(min_density, max_density), [cage1], 0.5)
add_channel_dim = gp.Stack(1)
stack = gp.Stack(5)
prepare_data = PrepareTrainingData()
train = gp.torch.Train(model,
loss,
optimizer,
inputs={'input': raw},
loss_inputs={
0: prediction,
1: out_cage_map
},
outputs={0: prediction})
pipeline = (source + normalize + gp.RandomLocation() + simulate_cages +
add_channel_dim + stack + prepare_data +
gp.PreCache(num_workers=40) + train +
gp.PrintProfilingStats(every=1))
print("PIPELINE")
print(pipeline)
# compute a valid input and output size
# for x and y:
#
# 124 -> 120 88 -> 84
# | |
# 60 -> 56 48 -> 44
# | |
# 28 -> 24
#
def create_train_pipeline(self, model):
optimizer = self.params['optimizer'](model.parameters(),
**self.params['optimizer_kwargs'])
filename = self.params['data_file']
datasets = self.params['dataset']
raw_0 = gp.ArrayKey('RAW_0')
points_0 = gp.GraphKey('POINTS_0')
locations_0 = gp.ArrayKey('LOCATIONS_0')
emb_0 = gp.ArrayKey('EMBEDDING_0')
raw_1 = gp.ArrayKey('RAW_1')
points_1 = gp.GraphKey('POINTS_1')
locations_1 = gp.ArrayKey('LOCATIONS_1')
emb_1 = gp.ArrayKey('EMBEDDING_1')
data = daisy.open_ds(filename, datasets[0])
source_roi = gp.Roi(data.roi.get_offset(), data.roi.get_shape())
voxel_size = gp.Coordinate(data.voxel_size)
# Get in and out shape
in_shape = gp.Coordinate(model.in_shape)
out_shape = gp.Coordinate(model.out_shape[2:])
is_2d = in_shape.dims() == 2
emb_voxel_size = voxel_size
cv_loss = ContrastiveVolumeLoss(self.params['temperature'],
self.params['point_density'],
out_shape * voxel_size)
# Add fake 3rd dim
if is_2d:
in_shape = gp.Coordinate((1, *in_shape))
out_shape = gp.Coordinate((1, *out_shape))
voxel_size = gp.Coordinate((1, *voxel_size))
source_roi = gp.Roi((0, *source_roi.get_offset()),
(data.shape[0], *source_roi.get_shape()))
in_shape = in_shape * voxel_size
out_shape = out_shape * voxel_size
logger.info(f"source roi: {source_roi}")
logger.info(f"in_shape: {in_shape}")
logger.info(f"out_shape: {out_shape}")
logger.info(f"voxel_size: {voxel_size}")
request = gp.BatchRequest()
request.add(raw_0, in_shape)
request.add(raw_1, in_shape)
request.add(points_0, out_shape)
request.add(points_1, out_shape)
request[locations_0] = gp.ArraySpec(nonspatial=True)
request[locations_1] = gp.ArraySpec(nonspatial=True)
snapshot_request = gp.BatchRequest()
snapshot_request[emb_0] = gp.ArraySpec(roi=request[points_0].roi)
snapshot_request[emb_1] = gp.ArraySpec(roi=request[points_1].roi)
random_point_generator = RandomPointGenerator(
density=self.params['point_density'], repetitions=2)
# Use volume to calculate probabilities, RandomSourceGenerator will
# normalize volumes to probablilties
probabilities = np.array([
np.product(daisy.open_ds(filename, dataset).shape)
for dataset in datasets
])
random_source_generator = RandomSourceGenerator(
num_sources=len(datasets),
probabilities=probabilities,
repetitions=2)
array_sources = tuple(
tuple(
gp.ZarrSource(
filename,
{raw: dataset},
# fake 3D data
array_specs={
raw:
gp.ArraySpec(roi=source_roi,
voxel_size=voxel_size,
interpolatable=True)
}) for dataset in datasets) for raw in [raw_0, raw_1])
# Choose a random dataset to pull from
array_sources = \
tuple(arrays +
RandomMultiBranchSource(random_source_generator) +
gp.Normalize(raw, self.params['norm_factor']) +
gp.Pad(raw, None)
for raw, arrays
in zip([raw_0, raw_1], array_sources))
point_sources = tuple(
(RandomPointSource(points_0,
random_point_generator=random_point_generator),
RandomPointSource(points_1,
random_point_generator=random_point_generator)))
# Merge the point and array sources together.
# There is one array and point source per branch.
sources = tuple((array_source, point_source) + gp.MergeProvider()
for array_source, point_source in zip(
array_sources, point_sources))
sources = tuple(
self._make_train_augmentation_pipeline(raw, source)
for raw, source in zip([raw_0, raw_1], sources))
pipeline = (sources + gp.MergeProvider() + gp.Crop(raw_0, source_roi) +
gp.Crop(raw_1, source_roi) + gp.RandomLocation() +
PrepareBatch(raw_0, raw_1, points_0, points_1, locations_0,
locations_1, is_2d) +
RejectArray(ensure_nonempty=locations_0) +
RejectArray(ensure_nonempty=locations_1))
if not is_2d:
pipeline = (pipeline + AddChannelDim(raw_0) + AddChannelDim(raw_1))
pipeline = (pipeline + gp.PreCache() + gp.torch.Train(
model,
cv_loss,
optimizer,
inputs={
'raw_0': raw_0,
'raw_1': raw_1
},
loss_inputs={
'emb_0': emb_0,
'emb_1': emb_1,
'locations_0': locations_0,
'locations_1': locations_1
},
outputs={
2: emb_0,
3: emb_1
},
array_specs={
emb_0: gp.ArraySpec(voxel_size=emb_voxel_size),
emb_1: gp.ArraySpec(voxel_size=emb_voxel_size)
},
checkpoint_basename=self.logdir + '/contrastive/checkpoints/model',
save_every=self.params['save_every'],
log_dir=self.logdir + "/contrastive",
log_every=self.log_every))
if is_2d:
pipeline = (
pipeline +
# everything is 3D, except emb_0 and emb_1
AddSpatialDim(emb_0) + AddSpatialDim(emb_1))
pipeline = (
pipeline +
# now everything is 3D
RemoveChannelDim(raw_0) + RemoveChannelDim(raw_1) +
RemoveChannelDim(emb_0) + RemoveChannelDim(emb_1) +
gp.Snapshot(output_dir=self.logdir + '/contrastive/snapshots',
output_filename='it{iteration}.hdf',
dataset_names={
raw_0: 'raw_0',
raw_1: 'raw_1',
locations_0: 'locations_0',
locations_1: 'locations_1',
emb_0: 'emb_0',
emb_1: 'emb_1'
},
additional_request=snapshot_request,
every=self.params['save_every']) +
gp.PrintProfilingStats(every=500))
return pipeline, request
def build_pipeline(parameter, augment=True):
voxel_size = gp.Coordinate(parameter['voxel_size'])
# Array Specifications.
raw = gp.ArrayKey('RAW')
gt_neurons = gp.ArrayKey('GT_NEURONS')
gt_postpre_vectors = gp.ArrayKey('GT_POSTPRE_VECTORS')
gt_post_indicator = gp.ArrayKey('GT_POST_INDICATOR')
post_loss_weight = gp.ArrayKey('POST_LOSS_WEIGHT')
vectors_mask = gp.ArrayKey('VECTORS_MASK')
pred_postpre_vectors = gp.ArrayKey('PRED_POSTPRE_VECTORS')
pred_post_indicator = gp.ArrayKey('PRED_POST_INDICATOR')
grad_syn_indicator = gp.ArrayKey('GRAD_SYN_INDICATOR')
grad_partner_vectors = gp.ArrayKey('GRAD_PARTNER_VECTORS')
# Points specifications
dummypostsyn = gp.PointsKey('DUMMYPOSTSYN')
postsyn = gp.PointsKey('POSTSYN')
presyn = gp.PointsKey('PRESYN')
trg_context = 140 # AddPartnerVectorMap context in nm - pre-post distance
with open('train_net_config.json', 'r') as f:
net_config = json.load(f)
input_size = gp.Coordinate(net_config['input_shape']) * voxel_size
output_size = gp.Coordinate(net_config['output_shape']) * voxel_size
request = gp.BatchRequest()
request.add(raw, input_size)
request.add(gt_neurons, output_size)
request.add(gt_postpre_vectors, output_size)
request.add(gt_post_indicator, output_size)
request.add(post_loss_weight, output_size)
request.add(vectors_mask, output_size)
request.add(dummypostsyn, output_size)
for (key, request_spec) in request.items():
print(key)
print(request_spec.roi)
request_spec.roi.contains(request_spec.roi)
# slkfdms
snapshot_request = gp.BatchRequest({
pred_post_indicator:
request[gt_postpre_vectors],
pred_postpre_vectors:
request[gt_postpre_vectors],
grad_syn_indicator:
request[gt_postpre_vectors],
grad_partner_vectors:
request[gt_postpre_vectors],
vectors_mask:
request[gt_postpre_vectors]
})
postsyn_rastersetting = gp.RasterizationSettings(
radius=parameter['blob_radius'],
mask=gt_neurons,
mode=parameter['blob_mode'])
pipeline = tuple([
create_source(sample, raw, presyn, postsyn, dummypostsyn, parameter,
gt_neurons) for sample in samples
])
pipeline += gp.RandomProvider()
if augment:
pipeline += gp.ElasticAugment([4, 40, 40], [0, 2, 2],
[0, math.pi / 2.0],
prob_slip=0.05,
prob_shift=0.05,
max_misalign=10,
subsample=8)
pipeline += gp.SimpleAugment(transpose_only=[1, 2], mirror_only=[1, 2])
pipeline += gp.IntensityAugment(raw,
0.9,
1.1,
-0.1,
0.1,
z_section_wise=True)
pipeline += gp.IntensityScaleShift(raw, 2, -1)
pipeline += gp.RasterizePoints(
postsyn, gt_post_indicator,
gp.ArraySpec(voxel_size=voxel_size, dtype=np.int32),
postsyn_rastersetting)
spec = gp.ArraySpec(voxel_size=voxel_size)
pipeline += AddPartnerVectorMap(
src_points=postsyn,
trg_points=presyn,
array=gt_postpre_vectors,
radius=parameter['d_blob_radius'],
trg_context=trg_context, # enlarge
array_spec=spec,
mask=gt_neurons,
pointmask=vectors_mask)
pipeline += gp.BalanceLabels(labels=gt_post_indicator,
scales=post_loss_weight,
slab=(-1, -1, -1),
clipmin=parameter['cliprange'][0],
clipmax=parameter['cliprange'][1])
if parameter['d_scale'] != 1:
pipeline += gp.IntensityScaleShift(gt_postpre_vectors,
scale=parameter['d_scale'],
shift=0)
pipeline += gp.PreCache(cache_size=40, num_workers=10)
pipeline += gp.tensorflow.Train(
'./train_net',
optimizer=net_config['optimizer'],
loss=net_config['loss'],
summary=net_config['summary'],
log_dir='./tensorboard/',
save_every=30000, # 10000
log_every=100,
inputs={
net_config['raw']:
raw,
net_config['gt_partner_vectors']:
gt_postpre_vectors,
net_config['gt_syn_indicator']:
gt_post_indicator,
net_config['vectors_mask']:
vectors_mask,
# Loss weights --> mask
net_config['indicator_weight']:
post_loss_weight, # Loss weights
},
outputs={
net_config['pred_partner_vectors']: pred_postpre_vectors,
net_config['pred_syn_indicator']: pred_post_indicator,
},
gradients={
net_config['pred_partner_vectors']: grad_partner_vectors,
net_config['pred_syn_indicator']: grad_syn_indicator,
},
)
# Visualize.
pipeline += gp.IntensityScaleShift(raw, 0.5, 0.5)
pipeline += gp.Snapshot(
{
raw: 'volumes/raw',
gt_neurons: 'volumes/labels/neuron_ids',
gt_post_indicator: 'volumes/gt_post_indicator',
gt_postpre_vectors: 'volumes/gt_postpre_vectors',
pred_postpre_vectors: 'volumes/pred_postpre_vectors',
pred_post_indicator: 'volumes/pred_post_indicator',
post_loss_weight: 'volumes/post_loss_weight',
grad_syn_indicator: 'volumes/post_indicator_gradients',
grad_partner_vectors: 'volumes/partner_vectors_gradients',
vectors_mask: 'volumes/vectors_mask'
},
every=1000,
output_filename='batch_{iteration}.hdf',
compression_type='gzip',
additional_request=snapshot_request)
pipeline += gp.PrintProfilingStats(every=100)
print("Starting training...")
