Example 03: Train a network for segmentation

This script trains a MS-D network for segmentation (i.e. labeling) Run generatedata.py first to generate required training data.

# Import code
import msdnet
import glob

# Define dilations in [1,10] as in paper.
dilations = msdnet.dilations.IncrementDilations(10)

# Create main network object for segmentation, with 100 layers,
# [1,10] dilations, 1 input channel, 5 output channels (one for each label),
# using the GPU (set gpu=False to use CPU)
n = msdnet.network.SegmentationMSDNet(100, dilations, 1, 5, gpu=True)

# Initialize network parameters
n.initialize()

# Define training data
# First, create lists of input files (noisy) and target files (labels)
flsin = sorted(glob.glob('train/noisy/*.tiff'))
flstg = sorted(glob.glob('train/label/*.tiff'))
# Create list of datapoints (i.e. input/target pairs)
dats = []
for i in range(len(flsin)):
    # Create datapoint with file names
    d = msdnet.data.ImageFileDataPoint(flsin[i],flstg[i])
    # Convert datapoint to one-hot, using labels 0, 1, 2, 3, and 4,
    # which are the labels given in each label TIFF file.
    d_oh = msdnet.data.OneHotDataPoint(d, [0,1,2,3,4])
    # Augment data by rotating and flipping
    d_augm = msdnet.data.RotateAndFlipDataPoint(d_oh)
    # Add augmented datapoint to list
    dats.append(d_augm)
# Note: The above can also be achieved using a utility function for such 'simple' cases:
# dats = msdnet.utils.load_simple_data('train/noisy/*.tiff', 'train/label/*.tiff', augment=True, labels=[0,1,2,3,4])

# Normalize input and output of network to zero mean and unit variance using
# training data images
n.normalizeinout(dats)

# Use image batches of a single image
bprov = msdnet.data.BatchProvider(dats,1)

# Define validation data (not using augmentation)
flsin = sorted(glob.glob('val/noisy/*.tiff'))
flstg = sorted(glob.glob('val/label/*.tiff'))
datsv = []
for i in range(len(flsin)):
    d = msdnet.data.ImageFileDataPoint(flsin[i],flstg[i])
    d_oh = msdnet.data.OneHotDataPoint(d, [0,1,2,3,4])
    datsv.append(d_oh)
# Note: The above can also be achieved using a utility function for such 'simple' cases:
# datsv = msdnet.utils.load_simple_data('train/noisy/*.tiff', 'train/label/*.tiff', augment=False, labels=[0,1,2,3,4])

# Validate with Mean-Squared Error
val = msdnet.validate.MSEValidation(datsv)

# Use ADAM training algorithms
t = msdnet.train.AdamAlgorithm(n)

# Log error metrics to console
consolelog = msdnet.loggers.ConsoleLogger()
# Log error metrics to file
filelog = msdnet.loggers.FileLogger('log_segm.txt')
# Log typical, worst, and best images to image files
imagelog = msdnet.loggers.ImageLabelLogger('log_segm', onlyifbetter=True)
# Log typical, worst, and best images to image files
# Output probability map for a single channel (in this case, channel 3)
singlechannellog = msdnet.loggers.ImageLogger('log_segm_singlechannel', chan_out=3, onlyifbetter=True)

# Train network until program is stopped manually
# Network parameters are saved in segm_params.h5
# Validation is run after every len(datsv) (=25)
# training steps.
msdnet.train.train(n, t, val, bprov, 'segm_params.h5',loggers=[consolelog,filelog,imagelog,singlechannellog], val_every=len(datsv))