
 'shape':(1200,1600),
 'num_inputs':1,
 'num_outputs':2,
 'num_epochs':1000,
 'batch_size':1,
 'warm_start':False,
 'bridge':'concat'}

Run a UNet prediction:

module = unet2d
func = SERVER_predict_unet2d
device = GPU
params = {'path': '/media/arl/DataII/Data/competition/colony',
 'image_dict': {'gfp':'fluorescence/GFP_pos12.tif',
                'rfp':'fluorescence/RFP_pos11.tif'},
 'name':'UNet2D_competition',
 'shape':(1200,1600),
 'num_inputs':2,
 'num_outputs':3}

Track cells