max_iteration = parameter['max_iteration']
with gp.build(pipeline) as b:
for i in range(max_iteration):
b.request_batch(request)
def make_pipeline(self):
raw = gp.ArrayKey('RAW')
pred_affs = gp.ArrayKey('PREDICTIONS')
source_shape = zarr.open(self.data_file)[self.dataset].shape
raw_roi = gp.Roi(np.zeros(len(source_shape[1:])), source_shape[1:])
data = daisy.open_ds(self.data_file, self.dataset)
source_roi = gp.Roi(data.roi.get_offset(), data.roi.get_shape())
voxel_size = gp.Coordinate(data.voxel_size)
# Get in and out shape
in_shape = gp.Coordinate(self.model.in_shape)
out_shape = gp.Coordinate(self.model.out_shape[2:])
is_2d = in_shape.dims() == 2
in_shape = in_shape * voxel_size
out_shape = out_shape * voxel_size
logger.info(f"source roi: {source_roi}")
logger.info(f"in_shape: {in_shape}")
logger.info(f"out_shape: {out_shape}")
logger.info(f"voxel_size: {voxel_size}")
request = gp.BatchRequest()
request.add(raw, in_shape)
request.add(pred_affs, out_shape)
context = (in_shape - out_shape) / 2
source = (gp.ZarrSource(self.data_file, {
raw: self.dataset,
},
array_specs={
raw:
gp.ArraySpec(roi=source_roi,
interpolatable=False)
}))
in_dims = len(self.model.in_shape)
if is_2d:
# 2D: [samples, y, x] or [samples, channels, y, x]
needs_channel_fix = (len(data.shape) - in_dims == 1)
if needs_channel_fix:
source = (source + AddChannelDim(raw, axis=1))
# raw [samples, channels, y, x]
else:
# 3D: [z, y, x] or [channel, z, y, x] or [sample, channel, z, y, x]
needs_channel_fix = (len(data.shape) - in_dims == 0)
needs_batch_fix = (len(data.shape) - in_dims <= 1)
if needs_channel_fix:
source = (source + AddChannelDim(raw, axis=0))
# Batch fix
if needs_batch_fix:
source = (source + AddChannelDim(raw))
# raw: [sample, channels, z, y, x]
with gp.build(source):
raw_roi = source.spec[raw].roi
logger.info(f"raw_roi: {raw_roi}")
pipeline = (source +
gp.Normalize(raw, factor=self.params['norm_factor']) +
gp.Pad(raw, context) + gp.PreCache() + gp.torch.Predict(
self.model,
inputs={'raw': raw},
outputs={0: pred_affs},
array_specs={pred_affs: gp.ArraySpec(roi=raw_roi)}))
pipeline = (pipeline + gp.ZarrWrite({
pred_affs: 'predictions',
},
output_dir=self.curr_log_dir,
output_filename='predictions.zarr',
compression_type='gzip') +
gp.Scan(request))
return pipeline, request, pred_affs
def train_until(**kwargs):
if tf.train.latest_checkpoint(kwargs['output_folder']):
trained_until = int(
tf.train.latest_checkpoint(kwargs['output_folder']).split('_')[-1])
else:
trained_until = 0
if trained_until >= kwargs['max_iteration']:
return
anchor = gp.ArrayKey('ANCHOR')
raw = gp.ArrayKey('RAW')
raw_cropped = gp.ArrayKey('RAW_CROPPED')
gt_threeclass = gp.ArrayKey('GT_THREECLASS')
loss_weights_threeclass = gp.ArrayKey('LOSS_WEIGHTS_THREECLASS')
pred_threeclass = gp.ArrayKey('PRED_THREECLASS')
pred_threeclass_gradients = gp.ArrayKey('PRED_THREECLASS_GRADIENTS')
with open(
os.path.join(kwargs['output_folder'],
kwargs['name'] + '_config.json'), 'r') as f:
net_config = json.load(f)
with open(
os.path.join(kwargs['output_folder'],
kwargs['name'] + '_names.json'), 'r') as f:
net_names = json.load(f)
voxel_size = gp.Coordinate(kwargs['voxel_size'])
input_shape_world = gp.Coordinate(net_config['input_shape']) * voxel_size
output_shape_world = gp.Coordinate(net_config['output_shape']) * voxel_size
# formulate the request for what a batch should (at least) contain
request = gp.BatchRequest()
request.add(raw, input_shape_world)
request.add(raw_cropped, output_shape_world)
request.add(gt_threeclass, output_shape_world)
request.add(anchor, output_shape_world)
request.add(loss_weights_threeclass, output_shape_world)
# when we make a snapshot for inspection (see below), we also want to
# request the predicted affinities and gradients of the loss wrt the
# affinities
snapshot_request = gp.BatchRequest()
snapshot_request.add(raw_cropped, output_shape_world)
snapshot_request.add(gt_threeclass, output_shape_world)
snapshot_request.add(pred_threeclass, output_shape_world)
# snapshot_request.add(pred_threeclass_gradients, output_shape_world)
if kwargs['input_format'] != "hdf" and kwargs['input_format'] != "zarr":
raise NotImplementedError("train node for {} not implemented".format(
kwargs['input_format']))
fls = []
shapes = []
for f in kwargs['data_files']:
fls.append(os.path.splitext(f)[0])
if kwargs['input_format'] == "hdf":
vol = h5py.File(f, 'r')['volumes/raw']
elif kwargs['input_format'] == "zarr":
vol = zarr.open(f, 'r')['volumes/raw']
print(f, vol.shape, vol.dtype)
shapes.append(vol.shape)
if vol.dtype != np.float32:
print("please convert to float32")
ln = len(fls)
print("first 5 files: ", fls[0:4])
# padR = 46
# padGT = 32
if kwargs['input_format'] == "hdf":
sourceNode = gp.Hdf5Source
elif kwargs['input_format'] == "zarr":
sourceNode = gp.ZarrSource
augmentation = kwargs['augmentation']
pipeline = (
tuple(
# read batches from the HDF5 file
sourceNode(
fls[t] + "." + kwargs['input_format'],
datasets={
raw: 'volumes/raw',
gt_threeclass: 'volumes/gt_threeclass',
anchor: 'volumes/gt_threeclass',
},
array_specs={
raw: gp.ArraySpec(interpolatable=True),
gt_threeclass: gp.ArraySpec(interpolatable=False),
anchor: gp.ArraySpec(interpolatable=False)
}
)
+ gp.MergeProvider()
+ gp.Pad(raw, None)
+ gp.Pad(gt_threeclass, None)
+ gp.Pad(anchor, gp.Coordinate((2,2,2)))
# chose a random location for each requested batch
+ gp.RandomLocation()
for t in range(ln)
) +
# chose a random source (i.e., sample) from the above
gp.RandomProvider() +
# elastically deform the batch
(gp.ElasticAugment(
augmentation['elastic']['control_point_spacing'],
augmentation['elastic']['jitter_sigma'],
[augmentation['elastic']['rotation_min']*np.pi/180.0,
augmentation['elastic']['rotation_max']*np.pi/180.0],
subsample=augmentation['elastic'].get('subsample', 1)) \
if augmentation.get('elastic') is not None else NoOp()) +
# apply transpose and mirror augmentations
gp.SimpleAugment(mirror_only=augmentation['simple'].get("mirror"),
transpose_only=augmentation['simple'].get("transpose")) +
# # scale and shift the intensity of the raw array
gp.IntensityAugment(
raw,
scale_min=augmentation['intensity']['scale'][0],
scale_max=augmentation['intensity']['scale'][1],
shift_min=augmentation['intensity']['shift'][0],
shift_max=augmentation['intensity']['shift'][1],
z_section_wise=False) +
# grow a boundary between labels
# TODO: check
# gp.GrowBoundary(
# gt_threeclass,
# steps=1,
# only_xy=False) +
gp.BalanceLabels(
gt_threeclass,
loss_weights_threeclass,
num_classes=3) +
# pre-cache batches from the point upstream
gp.PreCache(
cache_size=kwargs['cache_size'],
num_workers=kwargs['num_workers']) +
# perform one training iteration for each passing batch (here we use
# the tensor names earlier stored in train_net.config)
gp.tensorflow.Train(
os.path.join(kwargs['output_folder'], kwargs['name']),
optimizer=net_names['optimizer'],
summary=net_names['summaries'],
log_dir=kwargs['output_folder'],
loss=net_names['loss'],
inputs={
net_names['raw']: raw,
net_names['anchor']: anchor,
net_names['gt_threeclass']: gt_threeclass,
net_names['loss_weights_threeclass']: loss_weights_threeclass
},
outputs={
net_names['pred_threeclass']: pred_threeclass,
net_names['raw_cropped']: raw_cropped,
},
gradients={
net_names['pred_threeclass']: pred_threeclass_gradients,
},
save_every=kwargs['checkpoints']) +
# save the passing batch as an HDF5 file for inspection
gp.Snapshot(
{
raw: '/volumes/raw',
raw_cropped: 'volumes/raw_cropped',
gt_threeclass: '/volumes/gt_threeclass',
pred_threeclass: '/volumes/pred_threeclass',
},
output_dir=os.path.join(kwargs['output_folder'], 'snapshots'),
output_filename='batch_{iteration}.hdf',
every=kwargs['snapshots'],
additional_request=snapshot_request,
compression_type='gzip') +
# show a summary of time spend in each node every 10 iterations
gp.PrintProfilingStats(every=kwargs['profiling'])
)
#########
# TRAIN #
#########
print("Starting training...")
with gp.build(pipeline):
print(pipeline)
for i in range(trained_until, kwargs['max_iteration']):
# print("request", request)
start = time.time()
pipeline.request_batch(request)
time_of_iteration = time.time() - start
logger.info("Batch: iteration=%d, time=%f", i, time_of_iteration)
# exit()
print("Training finished")
def create_train_pipeline(self, model):
print(
f"Creating training pipeline with batch size {self.params['batch_size']}"
)
filename = self.params['data_file']
raw_dataset = self.params['dataset']['train']['raw']
gt_dataset = self.params['dataset']['train']['gt']
optimizer = self.params['optimizer'](model.parameters(),
**self.params['optimizer_kwargs'])
raw = gp.ArrayKey('RAW')
gt_labels = gp.ArrayKey('LABELS')
gt_aff = gp.ArrayKey('AFFINITIES')
predictions = gp.ArrayKey('PREDICTIONS')
emb = gp.ArrayKey('EMBEDDING')
raw_data = daisy.open_ds(filename, raw_dataset)
source_roi = gp.Roi(raw_data.roi.get_offset(),
raw_data.roi.get_shape())
source_voxel_size = gp.Coordinate(raw_data.voxel_size)
out_voxel_size = gp.Coordinate(raw_data.voxel_size)
# Get in and out shape
in_shape = gp.Coordinate(model.in_shape)
out_shape = gp.Coordinate(model.out_shape[2:])
is_2d = in_shape.dims() == 2
in_shape = in_shape * out_voxel_size
out_shape = out_shape * out_voxel_size
context = (in_shape - out_shape) / 2
gt_labels_out_shape = out_shape
# Add fake 3rd dim
if is_2d:
source_voxel_size = gp.Coordinate((1, *source_voxel_size))
source_roi = gp.Roi((0, *source_roi.get_offset()),
(raw_data.shape[0], *source_roi.get_shape()))
context = gp.Coordinate((0, *context))
aff_neighborhood = [[0, -1, 0], [0, 0, -1]]
gt_labels_out_shape = (1, *gt_labels_out_shape)
else:
aff_neighborhood = [[-1, 0, 0], [0, -1, 0], [0, 0, -1]]
logger.info(f"source roi: {source_roi}")
logger.info(f"in_shape: {in_shape}")
logger.info(f"out_shape: {out_shape}")
logger.info(f"voxel_size: {out_voxel_size}")
logger.info(f"context: {context}")
request = gp.BatchRequest()
request.add(raw, in_shape)
request.add(gt_aff, out_shape)
request.add(predictions, out_shape)
snapshot_request = gp.BatchRequest()
snapshot_request[emb] = gp.ArraySpec(
roi=gp.Roi((0, ) * in_shape.dims(),
gp.Coordinate((*model.base_encoder.out_shape[2:], )) *
out_voxel_size))
snapshot_request[gt_labels] = gp.ArraySpec(
roi=gp.Roi(context, gt_labels_out_shape))
source = (
gp.ZarrSource(filename, {
raw: raw_dataset,
gt_labels: gt_dataset
},
array_specs={
raw:
gp.ArraySpec(roi=source_roi,
voxel_size=source_voxel_size,
interpolatable=True),
gt_labels:
gp.ArraySpec(roi=source_roi,
voxel_size=source_voxel_size)
}) + gp.Normalize(raw, self.params['norm_factor']) +
gp.Pad(raw, context) + gp.Pad(gt_labels, context) +
gp.RandomLocation()
# raw : (l=1, h, w)
# gt_labels: (l=1, h, w)
)
source = self._augmentation_pipeline(raw, source)
pipeline = (
source +
# raw : (l=1, h, w)
# gt_labels: (l=1, h, w)
gp.AddAffinities(aff_neighborhood, gt_labels, gt_aff) +
SetDtype(gt_aff, np.float32) +
# raw : (l=1, h, w)
# gt_aff : (c=2, l=1, h, w)
AddChannelDim(raw)
# raw : (c=1, l=1, h, w)
# gt_aff : (c=2, l=1, h, w)
)
if is_2d:
pipeline = (
pipeline + RemoveSpatialDim(raw) + RemoveSpatialDim(gt_aff)
# raw : (c=1, h, w)
# gt_aff : (c=2, h, w)
)
pipeline = (
pipeline + gp.Stack(self.params['batch_size']) + gp.PreCache() +
# raw : (b, c=1, h, w)
# gt_aff : (b, c=2, h, w)
# (which is what train requires)
gp.torch.Train(
model,
self.loss,
optimizer,
inputs={'raw': raw},
loss_inputs={
0: predictions,
1: gt_aff
},
outputs={
0: predictions,
1: emb
},
array_specs={
predictions: gp.ArraySpec(voxel_size=out_voxel_size),
},
checkpoint_basename=self.logdir + '/checkpoints/model',
save_every=self.params['save_every'],
log_dir=self.logdir,
log_every=self.log_every) +
# everything is 2D at this point, plus extra dimensions for
# channels and batch
# raw : (b, c=1, h, w)
# gt_aff : (b, c=2, h, w)
# predictions: (b, c=2, h, w)
# Crop GT to look at labels
gp.Crop(gt_labels, gp.Roi(context, gt_labels_out_shape)) +
gp.Snapshot(output_dir=self.logdir + '/snapshots',
output_filename='it{iteration}.hdf',
dataset_names={
raw: 'raw',
gt_labels: 'gt_labels',
predictions: 'predictions',
gt_aff: 'gt_aff',
emb: 'emb'
},
additional_request=snapshot_request,
every=self.params['save_every']) +
gp.PrintProfilingStats(every=500))
return pipeline, request
def train_until(max_iteration):
# get the latest checkpoint
if tf.train.latest_checkpoint("."):
trained_until = int(tf.train.latest_checkpoint(".").split("_")[-1])
else:
trained_until = 0
if trained_until >= max_iteration:
return
# array keys for data sources
raw = gp.ArrayKey("RAW")
swcs = gp.PointsKey("SWCS")
labels = gp.ArrayKey("LABELS")
# array keys for base volume
raw_base = gp.ArrayKey("RAW_BASE")
labels_base = gp.ArrayKey("LABELS_BASE")
swc_base = gp.PointsKey("SWC_BASE")
# array keys for add volume
raw_add = gp.ArrayKey("RAW_ADD")
labels_add = gp.ArrayKey("LABELS_ADD")
swc_add = gp.PointsKey("SWC_ADD")
# array keys for fused volume
raw_fused = gp.ArrayKey("RAW_FUSED")
labels_fused = gp.ArrayKey("LABELS_FUSED")
swc_fused = gp.PointsKey("SWC_FUSED")
# output data
fg = gp.ArrayKey("FG")
labels_fg = gp.ArrayKey("LABELS_FG")
labels_fg_bin = gp.ArrayKey("LABELS_FG_BIN")
gradient_fg = gp.ArrayKey("GRADIENT_FG")
loss_weights = gp.ArrayKey("LOSS_WEIGHTS")
voxel_size = gp.Coordinate((10, 3, 3))
input_size = gp.Coordinate(net_config["input_shape"]) * voxel_size
output_size = gp.Coordinate(net_config["output_shape"]) * voxel_size
# add request
request = gp.BatchRequest()
request.add(raw_fused, input_size)
request.add(labels_fused, input_size)
request.add(swc_fused, input_size)
request.add(labels_fg, output_size)
request.add(labels_fg_bin, output_size)
request.add(loss_weights, output_size)
# add snapshot request
# request.add(fg, output_size)
# request.add(labels_fg, output_size)
request.add(gradient_fg, output_size)
request.add(raw_base, input_size)
request.add(raw_add, input_size)
request.add(labels_base, input_size)
request.add(labels_add, input_size)
request.add(swc_base, input_size)
request.add(swc_add, input_size)
data_sources = tuple(
(
gp.N5Source(
filename=str(
(
filename
/ "consensus-neurons-with-machine-centerpoints-labelled-as-swcs-carved.n5"
).absolute()
),
datasets={raw: "volume"},
array_specs={
raw: gp.ArraySpec(
interpolatable=True, voxel_size=voxel_size, dtype=np.uint16
)
},
),
MouselightSwcFileSource(
filename=str(
(
filename
/ "consensus-neurons-with-machine-centerpoints-labelled-as-swcs"
).absolute()
),
points=(swcs,),
scale=voxel_size,
transpose=(2, 1, 0),
transform_file=str((filename / "transform.txt").absolute()),
ignore_human_nodes=True
),
)
+ gp.MergeProvider()
+ gp.RandomLocation(
ensure_nonempty=swcs, ensure_centered=True
)
+ RasterizeSkeleton(
points=swcs,
array=labels,
array_spec=gp.ArraySpec(
interpolatable=False, voxel_size=voxel_size, dtype=np.uint32
),
)
+ GrowLabels(labels, radius=10)
# augment
+ gp.ElasticAugment(
[40, 10, 10],
[0.25, 1, 1],
[0, math.pi / 2.0],
subsample=4,
)
+ gp.SimpleAugment(mirror_only=[1, 2], transpose_only=[1, 2])
+ gp.Normalize(raw)
+ gp.IntensityAugment(raw, 0.9, 1.1, -0.001, 0.001)
for filename in Path(sample_dir).iterdir()
if "2018-08-01" in filename.name # or "2018-07-02" in filename.name
)
pipeline = (
data_sources
+ gp.RandomProvider()
+ GetNeuronPair(
swcs,
raw,
labels,
(swc_base, swc_add),
(raw_base, raw_add),
(labels_base, labels_add),
seperate_by=150,
shift_attempts=50,
request_attempts=10,
)
+ FusionAugment(
raw_base,
raw_add,
labels_base,
labels_add,
swc_base,
swc_add,
raw_fused,
labels_fused,
swc_fused,
blend_mode="labels_mask",
blend_smoothness=10,
num_blended_objects=0,
)
+ Crop(labels_fused, labels_fg)
+ BinarizeGt(labels_fg, labels_fg_bin)
+ gp.BalanceLabels(labels_fg_bin, loss_weights)
# train
+ gp.PreCache(cache_size=40, num_workers=10)
+ gp.tensorflow.Train(
"./train_net",
optimizer=net_names["optimizer"],
loss=net_names["loss"],
inputs={
net_names["raw"]: raw_fused,
net_names["labels_fg"]: labels_fg_bin,
net_names["loss_weights"]: loss_weights,
},
outputs={net_names["fg"]: fg},
gradients={net_names["fg"]: gradient_fg},
save_every=100000,
)
+ gp.Snapshot(
output_filename="snapshot_{iteration}.hdf",
dataset_names={
raw_fused: "volumes/raw_fused",
raw_base: "volumes/raw_base",
raw_add: "volumes/raw_add",
labels_fused: "volumes/labels_fused",
labels_base: "volumes/labels_base",
labels_add: "volumes/labels_add",
labels_fg_bin: "volumes/labels_fg_bin",
fg: "volumes/pred_fg",
gradient_fg: "volumes/gradient_fg",
},
every=100,
)
+ gp.PrintProfilingStats(every=10)
)
with gp.build(pipeline):
logging.info("Starting training...")
