def process(self, batch, request):
components = batch.volumes[VolumeType.GT_LABELS].data
dtype = components.dtype
simple_neighborhood = malis.mknhood3d()
affinities_from_components = malis.seg_to_affgraph(
components, simple_neighborhood)
components, _ = malis.connected_components_affgraph(
affinities_from_components, simple_neighborhood)
batch.volumes[VolumeType.GT_LABELS].data = components.astype(dtype)
def process(self, batch, request):
components = batch.arrays[self.labels].data
dtype = components.dtype
simple_neighborhood = malis.mknhood3d()
affinities_from_components = malis.seg_to_affgraph(
components, simple_neighborhood)
components, _ = malis.connected_components_affgraph(
affinities_from_components, simple_neighborhood)
batch.arrays[self.labels].data = components.astype(dtype)
def train():
#define many values needed for training
#creates some sort of offset, according to documentation "nhood is just the offset vector that the edge corresponds to"
affinity_neighborhood = malis.mknhood3d()
#Initializes Cafee solver parameters object with atributes filled below, probably should not mess with these without
#good reason
solver_parameters = SolverParameters()
solver_parameters.train_net = 'net.prototxt'
solver_parameters.base_lr = 1e-4
solver_parameters.momentum = 0.95
solver_parameters.momentum2 = 0.999
solver_parameters.delta = 1e-8
solver_parameters.weight_decay = 0.000005
solver_parameters.lr_policy = 'inv'
solver_parameters.gamma = 0.0001
solver_parameters.power = 0.75
solver_parameters.snapshot = 2000
solver_parameters.snapshot_prefix = 'net'
solver_parameters.type = 'Adam'
solver_parameters.resume_from = None
solver_parameters.train_state.add_stage('euclid')
#batch request states what is to be contained in the batch that is requested
request = BatchRequest()
request.add_array_request(ArrayKeys.RAW, (84, 268, 268))
request.add_array_request(ArrayKeys.GT_LABELS, (56, 56, 56))
request.add_array_request(ArrayKeys.GT_MASK, (56, 56, 56))
request.add_array_request(ArrayKeys.GT_IGNORE, (56, 56, 56))
request.add_array_request(ArrayKeys.GT_AFFINITIES, (56, 56, 56))
#import data, normalize it, pick random volumes within it
data_sources = tuple(
Hdf5Source('sample_' + s +
'_padded_20160501.aligned.filled.cropped.hdf',
datasets={
ArrayKeys.RAW: 'volumes/raw',
ArrayKeys.GT_LABELS:
'volumes/labels/neuron_ids_notransparency',
ArrayKeys.GT_MASK: 'volumes/labels/mask',
}) + Normalize() + RandomLocation()
#not sure what this line below does, somehow I think it defines how the random location does its sampling
for s in ['A', 'B', 'C'])
#atrifcat source, not sure whats being imported but after its inmported its randomly sampled volumetrically, normalized,
#randomly augmented for intesity, elasticity and randomly mirrored and transposed
artifact_source = (
Hdf5Source('sample_ABC_padded_20160501.defects.hdf',
datasets={
ArrayKeys.RAW: 'defect_sections/raw',
ArrayKeys.ALPHA_MASK: 'defect_sections/mask',
}) +
RandomLocation(min_masked=0.05, mask_array_key=ArrayKeys.ALPHA_MASK) +
Normalize() +
IntensityAugment(0.9, 1.1, -0.1, 0.1, z_section_wise=True) +
ElasticAugment([4, 40, 40], [0, 2, 2], [0, math.pi / 2.0]) +