for i in range(max_iteration - trained_until):
logging.info("requesting batch {}".format(i))
batch = pipeline.request_batch(request)
"""
def create_pipeline_3d(task, predictor, optimizer, batch_size, outdir,
snapshot_every):
raw_channels = max(1, task.data.raw.num_channels)
input_shape = predictor.input_shape
output_shape = predictor.output_shape
voxel_size = task.data.raw.train.voxel_size
# switch to world units
input_size = voxel_size * input_shape
output_size = voxel_size * output_shape
raw = gp.ArrayKey('RAW')
gt = gp.ArrayKey('GT')
target = gp.ArrayKey('TARGET')
weights = gp.ArrayKey('WEIGHTS')
prediction = gp.ArrayKey('PREDICTION')
channel_dims = 0 if raw_channels == 1 else 1
num_samples = task.data.raw.train.num_samples
assert num_samples == 0, (
"Multiple samples for 3D training not yet implemented")
sources = (task.data.raw.train.get_source(raw),
task.data.gt.train.get_source(gt))
pipeline = sources + gp.MergeProvider()
pipeline += gp.Pad(raw, None)
# raw: ([c,] d, h, w)
# gt: ([c,] d, h, w)
pipeline += gp.Normalize(raw)
# raw: ([c,] d, h, w)
# gt: ([c,] d, h, w)
pipeline += gp.RandomLocation()
# raw: ([c,] d, h, w)
# gt: ([c,] d, h, w)
for augmentation in eval(task.augmentations):
pipeline += augmentation
pipeline += predictor.add_target(gt, target)
# (don't care about gt anymore)
# raw: ([c,] d, h, w)
# target: ([c,] d, h, w)
weights_node = task.loss.add_weights(target, weights)
if weights_node:
pipeline += weights_node
loss_inputs = {0: prediction, 1: target, 2: weights}
else:
loss_inputs = {0: prediction, 1: target}
# raw: ([c,] d, h, w)
# target: ([c,] d, h, w)
# [weights: ([c,] d, h, w)]
if channel_dims == 0:
pipeline += AddChannelDim(raw)
# raw: (c, d, h, w)
# target: ([c,] d, h, w)
# [weights: ([c,] d, h, w)]
pipeline += gp.PreCache()
pipeline += gp.Stack(batch_size)
# raw: (b, c, d, h, w)
# target: (b, [c,] d, h, w)
# [weights: (b, [c,] d, h, w)]
pipeline += gp_torch.Train(model=predictor,
loss=task.loss,
optimizer=optimizer,
inputs={'x': raw},
loss_inputs=loss_inputs,
outputs={0: prediction},
save_every=1e6)
# raw: (b, c, d, h, w)
# target: (b, [c,] d, h, w)
# [weights: (b, [c,] d, h, w)]
# prediction: (b, [c,] d, h, w)
if snapshot_every > 0:
# get channels first
pipeline += TransposeDims(raw, (1, 0, 2, 3, 4))
if predictor.target_channels > 0:
pipeline += TransposeDims(target, (1, 0, 2, 3, 4))
if weights_node:
pipeline += TransposeDims(weights, (1, 0, 2, 3, 4))
if predictor.prediction_channels > 0:
pipeline += TransposeDims(prediction, (1, 0, 2, 3, 4))
# raw: (c, b, d, h, w)
# target: ([c,] b, d, h, w)
# [weights: ([c,] b, d, h, w)]
# prediction: ([c,] b, d, h, w)
if channel_dims == 0:
pipeline += RemoveChannelDim(raw)
# raw: ([c,] b, d, h, w)
# target: (c, b, d, h, w)
# [weights: ([c,] b, d, h, w)]
# prediction: (c, b, d, h, w)
pipeline += gp.Snapshot(dataset_names={
raw: 'raw',
target: 'target',
prediction: 'prediction',
weights: 'weights'
},
every=snapshot_every,
output_dir=os.path.join(outdir, 'snapshots'),
output_filename="{iteration}.hdf")
pipeline += gp.PrintProfilingStats(every=100)
request = gp.BatchRequest()
request.add(raw, input_size)
request.add(gt, output_size)
request.add(target, output_size)
if weights_node:
request.add(weights, output_size)
request.add(prediction, output_size)
return pipeline, request
def add_caching(pipeline, setup_config):
cache_size = setup_config["CACHE_SIZE"]
num_workers = setup_config["NUM_WORKERS"]
pipeline = pipeline + gp.PreCache(cache_size=cache_size,
num_workers=num_workers)
return pipeline
def create_pipeline_3d(
task, data, predictor, optimizer, batch_size, outdir, snapshot_every
):
raw_channels = max(1, data.raw.num_channels)
input_shape = gp.Coordinate(task.model.input_shape)
output_shape = gp.Coordinate(task.model.output_shape)
voxel_size = data.raw.train.voxel_size
task.predictor = task.predictor.to("cuda")
# switch to world units
input_size = voxel_size * input_shape
output_size = voxel_size * output_shape
raw = gp.ArrayKey("RAW")
gt = gp.ArrayKey("GT")
mask = gp.ArrayKey("MASK")
target = gp.ArrayKey("TARGET")
weights = gp.ArrayKey("WEIGHTS")
model_outputs = gp.ArrayKey("MODEL_OUTPUTS")
model_output_grads = gp.ArrayKey("MODEL_OUT_GRAD")
prediction = gp.ArrayKey("PREDICTION")
pred_gradients = gp.ArrayKey("PRED_GRADIENTS")
snapshot_dataset_names = {
raw: "raw",
model_outputs: "model_outputs",
model_output_grads: "model_out_grad",
target: "target",
prediction: "prediction",
pred_gradients: "pred_gradients",
weights: "weights",
}
aux_keys = {}
aux_grad_keys = {}
for name, _, _ in task.aux_tasks:
aux_keys[name] = (
gp.ArrayKey(f"{name.upper()}_PREDICTION"),
gp.ArrayKey(f"{name.upper()}_TARGET"),
None,
)
aux_grad_keys[name] = gp.ArrayKey(f"{name.upper()}_PRED_GRAD")
aux_pred, aux_target, _ = aux_keys[name]
snapshot_dataset_names[aux_pred] = f"{name}_pred"
snapshot_dataset_names[aux_target] = f"{name}_target"
aux_grad = aux_grad_keys[name]
snapshot_dataset_names[aux_grad] = f"{name}_aux_grad"
channel_dims = 0 if raw_channels == 1 else 1
num_samples = data.raw.train.num_samples
assert num_samples == 0, "Multiple samples for 3D training not yet implemented"
sources = (data.raw.train.get_source(raw), data.gt.train.get_source(gt))
pipeline = sources + gp.MergeProvider()
pipeline += gp.Pad(raw, input_shape / 2 * voxel_size)
# pipeline += gp.Pad(gt, input_shape / 2 * voxel_size)
# raw: ([c,] d, h, w)
# gt: ([c,] d, h, w)
pipeline += gp.Normalize(raw)
mask_node = task.loss.add_mask(gt, mask)
if mask_node is not None:
pipeline += mask_node
pipeline += gp.RandomLocation(min_masked=1e-6, mask=mask)
else:
# raw: ([c,] d, h, w)
# gt: ([c,] d, h, w)
pipeline += gp.RandomLocation()
# raw: ([c,] d, h, w)
# gt: ([c,] d, h, w)
for augmentation in eval(task.augmentations):
pipeline += augmentation
pipeline += predictor.add_target(gt, target)
# (don't care about gt anymore)
# raw: ([c,] d, h, w)
# target: ([c,] d, h, w)
weights_node = task.loss.add_weights(target, weights)
loss_inputs = []
if weights_node:
pipeline += weights_node
loss_inputs.append({0: prediction, 1: target, 2: weights})
else:
loss_inputs.append({0: prediction, 1: target})
head_outputs = []
head_gradients = []
for name, aux_predictor, aux_loss in task.aux_tasks:
aux_prediction, aux_target, aux_weights = aux_keys[name]
pipeline += aux_predictor.add_target(gt, aux_target)
aux_weights_node = aux_loss.add_weights(aux_target, aux_weights)
if aux_weights_node:
aux_weights = gp.ArrayKey(f"{name.upper()}_WEIGHTS")
aux_keys[name] = (
aux_prediction,
aux_target,
aux_weights,
)
pipeline += aux_weights_node
loss_inputs.append({0: aux_prediction, 1: aux_target, 2: aux_weights})
snapshot_dataset_names[aux_weights] = f"{name}_weights"
else:
loss_inputs.append({0: aux_prediction, 1: aux_target})
head_outputs.append({0: aux_prediction})
aux_pred_gradient = aux_grad_keys[name]
head_gradients.append({0: aux_pred_gradient})
# raw: ([c,] d, h, w)
# target: ([c,] d, h, w)
# [weights: ([c,] d, h, w)]
if channel_dims == 0:
pipeline += AddChannelDim(raw)
# raw: (c, d, h, w)
# target: ([c,] d, h, w)
# [weights: ([c,] d, h, w)]
pipeline += gp.PreCache()
pipeline += gp.Stack(batch_size)
# raw: (b, c, d, h, w)
# target: (b, [c,] d, h, w)
# [weights: (b, [c,] d, h, w)]
pipeline += Train(
model=task.model,
heads=[("opt", predictor)]
+ [(name, aux_pred) for name, aux_pred, _ in task.aux_tasks],
losses=[task.loss] + [loss for _, _, loss in task.aux_tasks],
optimizer=optimizer,
inputs={"x": raw},
outputs={0: model_outputs},
head_outputs=[{0: prediction}] + head_outputs,
loss_inputs=loss_inputs,
gradients=[{0: model_output_grads}, {0: pred_gradients}] + head_gradients,
save_every=1e6,
)
# raw: (b, c, d, h, w)
# target: (b, [c,] d, h, w)
# [weights: (b, [c,] d, h, w)]
# prediction: (b, [c,] d, h, w)
if snapshot_every > 0:
# get channels first
pipeline += TransposeDims(raw, (1, 0, 2, 3, 4))
if predictor.target_channels > 0:
pipeline += TransposeDims(target, (1, 0, 2, 3, 4))
if weights_node:
pipeline += TransposeDims(weights, (1, 0, 2, 3, 4))
if predictor.prediction_channels > 0:
pipeline += TransposeDims(prediction, (1, 0, 2, 3, 4))
# raw: (c, b, d, h, w)
# target: ([c,] b, d, h, w)
# [weights: ([c,] b, d, h, w)]
# prediction: ([c,] b, d, h, w)
if channel_dims == 0:
pipeline += RemoveChannelDim(raw)
# raw: ([c,] b, d, h, w)
# target: (c, b, d, h, w)
# [weights: ([c,] b, d, h, w)]
# prediction: (c, b, d, h, w)
pipeline += gp.Snapshot(
dataset_names=snapshot_dataset_names,
every=snapshot_every,
output_dir=os.path.join(outdir, "snapshots"),
output_filename="{iteration}.hdf",
)
pipeline += gp.PrintProfilingStats(every=10)
request = gp.BatchRequest()
request.add(raw, input_size)
request.add(gt, output_size)
if mask_node is not None:
request.add(mask, output_size)
request.add(target, output_size)
for name, _, _ in task.aux_tasks:
aux_pred, aux_target, aux_weight = aux_keys[name]
request.add(aux_pred, output_size)
request.add(aux_target, output_size)
if aux_weight is not None:
request.add(aux_weight, output_size)
aux_pred_grad = aux_grad_keys[name]
request.add(aux_pred_grad, output_size)
if weights_node:
request.add(weights, output_size)
request.add(prediction, output_size)
request.add(pred_gradients, output_size)
return pipeline, request
def train_until(max_iteration):
# get the latest checkpoint
if tf.train.latest_checkpoint("."):
trained_until = int(tf.train.latest_checkpoint(".").split("_")[-1])
else:
trained_until = 0
if trained_until >= max_iteration:
return
# array keys for fused volume
raw = gp.ArrayKey("RAW")
labels = gp.ArrayKey("LABELS")
labels_fg = gp.ArrayKey("LABELS_FG")
# array keys for base volume
raw_base = gp.ArrayKey("RAW_BASE")
labels_base = gp.ArrayKey("LABELS_BASE")
swc_base = gp.PointsKey("SWC_BASE")
swc_center_base = gp.PointsKey("SWC_CENTER_BASE")
# array keys for add volume
raw_add = gp.ArrayKey("RAW_ADD")
labels_add = gp.ArrayKey("LABELS_ADD")
swc_add = gp.PointsKey("SWC_ADD")
swc_center_add = gp.PointsKey("SWC_CENTER_ADD")
# output data
fg = gp.ArrayKey("FG")
gradient_fg = gp.ArrayKey("GRADIENT_FG")
loss_weights = gp.ArrayKey("LOSS_WEIGHTS")
voxel_size = gp.Coordinate((4, 1, 1))
input_size = gp.Coordinate(net_config["input_shape"]) * voxel_size
output_size = gp.Coordinate(net_config["output_shape"]) * voxel_size
# add request
request = gp.BatchRequest()
request.add(raw, input_size)
request.add(labels, output_size)
request.add(labels_fg, output_size)
request.add(loss_weights, output_size)
request.add(swc_center_base, output_size)
request.add(swc_center_add, output_size)
# add snapshot request
snapshot_request = gp.BatchRequest()
snapshot_request.add(fg, output_size)
snapshot_request.add(labels_fg, output_size)
snapshot_request.add(gradient_fg, output_size)
snapshot_request.add(raw_base, input_size)
snapshot_request.add(raw_add, input_size)
snapshot_request.add(labels_base, input_size)
snapshot_request.add(labels_add, input_size)
# data source for "base" volume
data_sources_base = tuple(
(
gp.Hdf5Source(
filename,
datasets={raw_base: "/volume"},
array_specs={
raw_base:
gp.ArraySpec(interpolatable=True,
voxel_size=voxel_size,
dtype=np.uint16)
},
channels_first=False,
),
SwcSource(
filename=filename,
dataset="/reconstruction",
points=(swc_center_base, swc_base),
scale=voxel_size,
),
) + gp.MergeProvider() +
gp.RandomLocation(ensure_nonempty=swc_center_base) + RasterizeSkeleton(
points=swc_base,
array=labels_base,
array_spec=gp.ArraySpec(
interpolatable=False, voxel_size=voxel_size, dtype=np.uint32),
radius=5.0,
) for filename in files)
# data source for "add" volume
data_sources_add = tuple(
(
gp.Hdf5Source(
file,
datasets={raw_add: "/volume"},
array_specs={
raw_add:
gp.ArraySpec(interpolatable=True,
voxel_size=voxel_size,
dtype=np.uint16)
},
channels_first=False,
),
SwcSource(
filename=file,
dataset="/reconstruction",
points=(swc_center_add, swc_add),
scale=voxel_size,
),
) + gp.MergeProvider() +
gp.RandomLocation(ensure_nonempty=swc_center_add) + RasterizeSkeleton(
points=swc_add,
array=labels_add,
array_spec=gp.ArraySpec(
interpolatable=False, voxel_size=voxel_size, dtype=np.uint32),
radius=5.0,
) for file in files)
data_sources = (
(data_sources_base + gp.RandomProvider()),
(data_sources_add + gp.RandomProvider()),
) + gp.MergeProvider()
pipeline = (
data_sources + FusionAugment(
raw_base,
raw_add,
labels_base,
labels_add,
raw,
labels,
blend_mode="labels_mask",
blend_smoothness=10,
num_blended_objects=0,
) +
# augment
gp.ElasticAugment([40, 10, 10], [0.25, 1, 1], [0, math.pi / 2.0],
subsample=4) +
gp.SimpleAugment(mirror_only=[1, 2], transpose_only=[1, 2]) +
gp.Normalize(raw) + gp.IntensityAugment(raw, 0.9, 1.1, -0.001, 0.001) +
BinarizeGt(labels, labels_fg) +
gp.BalanceLabels(labels_fg, loss_weights) +
# train
gp.PreCache(cache_size=40, num_workers=10) + gp.tensorflow.Train(
"./train_net",
optimizer=net_names["optimizer"],
loss=net_names["loss"],
inputs={
net_names["raw"]: raw,
net_names["labels_fg"]: labels_fg,
net_names["loss_weights"]: loss_weights,
},
outputs={net_names["fg"]: fg},
gradients={net_names["fg"]: gradient_fg},
save_every=100000,
) +
# visualize
gp.Snapshot(
output_filename="snapshot_{iteration}.hdf",
dataset_names={
raw: "volumes/raw",
raw_base: "volumes/raw_base",
raw_add: "volumes/raw_add",
labels: "volumes/labels",
labels_base: "volumes/labels_base",
labels_add: "volumes/labels_add",
fg: "volumes/fg",
labels_fg: "volumes/labels_fg",
gradient_fg: "volumes/gradient_fg",
},
additional_request=snapshot_request,
every=100,
) + gp.PrintProfilingStats(every=100))
with gp.build(pipeline):
print("Starting training...")
for i in range(max_iteration - trained_until):
pipeline.request_batch(request)
def train_until(max_iteration, name='train_net', output_folder='.', clip_max=2000):
# get the latest checkpoint
if tf.train.latest_checkpoint(output_folder):
trained_until = int(tf.train.latest_checkpoint(output_folder).split('_')[-1])
else:
trained_until = 0
if trained_until >= max_iteration:
return
with open(os.path.join(output_folder, name + '_config.json'), 'r') as f:
net_config = json.load(f)
with open(os.path.join(output_folder, name + '_names.json'), 'r') as f:
net_names = json.load(f)
# array keys
raw = gp.ArrayKey('RAW')
gt_instances = gp.ArrayKey('GT_INSTANCES')
gt_mask = gp.ArrayKey('GT_MASK')
pred_mask = gp.ArrayKey('PRED_MASK')
#loss_weights = gp.ArrayKey('LOSS_WEIGHTS')
loss_gradients = gp.ArrayKey('LOSS_GRADIENTS')
# array keys for base and add volume
raw_base = gp.ArrayKey('RAW_BASE')
gt_instances_base = gp.ArrayKey('GT_INSTANCES_BASE')
gt_mask_base = gp.ArrayKey('GT_MASK_BASE')
raw_add = gp.ArrayKey('RAW_ADD')
gt_instances_add = gp.ArrayKey('GT_INSTANCES_ADD')
gt_mask_add = gp.ArrayKey('GT_MASK_ADD')
voxel_size = gp.Coordinate((1, 1, 1))
input_shape = gp.Coordinate(net_config['input_shape'])
output_shape = gp.Coordinate(net_config['output_shape'])
context = gp.Coordinate(input_shape - output_shape) / 2
request = gp.BatchRequest()
request.add(raw, input_shape)
request.add(gt_instances, output_shape)
request.add(gt_mask, output_shape)
#request.add(loss_weights, output_shape)
request.add(raw_base, input_shape)
request.add(raw_add, input_shape)
request.add(gt_mask_base, output_shape)
request.add(gt_mask_add, output_shape)
snapshot_request = gp.BatchRequest()
snapshot_request.add(raw, input_shape)
#snapshot_request.add(raw_base, input_shape)
#snapshot_request.add(raw_add, input_shape)
snapshot_request.add(gt_mask, output_shape)
#snapshot_request.add(gt_mask_base, output_shape)
#snapshot_request.add(gt_mask_add, output_shape)
snapshot_request.add(pred_mask, output_shape)
snapshot_request.add(loss_gradients, output_shape)
# specify data source
# data source for base volume
data_sources_base = tuple()
for data_file in data_files:
current_path = os.path.join(data_dir, data_file)
with h5py.File(current_path, 'r') as f:
data_sources_base += tuple(
gp.Hdf5Source(
current_path,
datasets={
raw_base: sample + '/raw',
gt_instances_base: sample + '/gt',
gt_mask_base: sample + '/fg',
},
array_specs={
raw_base: gp.ArraySpec(interpolatable=True, dtype=np.uint16, voxel_size=voxel_size),
gt_instances_base: gp.ArraySpec(interpolatable=False, dtype=np.uint16, voxel_size=voxel_size),
gt_mask_base: gp.ArraySpec(interpolatable=False, dtype=np.bool, voxel_size=voxel_size),
}
) +
Convert(gt_mask_base, np.uint8) +
gp.Pad(raw_base, context) +
gp.Pad(gt_instances_base, context) +
gp.Pad(gt_mask_base, context) +
gp.RandomLocation(min_masked=0.005, mask=gt_mask_base)
#gp.Reject(gt_mask_base, min_masked=0.005, reject_probability=1.)
for sample in f)
data_sources_base += gp.RandomProvider()
# data source for add volume
data_sources_add = tuple()
for data_file in data_files:
current_path = os.path.join(data_dir, data_file)
with h5py.File(current_path, 'r') as f:
data_sources_add += tuple(
gp.Hdf5Source(
current_path,
datasets={
raw_add: sample + '/raw',
gt_instances_add: sample + '/gt',
gt_mask_add: sample + '/fg',
},
array_specs={
raw_add: gp.ArraySpec(interpolatable=True, dtype=np.uint16, voxel_size=voxel_size),
gt_instances_add: gp.ArraySpec(interpolatable=False, dtype=np.uint16, voxel_size=voxel_size),
gt_mask_add: gp.ArraySpec(interpolatable=False, dtype=np.bool, voxel_size=voxel_size),
}
) +
Convert(gt_mask_add, np.uint8) +
gp.Pad(raw_add, context) +
gp.Pad(gt_instances_add, context) +
gp.Pad(gt_mask_add, context) +
gp.RandomLocation() +
gp.Reject(gt_mask_add, min_masked=0.005, reject_probability=0.95)
for sample in f)
data_sources_add += gp.RandomProvider()
data_sources = tuple([data_sources_base, data_sources_add]) + gp.MergeProvider()
pipeline = (
data_sources +
nl.FusionAugment(
raw_base, raw_add, gt_instances_base, gt_instances_add, raw, gt_instances,
blend_mode='labels_mask', blend_smoothness=5, num_blended_objects=0
) +
BinarizeLabels(gt_instances, gt_mask) +
nl.Clip(raw, 0, clip_max) +
gp.Normalize(raw, factor=1.0/clip_max) +
gp.ElasticAugment(
control_point_spacing=[20, 20, 20],
jitter_sigma=[1, 1, 1],
rotation_interval=[0, math.pi/2.0],
subsample=4) +
gp.SimpleAugment(mirror_only=[1, 2], transpose_only=[1, 2]) +
gp.IntensityAugment(raw, 0.9, 1.1, -0.1, 0.1) +
gp.IntensityScaleShift(raw, 2, -1) +
#gp.BalanceLabels(gt_mask, loss_weights) +
# train
gp.PreCache(
cache_size=40,
num_workers=10) +
gp.tensorflow.Train(
os.path.join(output_folder, name),
optimizer=net_names['optimizer'],
loss=net_names['loss'],
inputs={
net_names['raw']: raw,
net_names['gt']: gt_mask,
#net_names['loss_weights']: loss_weights,
},
outputs={
net_names['pred']: pred_mask,
},
gradients={
net_names['output']: loss_gradients,
},
save_every=5000) +
# visualize
gp.Snapshot({
raw: 'volumes/raw',
pred_mask: 'volumes/pred_mask',
gt_mask: 'volumes/gt_mask',
#loss_weights: 'volumes/loss_weights',
loss_gradients: 'volumes/loss_gradients',
},
output_filename=os.path.join(output_folder, 'snapshots', 'batch_{iteration}.hdf'),
additional_request=snapshot_request,
every=2500) +
gp.PrintProfilingStats(every=1000)
)
with gp.build(pipeline):
print("Starting training...")
for i in range(max_iteration - trained_until):
pipeline.request_batch(request)
def train_until(**kwargs):
if tf.train.latest_checkpoint(kwargs['output_folder']):
trained_until = int(
tf.train.latest_checkpoint(kwargs['output_folder']).split('_')[-1])
else:
trained_until = 0
if trained_until >= kwargs['max_iteration']:
return
anchor = gp.ArrayKey('ANCHOR')
raw = gp.ArrayKey('RAW')
raw_cropped = gp.ArrayKey('RAW_CROPPED')
points = gp.PointsKey('POINTS')
gt_cp = gp.ArrayKey('GT_CP')
pred_cp = gp.ArrayKey('PRED_CP')
pred_cp_gradients = gp.ArrayKey('PRED_CP_GRADIENTS')
with open(
os.path.join(kwargs['output_folder'],
kwargs['name'] + '_config.json'), 'r') as f:
net_config = json.load(f)
with open(
os.path.join(kwargs['output_folder'],
kwargs['name'] + '_names.json'), 'r') as f:
net_names = json.load(f)
voxel_size = gp.Coordinate(kwargs['voxel_size'])
input_shape_world = gp.Coordinate(net_config['input_shape']) * voxel_size
output_shape_world = gp.Coordinate(net_config['output_shape']) * voxel_size
# formulate the request for what a batch should (at least) contain
request = gp.BatchRequest()
request.add(raw, input_shape_world)
request.add(raw_cropped, output_shape_world)
request.add(gt_cp, output_shape_world)
request.add(anchor, output_shape_world)
# when we make a snapshot for inspection (see below), we also want to
# request the predicted affinities and gradients of the loss wrt the
# affinities
snapshot_request = gp.BatchRequest()
snapshot_request.add(raw_cropped, output_shape_world)
snapshot_request.add(gt_cp, output_shape_world)
snapshot_request.add(pred_cp, output_shape_world)
# snapshot_request.add(pred_cp_gradients, output_shape_world)
if kwargs['input_format'] != "hdf" and kwargs['input_format'] != "zarr":
raise NotImplementedError("train node for %s not implemented yet",
kwargs['input_format'])
fls = []
shapes = []
mn = []
mx = []
for f in kwargs['data_files']:
fls.append(os.path.splitext(f)[0])
if kwargs['input_format'] == "hdf":
vol = h5py.File(f, 'r')['volumes/raw']
elif kwargs['input_format'] == "zarr":
vol = zarr.open(f, 'r')['volumes/raw']
print(f, vol.shape, vol.dtype)
shapes.append(vol.shape)
mn.append(np.min(vol))
mx.append(np.max(vol))
if vol.dtype != np.float32:
print("please convert to float32")
ln = len(fls)
print("first 5 files: ", fls[0:4])
if kwargs['input_format'] == "hdf":
sourceNode = gp.Hdf5Source
elif kwargs['input_format'] == "zarr":
sourceNode = gp.ZarrSource
augmentation = kwargs['augmentation']
sources = tuple(
(sourceNode(fls[t] + "." + kwargs['input_format'],
datasets={
raw: 'volumes/raw',
anchor: 'volumes/gt_fgbg',
},
array_specs={
raw: gp.ArraySpec(interpolatable=True),
anchor: gp.ArraySpec(interpolatable=False)
}),
gp.CsvIDPointsSource(fls[t] + ".csv",
points,
points_spec=gp.PointsSpec(
roi=gp.Roi(gp.Coordinate((
0, 0, 0)), gp.Coordinate(shapes[t]))))) +
gp.MergeProvider()
# + Clip(raw, mn=mn[t], mx=mx[t])
# + NormalizeMinMax(raw, mn=mn[t], mx=mx[t])
+ gp.Pad(raw, None) + gp.Pad(points, None)
# chose a random location for each requested batch
+ gp.RandomLocation() for t in range(ln))
pipeline = (
sources +
# chose a random source (i.e., sample) from the above
gp.RandomProvider() +
# elastically deform the batch
(gp.ElasticAugment(
augmentation['elastic']['control_point_spacing'],
augmentation['elastic']['jitter_sigma'],
[augmentation['elastic']['rotation_min']*np.pi/180.0,
augmentation['elastic']['rotation_max']*np.pi/180.0],
subsample=augmentation['elastic'].get('subsample', 1)) \
if augmentation.get('elastic') is not None else NoOp()) +
# apply transpose and mirror augmentations
gp.SimpleAugment(mirror_only=augmentation['simple'].get("mirror"),
transpose_only=augmentation['simple'].get("transpose")) +
# (gp.SimpleAugment(
# mirror_only=augmentation['simple'].get("mirror"),
# transpose_only=augmentation['simple'].get("transpose")) \
# if augmentation.get('simple') is not None and \
# augmentation.get('simple') != {} else NoOp()) +
# # scale and shift the intensity of the raw array
(gp.IntensityAugment(
raw,
scale_min=augmentation['intensity']['scale'][0],
scale_max=augmentation['intensity']['scale'][1],
shift_min=augmentation['intensity']['shift'][0],
shift_max=augmentation['intensity']['shift'][1],
z_section_wise=False) \
if augmentation.get('intensity') is not None and \
augmentation.get('intensity') != {} else NoOp()) +
gp.RasterizePoints(
points,
gt_cp,
array_spec=gp.ArraySpec(voxel_size=voxel_size),
settings=gp.RasterizationSettings(
radius=(2, 2, 2),
mode='peak')) +
# pre-cache batches from the point upstream
gp.PreCache(
cache_size=kwargs['cache_size'],
num_workers=kwargs['num_workers']) +
# perform one training iteration for each passing batch (here we use
# the tensor names earlier stored in train_net.config)
gp.tensorflow.Train(
os.path.join(kwargs['output_folder'], kwargs['name']),
optimizer=net_names['optimizer'],
summary=net_names['summaries'],
log_dir=kwargs['output_folder'],
loss=net_names['loss'],
inputs={
net_names['raw']: raw,
net_names['gt_cp']: gt_cp,
net_names['anchor']: anchor,
},
outputs={
net_names['pred_cp']: pred_cp,
net_names['raw_cropped']: raw_cropped,
},
gradients={
# net_names['pred_cp']: pred_cp_gradients,
},
save_every=kwargs['checkpoints']) +
# save the passing batch as an HDF5 file for inspection
gp.Snapshot(
{
raw: '/volumes/raw',
raw_cropped: 'volumes/raw_cropped',
gt_cp: '/volumes/gt_cp',
pred_cp: '/volumes/pred_cp',
# pred_cp_gradients: '/volumes/pred_cp_gradients',
},
output_dir=os.path.join(kwargs['output_folder'], 'snapshots'),
output_filename='batch_{iteration}.hdf',
every=kwargs['snapshots'],
additional_request=snapshot_request,
compression_type='gzip') +
# show a summary of time spend in each node every 10 iterations
gp.PrintProfilingStats(every=kwargs['profiling'])
)
#########
# TRAIN #
#########
print("Starting training...")