SimpleAugment(transpose_only_xy=True))
#Creates snapshots to be saved that contain LOSS_GRADIENT, this will be pulled from upstream provider
snapshot_request = BatchRequest()
snapshot_request.add_array_request(ArrayKeys.LOSS_GRADIENT, (56, 56, 56))
# creates "directed acyclic graph" DAG with each term being a source or a batch provider, the first temp is the source
# "data_soruces" and the remaining terms are batch_providers, when they get the request they send it up stream to the
# source and eventually return it
batch_provider_tree = (
data_sources + RandomProvider() +
ExcludeLabels([8094], ignore_mask_erode=12) +
ElasticAugment([4, 40, 40], [0, 2, 2], [0, math.pi / 2.0],
prob_slip=0.05,
prob_shift=0.05,
max_misalign=25) +
SimpleAugment(transpose_only_xy=True) +
GrowBoundary(steps=3, only_xy=True) +
AddGtAffinities(affinity_neighborhood) +
SplitAndRenumberSegmentationLabels() +
IntensityAugment(0.9, 1.1, -0.1, 0.1, z_section_wise=True) +
DefectAugment(prob_missing=0.03,
prob_low_contrast=0.01,
prob_artifact=0.03,
artifact_source=artifact_source,
contrast_scale=0.1) + ZeroOutConstSections() +
IntensityScaleShift(2, -1) + BalanceAffinityLabels() +
PreCache(cache_size=10, num_workers=5) +
Train(solver_parameters, use_gpu=0) +
Snapshot(every=10,
output_filename='batch_{id}.hdf',
additional_request=snapshot_request))
n = 10
print("Training for", n, "iterations")
#batch request is a training itteration, build takes the function created for batch provider tree and initializes things
with build(batch_provider_tree) as minibatch_maker:
for i in range(n):
minibatch_maker.request_batch(request)
print("Finished")
def train():
affinity_neighborhood = malis.mknhood3d()
solver_parameters = SolverParameters()
solver_parameters.train_net = 'net.prototxt'
solver_parameters.base_lr = 1e-4
solver_parameters.momentum = 0.95
solver_parameters.momentum2 = 0.999
solver_parameters.delta = 1e-8
solver_parameters.weight_decay = 0.000005
solver_parameters.lr_policy = 'inv'
solver_parameters.gamma = 0.0001
solver_parameters.power = 0.75
solver_parameters.snapshot = 2000
solver_parameters.snapshot_prefix = 'net'
solver_parameters.type = 'Adam'
solver_parameters.resume_from = None
solver_parameters.train_state.add_stage('euclid')
request = BatchRequest()
request.add_array_request(ArrayKeys.RAW, (84, 268, 268))
request.add_array_request(ArrayKeys.GT_LABELS, (56, 56, 56))
request.add_array_request(ArrayKeys.GT_MASK, (56, 56, 56))
request.add_array_request(ArrayKeys.GT_IGNORE, (56, 56, 56))
request.add_array_request(ArrayKeys.GT_AFFINITIES, (56, 56, 56))
data_sources = tuple(
Hdf5Source('sample_' + s +
'_padded_20160501.aligned.filled.cropped.hdf',
datasets={
ArrayKeys.RAW: 'volumes/raw',
ArrayKeys.GT_LABELS:
'volumes/labels/neuron_ids_notransparency',
ArrayKeys.GT_MASK: 'volumes/labels/mask',
}) + Normalize() + RandomLocation()
for s in ['A', 'B', 'C'])
artifact_source = (
Hdf5Source('sample_ABC_padded_20160501.defects.hdf',
datasets={
ArrayKeys.RAW: 'defect_sections/raw',
ArrayKeys.ALPHA_MASK: 'defect_sections/mask',
}) +
RandomLocation(min_masked=0.05, mask_array_key=ArrayKeys.ALPHA_MASK) +
Normalize() +
IntensityAugment(0.9, 1.1, -0.1, 0.1, z_section_wise=True) +