with gp.build(pipeline):
print(pipeline)
for i in range(trained_until, kwargs['max_iteration']):
# print("request", request)
start = time.time()
pipeline.request_batch(request)
time_of_iteration = time.time() - start
logger.info("Batch: iteration=%d, time=%f", i, time_of_iteration)
# exit()
print("Training finished")
def train_until(**kwargs):
print("cuda visibile devices", os.environ["CUDA_VISIBLE_DEVICES"])
if tf.train.latest_checkpoint(kwargs['output_folder']):
trained_until = int(
tf.train.latest_checkpoint(kwargs['output_folder']).split('_')[-1])
else:
trained_until = 0
if trained_until >= kwargs['max_iteration']:
return
raw = gp.ArrayKey('RAW')
raw_cropped = gp.ArrayKey('RAW_CROPPED')
gt_labels = gp.ArrayKey('GT_LABELS')
gt_instances = gp.ArrayKey('GT_INSTANCES')
gt_affs = gp.ArrayKey('GT_AFFS')
gt_numinst = gp.ArrayKey('GT_NUMINST')
gt_sample_mask = gp.ArrayKey('GT_SAMPLE_MASK')
pred_affs = gp.ArrayKey('PRED_AFFS')
pred_affs_gradients = gp.ArrayKey('PRED_AFFS_GRADIENTS')
pred_numinst = gp.ArrayKey('PRED_NUMINST')
with open(os.path.join(kwargs['output_folder'],
kwargs['name'] + '_config.json'), 'r') as f:
net_config = json.load(f)
with open(os.path.join(kwargs['output_folder'],
kwargs['name'] + '_names.json'), 'r') as f:
net_names = json.load(f)
voxel_size = gp.Coordinate(kwargs['voxel_size'])
input_shape_world = gp.Coordinate(net_config['input_shape'])*voxel_size
output_shape_world = gp.Coordinate(net_config['output_shape'])*voxel_size
context = gp.Coordinate(input_shape_world - output_shape_world) / 2
# formulate the request for what a batch should (at least) contain
request = gp.BatchRequest()
request.add(raw, input_shape_world)
request.add(raw_cropped, output_shape_world)
request.add(gt_labels, output_shape_world)
request.add(gt_instances, output_shape_world)
request.add(gt_sample_mask, output_shape_world)
request.add(gt_affs, output_shape_world)
if kwargs['overlapping_inst']:
request.add(gt_numinst, output_shape_world)
# request.add(loss_weights_affs, output_shape_world)
# when we make a snapshot for inspection (see below), we also want to
# request the predicted affinities and gradients of the loss wrt the
# affinities
snapshot_request = gp.BatchRequest()
snapshot_request.add(raw_cropped, output_shape_world)
snapshot_request.add(pred_affs, output_shape_world)
if kwargs['overlapping_inst']:
snapshot_request.add(pred_numinst, output_shape_world)
# snapshot_request.add(pred_affs_gradients, output_shape_world)
if kwargs['input_format'] != "hdf" and kwargs['input_format'] != "zarr":
raise NotImplementedError("train node for %s not implemented yet",
kwargs['input_format'])
raw_key = kwargs.get('raw_key', 'volumes/raw')
print('raw key: ', raw_key)
fls = []
shapes = []
for f in kwargs['data_files']:
fls.append(os.path.splitext(f)[0])
if kwargs['input_format'] == "hdf":
vol = h5py.File(f, 'r')[raw_key]
elif kwargs['input_format'] == "zarr":
vol = zarr.open(f, 'r')[raw_key]
# print(f, vol.shape, vol.dtype)
shapes.append(vol.shape)
if vol.dtype != np.float32:
print("please convert to float32")
ln = len(fls)
print("first 5 files: ", fls[0:4])
if kwargs['input_format'] == "hdf":
sourceNode = gp.Hdf5Source
elif kwargs['input_format'] == "zarr":
sourceNode = gp.ZarrSource
neighborhood = []
psH = np.array(kwargs['patchshape'])//2
for i in range(-psH[1], psH[1]+1, kwargs['patchstride'][1]):
for j in range(-psH[2], psH[2]+1, kwargs['patchstride'][2]):
neighborhood.append([i,j])
datasets = {
raw: raw_key,
gt_labels: 'volumes/gt_labels',
gt_instances: 'volumes/gt_instances'
}
array_specs = {
raw: gp.ArraySpec(interpolatable=True),
gt_labels: gp.ArraySpec(interpolatable=False),
gt_instances: gp.ArraySpec(interpolatable=False)
}
inputs = {
net_names['raw']: raw,
net_names['gt_affs']: gt_affs,
# net_names['loss_weights_affs']: loss_weights_affs,
}
outputs = {
net_names['pred_affs']: pred_affs,
net_names['raw_cropped']: raw_cropped,
}
snapshot = {
raw: '/volumes/raw',
raw_cropped: 'volumes/raw_cropped',
gt_affs: '/volumes/gt_affs',
pred_affs: '/volumes/pred_affs',
pred_affs_gradients: '/volumes/pred_affs_gradients',
}
if kwargs['overlapping_inst']:
datasets[gt_numinst] = 'volumes/gt_numinst'
array_specs[gt_numinst] = gp.ArraySpec(interpolatable=False)
inputs[net_names['gt_numinst']] = gt_numinst
outputs[net_names['pred_numinst']] = pred_numinst
snapshot[gt_numinst] = '/volumes/gt_numinst'
snapshot[pred_numinst] = '/volumes/pred_numinst'
augmentation = kwargs['augmentation']
sampling = kwargs['sampling']
source_fg = tuple(
sourceNode(
fls[t] + "." + kwargs['input_format'],
datasets=datasets,
array_specs=array_specs
) +
gp.Pad(raw, context) +
# chose a random location for each requested batch
nl.CountOverlap(gt_labels, gt_sample_mask, maxnuminst=1) +
gp.RandomLocation(
min_masked=sampling['min_masked'],
mask=gt_sample_mask
)
for t in range(ln)
)
source_fg += gp.RandomProvider()
source_overlap = tuple(
sourceNode(
fls[t] + "." + kwargs['input_format'],
datasets=datasets,
array_specs=array_specs
) +
gp.Pad(raw, context) +
# chose a random location for each requested batch
nl.MaskCloseDistanceToOverlap(
gt_labels, gt_sample_mask,
sampling['overlap_min_dist'],
sampling['overlap_max_dist']
) +
gp.RandomLocation(
min_masked=sampling['min_masked_overlap'],
mask=gt_sample_mask
)
for t in range(ln)
)
source_overlap += gp.RandomProvider()
pipeline = (
(source_fg, source_overlap) +
# chose a random source (i.e., sample) from the above
gp.RandomProvider(probabilities=[sampling['probability_fg'],
sampling['probability_overlap']]) +
# elastically deform the batch
gp.ElasticAugment(
augmentation['elastic']['control_point_spacing'],
augmentation['elastic']['jitter_sigma'],
[augmentation['elastic']['rotation_min']*np.pi/180.0,
augmentation['elastic']['rotation_max']*np.pi/180.0]) +
# apply transpose and mirror augmentations
gp.SimpleAugment(
mirror_only=augmentation['simple'].get("mirror"),
transpose_only=augmentation['simple'].get("transpose")) +
# # scale and shift the intensity of the raw array
gp.IntensityAugment(
raw,
scale_min=augmentation['intensity']['scale'][0],
scale_max=augmentation['intensity']['scale'][1],
shift_min=augmentation['intensity']['shift'][0],
shift_max=augmentation['intensity']['shift'][1],
z_section_wise=False) +
gp.IntensityScaleShift(raw, 2, -1) +
# convert labels into affinities between voxels
nl.AddAffinities(
neighborhood,
gt_labels,
gt_affs,
multiple_labels=kwargs['overlapping_inst']) +
# pre-cache batches from the point upstream
gp.PreCache(
cache_size=kwargs['cache_size'],
num_workers=kwargs['num_workers']) +
# perform one training iteration for each passing batch (here we use
# the tensor names earlier stored in train_net.config)
gp.tensorflow.Train(
os.path.join(kwargs['output_folder'], kwargs['name']),
optimizer=net_names['optimizer'],
summary=net_names['summaries'],
log_dir=kwargs['output_folder'],
loss=net_names['loss'],
inputs=inputs,
outputs=outputs,
gradients={
net_names['pred_affs']: pred_affs_gradients,
},
save_every=kwargs['checkpoints']) +
# save the passing batch as an HDF5 file for inspection
gp.Snapshot(
snapshot,
output_dir=os.path.join(kwargs['output_folder'], 'snapshots'),
output_filename='batch_{iteration}.hdf',
every=kwargs['snapshots'],
additional_request=snapshot_request,
compression_type='gzip') +
# show a summary of time spend in each node every 10 iterations
gp.PrintProfilingStats(every=kwargs['profiling'])
)
#########
# TRAIN #
#########
print("Starting training...")
with gp.build(pipeline):
print(pipeline)
for i in range(trained_until, kwargs['max_iteration']):
# print("request", request)
start = time.time()
pipeline.request_batch(request)
time_of_iteration = time.time() - start
logger.info(
"Batch: iteration=%d, time=%f",
i, time_of_iteration)
# exit()
print("Training finished")
def make_pipeline(self):
raw = gp.ArrayKey('RAW')
embs = gp.ArrayKey('EMBS')
source_shape = zarr.open(self.data_file)[self.dataset].shape
raw_roi = gp.Roi(np.zeros(len(source_shape[1:])), source_shape[1:])
data = daisy.open_ds(self.data_file, self.dataset)
source_roi = gp.Roi(data.roi.get_offset(), data.roi.get_shape())
voxel_size = gp.Coordinate(data.voxel_size)
# Get in and out shape
in_shape = gp.Coordinate(self.model.in_shape)
out_shape = gp.Coordinate(self.model.out_shape[2:])
is_2d = in_shape.dims() == 2
logger.info(f"source roi: {source_roi}")
logger.info(f"in_shape: {in_shape}")
logger.info(f"out_shape: {out_shape}")
logger.info(f"voxel_size: {voxel_size}")
in_shape = in_shape * voxel_size
out_shape = out_shape * voxel_size
logger.info(f"source roi: {source_roi}")
logger.info(f"in_shape: {in_shape}")
logger.info(f"out_shape: {out_shape}")
logger.info(f"voxel_size: {voxel_size}")
request = gp.BatchRequest()
request.add(raw, in_shape)
request.add(embs, out_shape)
context = (in_shape - out_shape) / 2
source = (gp.ZarrSource(self.data_file, {
raw: self.dataset,
},
array_specs={
raw:
gp.ArraySpec(roi=source_roi,
interpolatable=False)
}))
if is_2d:
source = (source + AddChannelDim(raw, axis=1))
else:
source = (source + AddChannelDim(raw, axis=0) + AddChannelDim(raw))
source = (
source
# raw : (c=1, roi)
)
with gp.build(source):
raw_roi = source.spec[raw].roi
logger.info(f"raw_roi: {raw_roi}")
pipeline = (
source + gp.Normalize(raw, factor=self.params['norm_factor']) +
gp.Pad(raw, context) + gp.PreCache() +
gp.torch.Predict(self.model,
inputs={'raw': raw},
outputs={0: embs},
array_specs={embs: gp.ArraySpec(roi=raw_roi)}))
pipeline = (pipeline +
gp.ZarrWrite({
embs: 'embs',
},
output_dir=self.curr_log_dir,
output_filename=self.dataset + '_embs.zarr',
compression_type='gzip') + gp.Scan(request))
return pipeline, request, embs
def build_pipeline(
data_dir,
model,
save_every,
batch_size,
input_size,
output_size,
raw,
labels,
affs,
affs_predicted,
lr=1e-5):
dataset_shape = zarr.open(str(data_dir))['train/raw'].shape
num_samples = dataset_shape[0]
sample_size = dataset_shape[1:]
loss = torch.nn.MSELoss()
optimizer = RAdam(model.parameters(), lr=lr)
pipeline = (
gp.ZarrSource(
data_dir,
{
raw: 'train/raw',
labels: 'train/gt'
},
array_specs={
raw: gp.ArraySpec(
roi=gp.Roi((0, 0, 0), (num_samples,) + sample_size),
voxel_size=(1, 1, 1)),
labels: gp.ArraySpec(
roi=gp.Roi((0, 0, 0), (num_samples,) + sample_size),
voxel_size=(1, 1, 1))
}) +
# raw: (d=1, h, w)
# labels: (d=1, fmap_inc_factors=5h, w)
gp.RandomLocation() +
# raw: (d=1, h, w)
# labels: (d=1, h, w)
gp.AddAffinities(
affinity_neighborhood=[(0, 1, 0), (0, 0, 1)],
labels=labels,
affinities=affs) +
gp.Normalize(affs, factor=1.0) +
# raw: (d=1, h, w)
# affs: (c=2, d=1, h, w)
Squash(dim=-3) +
# get rid of z dim
# raw: (h, w)
# affs: (c=2, h, w)
AddChannelDim(raw) +
# raw: (c=1, h, w)
# affs: (c=2, h, w)
gp.PreCache() +
gp.Stack(batch_size) +
# raw: (b=10, c=1, h, w)
# affs: (b=10, c=2, h, w)
Train(
model=model,
loss=loss,
optimizer=optimizer,
inputs={'x': raw},
target=affs,
output=affs_predicted,
save_every=save_every,
log_dir='log') +
# raw: (b=10, c=1, h, w)
# affs: (b=10, c=2, h, w)
# affs_predicted: (b=10, c=2, h, w)
TransposeDims(raw,(1, 0, 2, 3)) +
TransposeDims(affs,(1, 0, 2, 3)) +
TransposeDims(affs_predicted,(1, 0, 2, 3)) +
# raw: (c=1, b=10, h, w)
# affs: (c=2, b=10, h, w)
# affs_predicted: (c=2, b=10, h, w)
RemoveChannelDim(raw) +
# raw: (b=10, h, w)
# affs: (c=2, b=10, h, w)
# affs_predicted: (c=2, b=10, h, w)
gp.Snapshot(
dataset_names={
raw: 'raw',
labels: 'labels',
affs: 'affs',
affs_predicted: 'affs_predicted'
},
every=100) +
gp.PrintProfilingStats(every=100)
)
return pipeline
def train_until(max_iteration):
in_channels = 1
num_fmaps = 12
fmap_inc_factors = 6
downsample_factors = [(1, 3, 3), (1, 3, 3), (3, 3, 3)]
unet = UNet(in_channels,
num_fmaps,
fmap_inc_factors,
downsample_factors,
constant_upsample=True)
model = Convolve(unet, 12, 1)
loss = torch.nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-6)
# start of gunpowder part:
raw = gp.ArrayKey('RAW')
points = gp.GraphKey('POINTS')
groundtruth = gp.ArrayKey('RASTER')
prediction = gp.ArrayKey('PRED_POINT')
grad = gp.ArrayKey('GRADIENT')
voxel_size = gp.Coordinate((40, 4, 4))
input_shape = (96, 430, 430)
output_shape = (60, 162, 162)
input_size = gp.Coordinate(input_shape) * voxel_size
output_size = gp.Coordinate(output_shape) * voxel_size
request = gp.BatchRequest()
request.add(raw, input_size)
request.add(points, output_size)
request.add(groundtruth, output_size)
request.add(prediction, output_size)
request.add(grad, output_size)
pos_sources = tuple(
gp.ZarrSource(filename, {raw: 'volumes/raw'},
{raw: gp.ArraySpec(interpolatable=True)}) +
AddCenterPoint(points, raw) + gp.Pad(raw, None) +
gp.RandomLocation(ensure_nonempty=points)
for filename in pos_samples) + gp.RandomProvider()
neg_sources = tuple(
gp.ZarrSource(filename, {raw: 'volumes/raw'},
{raw: gp.ArraySpec(interpolatable=True)}) +
AddNoPoint(points, raw) + gp.RandomLocation()
for filename in neg_samples) + gp.RandomProvider()
data_sources = (pos_sources, neg_sources)
data_sources += gp.RandomProvider(probabilities=[0.9, 0.1])
data_sources += gp.Normalize(raw)
train_pipeline = data_sources
train_pipeline += gp.ElasticAugment(control_point_spacing=[4, 40, 40],
jitter_sigma=[0, 2, 2],
rotation_interval=[0, math.pi / 2.0],
prob_slip=0.05,
prob_shift=0.05,
max_misalign=10,
subsample=8)
train_pipeline += gp.SimpleAugment(transpose_only=[1, 2])
train_pipeline += gp.IntensityAugment(raw, 0.9, 1.1, -0.1, 0.1, \
z_section_wise=True)
train_pipeline += gp.RasterizePoints(
points,
groundtruth,
array_spec=gp.ArraySpec(voxel_size=voxel_size),
settings=gp.RasterizationSettings(radius=(100, 100, 100), mode='peak'))
train_pipeline += gp.PreCache(cache_size=40, num_workers=10)
train_pipeline += Reshape(raw, (1, 1) + input_shape)
train_pipeline += Reshape(groundtruth, (1, 1) + output_shape)
train_pipeline += gp_torch.Train(model=model,
loss=loss,
optimizer=optimizer,
inputs={'x': raw},
outputs={0: prediction},
loss_inputs={
0: prediction,
1: groundtruth
},
gradients={0: grad},
save_every=1000,
log_dir='log')
train_pipeline += Reshape(raw, input_shape)
train_pipeline += Reshape(groundtruth, output_shape)
train_pipeline += Reshape(prediction, output_shape)
train_pipeline += Reshape(grad, output_shape)
train_pipeline += gp.Snapshot(
{
raw: 'volumes/raw',
groundtruth: 'volumes/groundtruth',
prediction: 'volumes/prediction',
grad: 'volumes/gradient'
},
every=500,
output_filename='test_{iteration}.hdf')
train_pipeline += gp.PrintProfilingStats(every=10)
with gp.build(train_pipeline):
for i in range(max_iteration):
train_pipeline.request_batch(request)
def create_pipeline_2d(task, predictor, optimizer, batch_size, outdir,
snapshot_every):
raw_channels = task.data.raw.num_channels
filename = task.data.raw.train.filename
input_shape = predictor.input_shape
output_shape = predictor.output_shape
dataset_shape = task.data.raw.train.shape
dataset_roi = task.data.raw.train.roi
voxel_size = task.data.raw.train.voxel_size
# switch to world units
input_size = voxel_size * input_shape
output_size = voxel_size * output_shape
raw = gp.ArrayKey('RAW')
gt = gp.ArrayKey('GT')
target = gp.ArrayKey('TARGET')
weights = gp.ArrayKey('WEIGHTS')
prediction = gp.ArrayKey('PREDICTION')
channel_dims = 0 if raw_channels == 1 else 1
data_dims = len(dataset_shape) - channel_dims
if data_dims == 3:
num_samples = dataset_shape[0]
sample_shape = dataset_shape[channel_dims + 1:]
else:
raise RuntimeError("For 2D training, please provide a 3D array where "
"the first dimension indexes the samples.")