ElasticAugment([4, 40, 40], [0, 2, 2], [0, math.pi / 2.0]) +
SimpleAugment(transpose_only_xy=True))
snapshot_request = BatchRequest()
snapshot_request.add_array_request(ArrayKeys.LOSS_GRADIENT, (56, 56, 56))
batch_provider_tree = (
data_sources + RandomProvider() +
ExcludeLabels([8094], ignore_mask_erode=12) +
ElasticAugment([4, 40, 40], [0, 2, 2], [0, math.pi / 2.0],
prob_slip=0.05,
prob_shift=0.05,
max_misalign=25) +
SimpleAugment(transpose_only_xy=True) +
GrowBoundary(steps=3, only_xy=True) +
AddGtAffinities(affinity_neighborhood) +
SplitAndRenumberSegmentationLabels() +
IntensityAugment(0.9, 1.1, -0.1, 0.1, z_section_wise=True) +
DefectAugment(prob_missing=0.03,
prob_low_contrast=0.01,
prob_artifact=0.03,
artifact_source=artifact_source,
contrast_scale=0.1) + ZeroOutConstSections() +
IntensityScaleShift(2, -1) + BalanceAffinityLabels() +
PreCache(cache_size=10, num_workers=5) +
Train(solver_parameters, use_gpu=0) +
Snapshot(every=10,
output_filename='batch_{id}.hdf',
additional_request=snapshot_request))
n = 10
print("Training for", n, "iterations")
with build(batch_provider_tree) as minibatch_maker:
for i in range(n):
minibatch_maker.request_batch(request)
print("Finished")
def train():
affinity_neighborhood = malis.mknhood3d()
solver_parameters = SolverParameters()
solver_parameters.train_net = 'net.prototxt'
solver_parameters.base_lr = 1e-4
solver_parameters.momentum = 0.95
solver_parameters.momentum2 = 0.999
solver_parameters.delta = 1e-8
solver_parameters.weight_decay = 0.000005
solver_parameters.lr_policy = 'inv'
solver_parameters.gamma = 0.0001
solver_parameters.power = 0.75
solver_parameters.snapshot = 2000
solver_parameters.snapshot_prefix = 'net'
solver_parameters.type = 'Adam'
solver_parameters.resume_from = None
solver_parameters.train_state.add_stage('euclid')
request = BatchRequest()
request.add_volume_request(VolumeTypes.RAW, (84,268,268))
request.add_volume_request(VolumeTypes.GT_LABELS, (56,56,56))
request.add_volume_request(VolumeTypes.GT_AFFINITIES, (56,56,56))
request.add_volume_request(VolumeTypes.LOSS_SCALE, (56,56,56))
data_sources = tuple([
Hdf5Source(
'/groups/turaga/home/moreheadm/code/gunpowder/examples/vcn/VCN_train1.hdf5',
datasets = {
VolumeTypes.RAW: 'volume/raw',
VolumeTypes.GT_LABELS: 'volume/labels',
}
) +
Normalize() +
RandomLocation()
])
batch_provider_tree = (
data_sources +
RandomProvider() +
ElasticAugment([4,40,40], [0,2,2], [0,math.pi/2.0], prob_slip=0.05,prob_shift=0.05,max_misalign=25) +
SimpleAugment(transpose_only_xy=True) +
GrowBoundary(steps=3, only_xy=True) +
AddGtAffinities(affinity_neighborhood) +
SplitAndRenumberSegmentationLabels() +
IntensityAugment(0.9, 1.1, -0.1, 0.1, z_section_wise=True) +
DefectAugment(
prob_missing=0.03,
prob_low_contrast=0.01,
contrast_scale=0.1) +
ZeroOutConstSections() +
IntensityScaleShift(2,-1) +
BalanceAffinityLabels() +
PreCache(
request,
cache_size=10,
num_workers=5) +
Train(solver_parameters, use_gpu=0) +
Snapshot(every=10, output_filename='batch_{id}.hdf')
)
n = 10
print("Training for", n, "iterations")
with build(batch_provider_tree) as minibatch_maker:
for i in range(n):
minibatch_maker.request_batch(request)
print("Finished")