sample_shape = gp.Coordinate(sample_shape)
sample_size = sample_shape * voxel_size
# overwrite source ROI to treat samples as z dimension
spec = gp.ArraySpec(roi=gp.Roi((0, ) + dataset_roi.get_begin(),
(num_samples, ) + sample_size),
voxel_size=(1, ) + voxel_size)
sources = (task.data.raw.train.get_source(raw, overwrite_spec=spec),
task.data.gt.train.get_source(gt, overwrite_spec=spec))
pipeline = sources + gp.MergeProvider()
pipeline += gp.Pad(raw, None)
pipeline += gp.Normalize(raw)
# raw: ([c,] d=1, h, w)
# gt: ([c,] d=1, h, w)
pipeline += gp.RandomLocation()
# raw: ([c,] d=1, h, w)
# gt: ([c,] d=1, h, w)
for augmentation in eval(task.augmentations):
pipeline += augmentation
pipeline += predictor.add_target(gt, target)
# (don't care about gt anymore)
# raw: ([c,] d=1, h, w)
# target: ([c,] d=1, h, w)
weights_node = task.loss.add_weights(target, weights)
if weights_node:
pipeline += weights_node
loss_inputs = {0: prediction, 1: target, 2: weights}
else:
loss_inputs = {0: prediction, 1: target}
# raw: ([c,] d=1, h, w)
# target: ([c,] d=1, h, w)
# [weights: ([c,] d=1, h, w)]
# get rid of z dim:
pipeline += Squash(dim=-3)
# raw: ([c,] h, w)
# target: ([c,] h, w)
# [weights: ([c,] h, w)]
if channel_dims == 0:
pipeline += AddChannelDim(raw)
# raw: (c, h, w)
# target: ([c,] h, w)
# [weights: ([c,] h, w)]
pipeline += gp.PreCache()
pipeline += gp.Stack(batch_size)
# raw: (b, c, h, w)
# target: (b, [c,] h, w)
# [weights: (b, [c,] h, w)]
pipeline += gp_torch.Train(model=predictor,
loss=task.loss,
optimizer=optimizer,
inputs={'x': raw},
loss_inputs=loss_inputs,
outputs={0: prediction},
save_every=1e6)
# raw: (b, c, h, w)
# target: (b, [c,] h, w)
# [weights: (b, [c,] h, w)]
# prediction: (b, [c,] h, w)
if snapshot_every > 0:
# get channels first
pipeline += TransposeDims(raw, (1, 0, 2, 3))
if predictor.target_channels > 0:
pipeline += TransposeDims(target, (1, 0, 2, 3))
if weights_node:
pipeline += TransposeDims(weights, (1, 0, 2, 3))
if predictor.prediction_channels > 0:
pipeline += TransposeDims(prediction, (1, 0, 2, 3))
# raw: (c, b, h, w)
# target: ([c,] b, h, w)
# [weights: ([c,] b, h, w)]
# prediction: ([c,] b, h, w)
if channel_dims == 0:
pipeline += RemoveChannelDim(raw)
# raw: ([c,] b, h, w)
# target: ([c,] b, h, w)
# [weights: ([c,] b, h, w)]
# prediction: ([c,] b, h, w)
pipeline += gp.Snapshot(dataset_names={
raw: 'raw',
target: 'target',
prediction: 'prediction',
weights: 'weights'
},
every=snapshot_every,
output_dir=os.path.join(outdir, 'snapshots'),
output_filename="{iteration}.hdf")
pipeline += gp.PrintProfilingStats(every=100)
request = gp.BatchRequest()
request.add(raw, input_size)
request.add(gt, output_size)
request.add(target, output_size)
if weights_node:
request.add(weights, output_size)
request.add(prediction, output_size)
return pipeline, request
def train(n_iterations):
point_trees = gp.PointsKey("POINT_TREES")
labels = gp.ArrayKey("LABELS")
raw = gp.ArrayKey("RAW")
gt_fg = gp.ArrayKey("GT_FG")
embedding = gp.ArrayKey("EMBEDDING")
fg = gp.ArrayKey("FG")
maxima = gp.ArrayKey("MAXIMA")
gradient_embedding = gp.ArrayKey("GRADIENT_EMBEDDING")
gradient_fg = gp.ArrayKey("GRADIENT_FG")
emst = gp.ArrayKey("EMST")
edges_u = gp.ArrayKey("EDGES_U")
edges_v = gp.ArrayKey("EDGES_V")
request = gp.BatchRequest()
request.add(raw, INPUT_SHAPE, voxel_size=gp.Coordinate((1, 1)))
request.add(labels, OUTPUT_SHAPE, voxel_size=gp.Coordinate((1, 1)))
request.add(point_trees, INPUT_SHAPE)
snapshot_request = gp.BatchRequest()
snapshot_request.add(raw, INPUT_SHAPE)
snapshot_request.add(embedding, OUTPUT_SHAPE, voxel_size=gp.Coordinate((1, 1)))
snapshot_request.add(fg, OUTPUT_SHAPE, voxel_size=gp.Coordinate((1, 1)))
snapshot_request.add(gt_fg, OUTPUT_SHAPE, voxel_size=gp.Coordinate((1, 1)))
snapshot_request.add(maxima, OUTPUT_SHAPE, voxel_size=gp.Coordinate((1, 1)))
snapshot_request.add(
gradient_embedding, OUTPUT_SHAPE, voxel_size=gp.Coordinate((1, 1))
)
snapshot_request.add(gradient_fg, OUTPUT_SHAPE, voxel_size=gp.Coordinate((1, 1)))
snapshot_request[emst] = gp.ArraySpec()
snapshot_request[edges_u] = gp.ArraySpec()
snapshot_request[edges_v] = gp.ArraySpec()
pipeline = (
nl.SyntheticLightLike(
point_trees,
dims=2,
r=SKEL_GEN_RADIUS,
n_obj=N_OBJS,
thetas=THETAS,
split_ps=SPLIT_PS,
)
# + gp.SimpleAugment()
# + gp.ElasticAugment([10, 10], [0.1, 0.1], [0, 2.0 * math.pi], spatial_dims=2)
+ nl.RasterizeSkeleton(
point_trees,
labels,
gp.ArraySpec(
roi=gp.Roi((None,) * 2, (None,) * 2),
voxel_size=gp.Coordinate((1, 1)),
dtype=np.uint64,
),
)
+ gp.Copy(labels, raw)
+ nl.GrowLabels(labels, radii=LABEL_RADII)
+ nl.GrowLabels(raw, radii=RAW_RADII)
+ LabelToFloat32(raw, intensities=RAW_INTENSITIES)
+ gp.NoiseAugment(raw, var=NOISE_VAR)
+ gp.PreCache(cache_size=100, num_workers=20)
+ gp.tensorflow.Train(
"train_net",
optimizer=add_loss,
loss=None,
inputs={tensor_names["raw"]: raw, tensor_names["gt_labels"]: labels},
outputs={
tensor_names["embedding"]: embedding,
tensor_names["fg"]: fg,
"maxima:0": maxima,
"gt_fg:0": gt_fg,
emst_name: emst,
edges_u_name: edges_u,
edges_v_name: edges_v,
},
gradients={
tensor_names["embedding"]: gradient_embedding,
tensor_names["fg"]: gradient_fg,
},
)
+ gp.Snapshot(
output_filename="{iteration}.hdf",
dataset_names={
raw: "volumes/raw",
labels: "volumes/labels",
point_trees: "point_trees",
embedding: "volumes/embedding",
fg: "volumes/fg",
maxima: "volumes/maxima",
gt_fg: "volumes/gt_fg",
gradient_embedding: "volumes/gradient_embedding",
gradient_fg: "volumes/gradient_fg",
emst: "emst",
edges_u: "edges_u",
edges_v: "edges_v",
},
dataset_dtypes={maxima: np.float32, gt_fg: np.float32},
every=100,
additional_request=snapshot_request,
)
# + gp.PrintProfilingStats(every=10)
)
with gp.build(pipeline):
for i in range(n_iterations):
pipeline.request_batch(request)
mask: 'volumes/masks'
})
# define output snapshot
snapshot_request = BatchRequest()
snapshot_request.add(grad, (output_shape[0] * voxel_size[0], output_shape[1] *
voxel_size[1], output_shape[2] * voxel_size[2]))
snapshot_request.add(prediction,
(output_shape[0] * voxel_size[0], output_shape[1] *
voxel_size[1], output_shape[2] * voxel_size[2]))
# define pipeline
training_pipeline = (
source + gp.RandomLocation() +
# gp.SimpleAugment() +
gp.PreCache(cache_size=300, num_workers=80) +
# uses metagraph file 'unet' which contains 3D u-net
gp.tensorflow.Train('unet',
optimizer=optimizer.name,
loss=loss.name,
save_every=100,
checkpoint_dir="./checkpoints/",
inputs={
raw_tf.name: raw,
gt_binaries.name: mask,
},
outputs={binary.name: prediction},
gradients={unet.name: grad}) + Snapshot(
{
raw: 'volumes/raw',
gt: 'volumes/gt',
def train_until(max_iteration):
# get the latest checkpoint
if tf.train.latest_checkpoint('.'):
trained_until = int(tf.train.latest_checkpoint('.').split('_')[-1])
else:
trained_until = 0
if trained_until >= max_iteration:
return
# array keys for fused volume
raw = gp.ArrayKey('RAW')
labels = gp.ArrayKey('LABELS')
labels_fg = gp.ArrayKey('LABELS_FG')
# array keys for base volume
raw_base = gp.ArrayKey('RAW_BASE')
labels_base = gp.ArrayKey('LABELS_BASE')
swc_base = gp.PointsKey('SWC_BASE')
swc_center_base = gp.PointsKey('SWC_CENTER_BASE')
# array keys for add volume
raw_add = gp.ArrayKey('RAW_ADD')
labels_add = gp.ArrayKey('LABELS_ADD')
swc_add = gp.PointsKey('SWC_ADD')
swc_center_add = gp.PointsKey('SWC_CENTER_ADD')
# output data
fg = gp.ArrayKey('FG')
gradient_fg = gp.ArrayKey('GRADIENT_FG')
loss_weights = gp.ArrayKey('LOSS_WEIGHTS')
voxel_size = gp.Coordinate((3, 3, 3))
input_size = gp.Coordinate(net_config['input_shape']) * voxel_size
output_size = gp.Coordinate(net_config['output_shape']) * voxel_size
# add request
request = gp.BatchRequest()
request.add(raw, input_size)
request.add(labels, output_size)
request.add(labels_fg, output_size)
request.add(loss_weights, output_size)
request.add(swc_center_base, output_size)
request.add(swc_base, input_size)
request.add(swc_center_add, output_size)
request.add(swc_add, input_size)
# add snapshot request
snapshot_request = gp.BatchRequest()
snapshot_request.add(fg, output_size)
snapshot_request.add(labels_fg, output_size)
snapshot_request.add(gradient_fg, output_size)
snapshot_request.add(raw_base, input_size)
snapshot_request.add(raw_add, input_size)
snapshot_request.add(labels_base, input_size)
snapshot_request.add(labels_add, input_size)
# data source for "base" volume
data_sources_base = tuple()
data_sources_base += tuple(
(gp.Hdf5Source(file,
datasets={
raw_base: '/volume',
},
array_specs={
raw_base:
gp.ArraySpec(interpolatable=True,
voxel_size=voxel_size,
dtype=np.uint16),
},
channels_first=False),
SwcSource(filename=file,
dataset='/reconstruction',
points=(swc_center_base, swc_base),
scale=voxel_size)) + gp.MergeProvider() +
gp.RandomLocation(ensure_nonempty=swc_center_base) + RasterizeSkeleton(
points=swc_base,
array=labels_base,
array_spec=gp.ArraySpec(
interpolatable=False, voxel_size=voxel_size, dtype=np.uint32),
iteration=10) for file in files)
data_sources_base += gp.RandomProvider()
# data source for "add" volume
data_sources_add = tuple()
data_sources_add += tuple(
(gp.Hdf5Source(file,
datasets={
raw_add: '/volume',
},
array_specs={
raw_add:
gp.ArraySpec(interpolatable=True,
voxel_size=voxel_size,
dtype=np.uint16),
},
channels_first=False),
SwcSource(filename=file,
dataset='/reconstruction',
points=(swc_center_add, swc_add),
scale=voxel_size)) + gp.MergeProvider() +
gp.RandomLocation(ensure_nonempty=swc_center_add) + RasterizeSkeleton(
points=swc_add,
array=labels_add,
array_spec=gp.ArraySpec(
interpolatable=False, voxel_size=voxel_size, dtype=np.uint32),
iteration=1) for file in files)
data_sources_add += gp.RandomProvider()
data_sources = tuple([data_sources_base, data_sources_add
]) + gp.MergeProvider()
pipeline = (
data_sources + FusionAugment(raw_base,
raw_add,
labels_base,
labels_add,
raw,
labels,
blend_mode='labels_mask',
blend_smoothness=10,
num_blended_objects=0) +
# augment
gp.ElasticAugment([10, 10, 10], [1, 1, 1], [0, math.pi / 2.0],
subsample=8) +
gp.SimpleAugment(mirror_only=[2], transpose_only=[]) +
gp.Normalize(raw) + gp.IntensityAugment(raw, 0.9, 1.1, -0.001, 0.001) +
BinarizeGt(labels, labels_fg) +
gp.BalanceLabels(labels_fg, loss_weights) +
# train
gp.PreCache(cache_size=40, num_workers=10) +
gp.tensorflow.Train('./train_net',
optimizer=net_names['optimizer'],
loss=net_names['loss'],
inputs={
net_names['raw']: raw,
net_names['labels_fg']: labels_fg,
net_names['loss_weights']: loss_weights,
},
outputs={
net_names['fg']: fg,
},
gradients={
net_names['fg']: gradient_fg,
},
save_every=100) +
# visualize
gp.Snapshot(output_filename='snapshot_{iteration}.hdf',
dataset_names={
raw: 'volumes/raw',
raw_base: 'volumes/raw_base',
raw_add: 'volumes/raw_add',
labels: 'volumes/labels',
labels_base: 'volumes/labels_base',
labels_add: 'volumes/labels_add',
fg: 'volumes/fg',
labels_fg: 'volumes/labels_fg',
gradient_fg: 'volumes/gradient_fg',
},
additional_request=snapshot_request,
every=10) + gp.PrintProfilingStats(every=100))
with gp.build(pipeline):
print("Starting training...")
for i in range(max_iteration - trained_until):
pipeline.request_batch(request)
def train(iterations):
##################
# DECLARE ARRAYS #
##################
# raw intensities
raw = gp.ArrayKey('RAW')
# objects labelled with unique IDs
gt_labels = gp.ArrayKey('LABELS')
# array of per-voxel affinities to direct neighbors
gt_affs = gp.ArrayKey('AFFINITIES')
# weights to use to balance the loss
loss_weights = gp.ArrayKey('LOSS_WEIGHTS')
# the predicted affinities
pred_affs = gp.ArrayKey('PRED_AFFS')
# the gredient of the loss wrt to the predicted affinities
pred_affs_gradients = gp.ArrayKey('PRED_AFFS_GRADIENTS')
####################
# DECLARE REQUESTS #
####################
with open('train_net_config.json', 'r') as f:
net_config = json.load(f)
# get the input and output size in world units (nm, in this case)
voxel_size = gp.Coordinate((8, 8, 8))
input_size = gp.Coordinate(net_config['input_shape']) * voxel_size
output_size = gp.Coordinate(net_config['output_shape']) * voxel_size
# formulate the request for what a batch should (at least) contain
request = gp.BatchRequest()
request.add(raw, input_size)
request.add(gt_affs, output_size)
request.add(loss_weights, output_size)
# when we make a snapshot for inspection (see below), we also want to
# request the predicted affinities and gradients of the loss wrt the
# affinities
snapshot_request = gp.BatchRequest()
snapshot_request[pred_affs] = request[gt_affs]
snapshot_request[pred_affs_gradients] = request[gt_affs]
##############################
# ASSEMBLE TRAINING PIPELINE #
##############################
pipeline = (
# a tuple of sources, one for each sample (A, B, and C) provided by the
# CREMI challenge
tuple(
# read batches from the HDF5 file
gp.Hdf5Source(os.path.join(data_dir, 'fib.hdf'),
datasets={
raw: 'volumes/raw',
gt_labels: 'volumes/labels/neuron_ids'
}) +
# convert raw to float in [0, 1]
gp.Normalize(raw) +
# chose a random location for each requested batch
gp.RandomLocation()) +
# chose a random source (i.e., sample) from the above
gp.RandomProvider() +
# elastically deform the batch
gp.ElasticAugment([8, 8, 8], [0, 2, 2], [0, math.pi / 2.0],
prob_slip=0.05,
prob_shift=0.05,
max_misalign=25) +
# apply transpose and mirror augmentations
gp.SimpleAugment(transpose_only=[1, 2]) +
# scale and shift the intensity of the raw array
gp.IntensityAugment(raw,
scale_min=0.9,
scale_max=1.1,
shift_min=-0.1,
shift_max=0.1,
z_section_wise=True) +
# grow a boundary between labels
gp.GrowBoundary(gt_labels, steps=3, only_xy=True) +
# convert labels into affinities between voxels
gp.AddAffinities([[-1, 0, 0], [0, -1, 0], [0, 0, -1]], gt_labels,
gt_affs) +
# create a weight array that balances positive and negative samples in
# the affinity array
gp.BalanceLabels(gt_affs, loss_weights) +
# pre-cache batches from the point upstream
gp.PreCache(cache_size=10, num_workers=5) +
# perform one training iteration for each passing batch (here we use
# the tensor names earlier stored in train_net.config)
gp.tensorflow.Train(
'train_net',
net_config['optimizer'],
net_config['loss'],
inputs={
net_config['raw']: raw,
net_config['gt_affs']: gt_affs,
net_config['loss_weights']: loss_weights
},
outputs={net_config['pred_affs']: pred_affs},
gradients={net_config['pred_affs']: pred_affs_gradients},
save_every=10000) +
# save the passing batch as an HDF5 file for inspection
gp.Snapshot(
{
raw: '/volumes/raw',
gt_labels: '/volumes/labels/neuron_ids',
gt_affs: '/volumes/labels/affs',
pred_affs: '/volumes/pred_affs',
pred_affs_gradients: '/volumes/pred_affs_gradients'
},
output_dir='snapshots',
output_filename='batch_{iteration}.hdf',
every=1000,
additional_request=snapshot_request,
compression_type='gzip') +
# show a summary of time spend in each node every 10 iterations
gp.PrintProfilingStats(every=1000))
#########
# TRAIN #
#########
print("Training for", iterations, "iterations")
with gp.build(pipeline):
for i in range(iterations):
pipeline.request_batch(request)
print("Finished")
def build_batch_provider(self, datasets, model, task, snapshot_container=None):
input_shape = Coordinate(model.input_shape)
output_shape = Coordinate(model.output_shape)
# get voxel sizes
raw_voxel_size = datasets[0].raw.voxel_size
prediction_voxel_size = model.scale(raw_voxel_size)
# define input and output size:
# switch to world units
input_size = raw_voxel_size * input_shape
output_size = prediction_voxel_size * output_shape
# padding of groundtruth/mask
gt_mask_padding = output_size + task.predictor.padding(prediction_voxel_size)
# define keys:
raw_key = gp.ArrayKey("RAW")
gt_key = gp.ArrayKey("GT")
mask_key = gp.ArrayKey("MASK")
target_key = gp.ArrayKey("TARGET")
weight_key = gp.ArrayKey("WEIGHT")
# Get source nodes
dataset_sources = []
for dataset in datasets:
raw_source = DaCapoArraySource(dataset.raw, raw_key)
raw_source += gp.Pad(raw_key, None, 0)
gt_source = DaCapoArraySource(dataset.gt, gt_key)
gt_source += gp.Pad(gt_key, gt_mask_padding, 0)
if dataset.mask is not None:
mask_source = DaCapoArraySource(dataset.mask, mask_key)
else:
# Always provide a mask. By default it is simply an array
# of ones with the same shape/roi as gt. Avoids making us
# specially handle no mask case and allows padding of the
# ground truth without worrying about training on incorrect
# data.
mask_source = DaCapoArraySource(OnesArray.like(dataset.gt), mask_key)
mask_source += gp.Pad(mask_key, gt_mask_padding, 0)
array_sources = [raw_source, gt_source, mask_source]
dataset_source = (
tuple(array_sources) + gp.MergeProvider() + gp.RandomLocation()
)
dataset_sources.append(dataset_source)
pipeline = tuple(dataset_sources) + gp.RandomProvider()
for augment in self.augments:
pipeline += augment.node(raw_key, gt_key, mask_key)
pipeline += gp.Reject(mask_key, min_masked=self.min_masked)
# Add predictor nodes to pipeline
pipeline += DaCapoTargetFilter(
task.predictor,
gt_key=gt_key,
target_key=target_key,
weights_key=weight_key,
mask_key=mask_key,
)
# Trainer attributes:
if self.num_data_fetchers > 1:
pipeline += gp.PreCache(num_workers=self.num_data_fetchers)
# stack to create a batch dimension
pipeline += gp.Stack(self.batch_size)
# print profiling stats
pipeline += gp.PrintProfilingStats(every=self.print_profiling)
# generate request for all necessary inputs to training
request = gp.BatchRequest()
request.add(raw_key, input_size)
request.add(target_key, output_size)
request.add(weight_key, output_size)
# request additional keys for snapshots
request.add(gt_key, output_size)
request.add(mask_key, output_size)
self._request = request
self._pipeline = pipeline
self._raw_key = raw_key
self._gt_key = gt_key
self._mask_key = mask_key
self._weight_key = weight_key
self._target_key = target_key
self._loss = task.loss
self.snapshot_container = snapshot_container
def create_train_pipeline(self, model):
print(f"Creating training pipeline with batch size \
{self.params['batch_size']}")
filename = self.params['data_file']
raw_dataset = self.params['dataset']['train']['raw']
gt_dataset = self.params['dataset']['train']['gt']
optimizer = self.params['optimizer'](model.parameters(),
**self.params['optimizer_kwargs'])
raw = gp.ArrayKey('RAW')
gt_labels = gp.ArrayKey('LABELS')
points = gp.GraphKey("POINTS")
locations = gp.ArrayKey("LOCATIONS")
predictions = gp.ArrayKey('PREDICTIONS')
emb = gp.ArrayKey('EMBEDDING')
raw_data = daisy.open_ds(filename, raw_dataset)
source_roi = gp.Roi(raw_data.roi.get_offset(),
raw_data.roi.get_shape())
source_voxel_size = gp.Coordinate(raw_data.voxel_size)
out_voxel_size = gp.Coordinate(raw_data.voxel_size)
# Get in and out shape
in_shape = gp.Coordinate(model.in_shape)
out_roi = gp.Coordinate(model.base_encoder.out_shape[2:])
is_2d = in_shape.dims() == 2
in_shape = in_shape * out_voxel_size
out_roi = out_roi * out_voxel_size
out_shape = gp.Coordinate(
(self.params["num_points"], *model.out_shape[2:]))
context = (in_shape - out_roi) / 2
gt_labels_out_shape = out_roi
# Add fake 3rd dim
if is_2d:
source_voxel_size = gp.Coordinate((1, *source_voxel_size))
source_roi = gp.Roi((0, *source_roi.get_offset()),
(raw_data.shape[0], *source_roi.get_shape()))
context = gp.Coordinate((0, *context))
gt_labels_out_shape = (1, *gt_labels_out_shape)
points_roi = out_voxel_size * tuple((*self.params["point_roi"], ))
points_pad = (0, *points_roi)
context = gp.Coordinate((0, None, None))
else:
points_roi = source_voxel_size * tuple(self.params["point_roi"])
points_pad = points_roi
context = gp.Coordinate((None, None, None))
logger.info(f"source roi: {source_roi}")
logger.info(f"in_shape: {in_shape}")
logger.info(f"out_shape: {out_shape}")
logger.info(f"voxel_size: {out_voxel_size}")
logger.info(f"context: {context}")
logger.info(f"out_voxel_size: {out_voxel_size}")
request = gp.BatchRequest()
request.add(raw, in_shape)
request.add(points, points_roi)
request.add(gt_labels, out_roi)
request[locations] = gp.ArraySpec(nonspatial=True)
request[predictions] = gp.ArraySpec(nonspatial=True)
snapshot_request = gp.BatchRequest()
snapshot_request[emb] = gp.ArraySpec(
roi=gp.Roi((0, ) * in_shape.dims(),
gp.Coordinate((*model.base_encoder.out_shape[2:], )) *
out_voxel_size))
source = (
(gp.ZarrSource(filename, {
raw: raw_dataset,
gt_labels: gt_dataset
},
array_specs={
raw:
gp.ArraySpec(roi=source_roi,
voxel_size=source_voxel_size,
interpolatable=True),
gt_labels:
gp.ArraySpec(roi=source_roi,
voxel_size=source_voxel_size)
}),
PointsLabelsSource(points, self.data, scale=source_voxel_size)) +
gp.MergeProvider() + gp.Pad(raw, context) +
gp.Pad(gt_labels, context) + gp.Pad(points, points_pad) +
gp.RandomLocation(ensure_nonempty=points) +
gp.Normalize(raw, self.params['norm_factor'])
# raw : (source_roi)
# gt_labels: (source_roi)
# points : (c=1, source_locations_shape)
# If 2d then source_roi = (1, input_shape) in order to select a RL
)
source = self._augmentation_pipeline(raw, source)
pipeline = (
source +
# Batches seem to be rejected because points are chosen near the
# edge of the points ROI and the augmentations remove them.
# TODO: Figure out if this is an actual issue, and if anything can
# be done.
gp.Reject(ensure_nonempty=points) + SetDtype(gt_labels, np.int64) +
# raw : (source_roi)
# gt_labels: (source_roi)
# points : (c=1, source_locations_shape)
AddChannelDim(raw) + AddChannelDim(gt_labels)
# raw : (c=1, source_roi)
# gt_labels: (c=2, source_roi)
# points : (c=1, source_locations_shape)
)
if is_2d:
pipeline = (
# Remove extra dim the 2d roi had
pipeline + RemoveSpatialDim(raw) +
RemoveSpatialDim(gt_labels) + RemoveSpatialDim(points)
# raw : (c=1, roi)
# gt_labels: (c=1, roi)
# points : (c=1, locations_shape)
)
pipeline = (
pipeline +
FillLocations(raw, points, locations, is_2d=False, max_points=1) +
gp.Stack(self.params['batch_size']) + gp.PreCache() +
# raw : (b, c=1, roi)
# gt_labels: (b, c=1, roi)
# locations: (b, c=1, locations_shape)
# (which is what train requires)
gp.torch.Train(
model,
self.loss,
optimizer,
inputs={
'raw': raw,
'points': locations
},
loss_inputs={
0: predictions,
1: gt_labels,
2: locations
},
outputs={
0: predictions,
1: emb
},
array_specs={
predictions: gp.ArraySpec(nonspatial=True),
emb: gp.ArraySpec(voxel_size=out_voxel_size)
},
checkpoint_basename=self.logdir + '/checkpoints/model',
save_every=self.params['save_every'],
log_dir=self.logdir,
log_every=self.log_every) +
# everything is 2D at this point, plus extra dimensions for
# channels and batch
# raw : (b, c=1, roi)
# gt_labels : (b, c=1, roi)
# predictions: (b, num_points)
gp.Snapshot(output_dir=self.logdir + '/snapshots',
output_filename='it{iteration}.hdf',
dataset_names={
raw: 'raw',
gt_labels: 'gt_labels',
predictions: 'predictions',
emb: 'emb'
},
additional_request=snapshot_request,
every=self.params['save_every']) +
InspectBatch('END') + gp.PrintProfilingStats(every=500))
return pipeline, request
def random_point_pairs_pipeline(model,
loss,
optimizer,
dataset,
augmentation_parameters,
point_density,
out_dir,
normalize_factor=None,
checkpoint_interval=5000,
snapshot_interval=5000):
raw_0 = gp.ArrayKey('RAW_0')
points_0 = gp.GraphKey('POINTS_0')
locations_0 = gp.ArrayKey('LOCATIONS_0')
emb_0 = gp.ArrayKey('EMBEDDING_0')
raw_1 = gp.ArrayKey('RAW_1')
points_1 = gp.GraphKey('POINTS_1')
locations_1 = gp.ArrayKey('LOCATIONS_1')
emb_1 = gp.ArrayKey('EMBEDDING_1')
# TODO parse this key from somewhere
key = 'train/raw/0'
data = daisy.open_ds(dataset.filename, key)
source_roi = gp.Roi(data.roi.get_offset(), data.roi.get_shape())
voxel_size = gp.Coordinate(data.voxel_size)
emb_voxel_size = voxel_size
# Get in and out shape
in_shape = gp.Coordinate(model.in_shape)
out_shape = gp.Coordinate(model.out_shape)
logger.info(f"source roi: {source_roi}")
logger.info(f"in_shape: {in_shape}")
logger.info(f"out_shape: {out_shape}")
logger.info(f"voxel_size: {voxel_size}")
request = gp.BatchRequest()
request.add(raw_0, in_shape)
request.add(raw_1, in_shape)
request.add(points_0, out_shape)
request.add(points_1, out_shape)
request[locations_0] = gp.ArraySpec(nonspatial=True)
request[locations_1] = gp.ArraySpec(nonspatial=True)
snapshot_request = gp.BatchRequest()
snapshot_request[emb_0] = gp.ArraySpec(roi=request[points_0].roi)
snapshot_request[emb_1] = gp.ArraySpec(roi=request[points_1].roi)
# Let's hardcode this for now
# TODO read actual number from zarr file keys
n_samples = 447
batch_size = 1
dim = 2
padding = (100, 100)
sources = []
for i in range(n_samples):
ds_key = f'train/raw/{i}'
image_sources = tuple(
gp.ZarrSource(
dataset.filename, {raw: ds_key},
{raw: gp.ArraySpec(interpolatable=True, voxel_size=(1, 1))}) +
gp.Pad(raw, None) for raw in [raw_0, raw_1])
random_point_generator = RandomPointGenerator(density=point_density,
repetitions=2)
point_sources = tuple(
(RandomPointSource(points_0,
dim,
random_point_generator=random_point_generator),
RandomPointSource(points_1,
dim,
random_point_generator=random_point_generator)))
# TODO: get augmentation parameters from some config file!
points_and_image_sources = tuple(
(img_source, point_source) + gp.MergeProvider() + \
gp.SimpleAugment() + \
gp.ElasticAugment(
spatial_dims=2,
control_point_spacing=(10, 10),
jitter_sigma=(0.0, 0.0),
rotation_interval=(0, math.pi/2)) + \
gp.IntensityAugment(r,
scale_min=0.8,
scale_max=1.2,
shift_min=-0.2,
shift_max=0.2,
clip=False) + \
gp.NoiseAugment(r, var=0.01, clip=False)
for r, img_source, point_source
in zip([raw_0, raw_1], image_sources, point_sources))
sample_source = points_and_image_sources + gp.MergeProvider()
data = daisy.open_ds(dataset.filename, ds_key)
source_roi = gp.Roi(data.roi.get_offset(), data.roi.get_shape())
sample_source += gp.Crop(raw_0, source_roi)
sample_source += gp.Crop(raw_1, source_roi)
sample_source += gp.Pad(raw_0, padding)
sample_source += gp.Pad(raw_1, padding)
sample_source += gp.RandomLocation()
sources.append(sample_source)
sources = tuple(sources)
pipeline = sources + gp.RandomProvider()
pipeline += gp.Unsqueeze([raw_0, raw_1])
pipeline += PrepareBatch(raw_0, raw_1, points_0, points_1, locations_0,
locations_1)
# How does prepare batch relate to Stack?????
pipeline += RejectArray(ensure_nonempty=locations_1)
pipeline += RejectArray(ensure_nonempty=locations_0)
# batch content
# raw_0: (1, h, w)
# raw_1: (1, h, w)
# locations_0: (n, 2)
# locations_1: (n, 2)
pipeline += gp.Stack(batch_size)
# batch content
# raw_0: (b, 1, h, w)
# raw_1: (b, 1, h, w)
# locations_0: (b, n, 2)
# locations_1: (b, n, 2)
pipeline += gp.PreCache(num_workers=10)
pipeline += gp.torch.Train(
model,
loss,
optimizer,
inputs={
'raw_0': raw_0,
'raw_1': raw_1
},
loss_inputs={
'emb_0': emb_0,
'emb_1': emb_1,
'locations_0': locations_0,
'locations_1': locations_1
},
outputs={
2: emb_0,
3: emb_1
},
array_specs={
emb_0: gp.ArraySpec(voxel_size=emb_voxel_size),
emb_1: gp.ArraySpec(voxel_size=emb_voxel_size)
},
checkpoint_basename=os.path.join(out_dir, 'model'),
save_every=checkpoint_interval)
pipeline += gp.Snapshot(
{
raw_0: 'raw_0',
raw_1: 'raw_1',
emb_0: 'emb_0',
emb_1: 'emb_1',
# locations_0 : 'locations_0',
# locations_1 : 'locations_1',
},
every=snapshot_interval,
additional_request=snapshot_request)
return pipeline, request
def train_simple_pipeline(n_iterations, setup_config, mknet_tensor_names,
loss_tensor_names):
input_shape = gp.Coordinate(setup_config["INPUT_SHAPE"])
output_shape = gp.Coordinate(setup_config["OUTPUT_SHAPE"])
voxel_size = gp.Coordinate(setup_config["VOXEL_SIZE"])
num_iterations = setup_config["NUM_ITERATIONS"]
cache_size = setup_config["CACHE_SIZE"]
num_workers = setup_config["NUM_WORKERS"]
snapshot_every = setup_config["SNAPSHOT_EVERY"]
checkpoint_every = setup_config["CHECKPOINT_EVERY"]
profile_every = setup_config["PROFILE_EVERY"]
seperate_by = setup_config["SEPERATE_BY"]
gap_crossing_dist = setup_config["GAP_CROSSING_DIST"]
match_distance_threshold = setup_config["MATCH_DISTANCE_THRESHOLD"]
point_balance_radius = setup_config["POINT_BALANCE_RADIUS"]
neuron_radius = setup_config["NEURON_RADIUS"]
samples_path = Path(setup_config["SAMPLES_PATH"])
mongo_url = setup_config["MONGO_URL"]
input_size = input_shape * voxel_size
output_size = output_shape * voxel_size
# voxels have size ~= 1 micron on z axis
# use this value to scale anything that depends on world unit distance
micron_scale = voxel_size[0]
seperate_distance = (np.array(seperate_by)).tolist()
# array keys for data sources
raw = gp.ArrayKey("RAW")
consensus = gp.PointsKey("CONSENSUS")
skeletonization = gp.PointsKey("SKELETONIZATION")
matched = gp.PointsKey("MATCHED")
labels = gp.ArrayKey("LABELS")
labels_fg = gp.ArrayKey("LABELS_FG")
labels_fg_bin = gp.ArrayKey("LABELS_FG_BIN")
loss_weights = gp.ArrayKey("LOSS_WEIGHTS")
# tensorflow tensors
gt_fg = gp.ArrayKey("GT_FG")
fg_pred = gp.ArrayKey("FG_PRED")
embedding = gp.ArrayKey("EMBEDDING")
fg = gp.ArrayKey("FG")
maxima = gp.ArrayKey("MAXIMA")
gradient_embedding = gp.ArrayKey("GRADIENT_EMBEDDING")
gradient_fg = gp.ArrayKey("GRADIENT_FG")
emst = gp.ArrayKey("EMST")
edges_u = gp.ArrayKey("EDGES_U")
edges_v = gp.ArrayKey("EDGES_V")
ratio_pos = gp.ArrayKey("RATIO_POS")
ratio_neg = gp.ArrayKey("RATIO_NEG")
dist = gp.ArrayKey("DIST")
num_pos_pairs = gp.ArrayKey("NUM_POS")
num_neg_pairs = gp.ArrayKey("NUM_NEG")
# add request
request = gp.BatchRequest()
request.add(labels_fg, output_size)
request.add(labels_fg_bin, output_size)
request.add(loss_weights, output_size)
request.add(raw, input_size)
request.add(labels, input_size)
request.add(matched, input_size)
request.add(skeletonization, input_size)
request.add(consensus, input_size)
# add snapshot request
snapshot_request = gp.BatchRequest()
request.add(labels_fg, output_size)
# tensorflow requests
# snapshot_request.add(raw, input_size) # input_size request for positioning
# snapshot_request.add(embedding, output_size, voxel_size=voxel_size)
# snapshot_request.add(fg, output_size, voxel_size=voxel_size)
# snapshot_request.add(gt_fg, output_size, voxel_size=voxel_size)
# snapshot_request.add(fg_pred, output_size, voxel_size=voxel_size)
# snapshot_request.add(maxima, output_size, voxel_size=voxel_size)
# snapshot_request.add(gradient_embedding, output_size, voxel_size=voxel_size)
# snapshot_request.add(gradient_fg, output_size, voxel_size=voxel_size)
# snapshot_request[emst] = gp.ArraySpec()
# snapshot_request[edges_u] = gp.ArraySpec()
# snapshot_request[edges_v] = gp.ArraySpec()
# snapshot_request[ratio_pos] = gp.ArraySpec()
# snapshot_request[ratio_neg] = gp.ArraySpec()
# snapshot_request[dist] = gp.ArraySpec()
# snapshot_request[num_pos_pairs] = gp.ArraySpec()
# snapshot_request[num_neg_pairs] = gp.ArraySpec()
data_sources = tuple(
(
gp.N5Source(
filename=str((sample /
"fluorescence-near-consensus.n5").absolute()),
datasets={raw: "volume"},
array_specs={
raw:
gp.ArraySpec(interpolatable=True,
voxel_size=voxel_size,
dtype=np.uint16)
},
),
gp.DaisyGraphProvider(
f"mouselight-{sample.name}-consensus",
mongo_url,
points=[consensus],
directed=True,
node_attrs=[],
edge_attrs=[],
),
gp.DaisyGraphProvider(
f"mouselight-{sample.name}-skeletonization",
mongo_url,
points=[skeletonization],
directed=False,
node_attrs=[],
edge_attrs=[],
),
) + gp.MergeProvider() + gp.RandomLocation(
ensure_nonempty=consensus,
ensure_centered=True,
point_balance_radius=point_balance_radius * micron_scale,
) + TopologicalMatcher(
skeletonization,
consensus,
matched,
failures=Path("matching_failures_slow"),
match_distance_threshold=match_distance_threshold * micron_scale,
max_gap_crossing=gap_crossing_dist * micron_scale,
try_complete=False,
use_gurobi=True,
) + RejectIfEmpty(matched) + RasterizeSkeleton(
points=matched,
array=labels,
array_spec=gp.ArraySpec(
interpolatable=False, voxel_size=voxel_size, dtype=np.uint32),
) + GrowLabels(labels, radii=[neuron_radius * micron_scale])
# TODO: Do these need to be scaled by world units?
+ gp.ElasticAugment(
[40, 10, 10],
[0.25, 1, 1],
[0, math.pi / 2.0],
subsample=4,
use_fast_points_transform=True,
recompute_missing_points=False,
)
# + gp.SimpleAugment(mirror_only=[1, 2], transpose_only=[1, 2])
+ gp.Normalize(raw) + gp.IntensityAugment(raw, 0.9, 1.1, -0.001, 0.001)
for sample in samples_path.iterdir()
if sample.name in ("2018-07-02", "2018-08-01"))
pipeline = (
data_sources + gp.RandomProvider() + Crop(labels, labels_fg) +
BinarizeGt(labels_fg, labels_fg_bin) +
gp.BalanceLabels(labels_fg_bin, loss_weights) +
gp.PreCache(cache_size=cache_size, num_workers=num_workers) +
gp.tensorflow.Train(
"train_net",
optimizer=create_custom_loss(mknet_tensor_names, setup_config),
loss=None,
inputs={
mknet_tensor_names["loss_weights"]: loss_weights,
mknet_tensor_names["raw"]: raw,
mknet_tensor_names["gt_labels"]: labels_fg,
},
outputs={
mknet_tensor_names["embedding"]: embedding,
mknet_tensor_names["fg"]: fg,
loss_tensor_names["fg_pred"]: fg_pred,
loss_tensor_names["maxima"]: maxima,
loss_tensor_names["gt_fg"]: gt_fg,
loss_tensor_names["emst"]: emst,
loss_tensor_names["edges_u"]: edges_u,
loss_tensor_names["edges_v"]: edges_v,
loss_tensor_names["ratio_pos"]: ratio_pos,
loss_tensor_names["ratio_neg"]: ratio_neg,
loss_tensor_names["dist"]: dist,
loss_tensor_names["num_pos_pairs"]: num_pos_pairs,
loss_tensor_names["num_neg_pairs"]: num_neg_pairs,
},
gradients={
mknet_tensor_names["embedding"]: gradient_embedding,
mknet_tensor_names["fg"]: gradient_fg,
},
save_every=checkpoint_every,
summary="Merge/MergeSummary:0",
log_dir="tensorflow_logs",
) + gp.PrintProfilingStats(every=profile_every) + gp.Snapshot(
additional_request=snapshot_request,
output_filename="snapshot_{}_{}.hdf".format(
int(np.min(seperate_distance)), "{id}"),
dataset_names={
# raw data
raw: "volumes/raw",
# labeled data
labels: "volumes/labels",
# trees
skeletonization: "points/skeletonization",
consensus: "points/consensus",
matched: "points/matched",
# output volumes
embedding: "volumes/embedding",
fg: "volumes/fg",
maxima: "volumes/maxima",
gt_fg: "volumes/gt_fg",
fg_pred: "volumes/fg_pred",
gradient_embedding: "volumes/gradient_embedding",
gradient_fg: "volumes/gradient_fg",
# output trees
emst: "emst",
edges_u: "edges_u",
edges_v: "edges_v",
# output debug data
ratio_pos: "ratio_pos",
ratio_neg: "ratio_neg",
dist: "dist",
num_pos_pairs: "num_pos_pairs",
num_neg_pairs: "num_neg_pairs",
loss_weights: "volumes/loss_weights",
},
every=snapshot_every,
))
with gp.build(pipeline):
for _ in range(num_iterations):
pipeline.request_batch(request)
def train(n_iterations, setup_config, mknet_tensor_names, loss_tensor_names):
# Network hyperparams
INPUT_SHAPE = setup_config["INPUT_SHAPE"]
OUTPUT_SHAPE = setup_config["OUTPUT_SHAPE"]
# Skeleton generation hyperparams
SKEL_GEN_RADIUS = setup_config["SKEL_GEN_RADIUS"]
THETAS = np.array(setup_config["THETAS"]) * math.pi
SPLIT_PS = setup_config["SPLIT_PS"]
NOISE_VAR = setup_config["NOISE_VAR"]
N_OBJS = setup_config["N_OBJS"]
# Skeleton variation hyperparams
LABEL_RADII = setup_config["LABEL_RADII"]
RAW_RADII = setup_config["RAW_RADII"]
RAW_INTENSITIES = setup_config["RAW_INTENSITIES"]
# Training hyperparams
CACHE_SIZE = setup_config["CACHE_SIZE"]
NUM_WORKERS = setup_config["NUM_WORKERS"]
SNAPSHOT_EVERY = setup_config["SNAPSHOT_EVERY"]
CHECKPOINT_EVERY = setup_config["CHECKPOINT_EVERY"]
point_trees = gp.PointsKey("POINT_TREES")
labels = gp.ArrayKey("LABELS")
raw = gp.ArrayKey("RAW")
gt_fg = gp.ArrayKey("GT_FG")
embedding = gp.ArrayKey("EMBEDDING")
fg = gp.ArrayKey("FG")
maxima = gp.ArrayKey("MAXIMA")
gradient_embedding = gp.ArrayKey("GRADIENT_EMBEDDING")
gradient_fg = gp.ArrayKey("GRADIENT_FG")
# tensorflow tensors
emst = gp.ArrayKey("EMST")
edges_u = gp.ArrayKey("EDGES_U")
edges_v = gp.ArrayKey("EDGES_V")
ratio_pos = gp.ArrayKey("RATIO_POS")
ratio_neg = gp.ArrayKey("RATIO_NEG")
dist = gp.ArrayKey("DIST")
num_pos_pairs = gp.ArrayKey("NUM_POS")
num_neg_pairs = gp.ArrayKey("NUM_NEG")
request = gp.BatchRequest()
request.add(raw, INPUT_SHAPE, voxel_size=gp.Coordinate((1, 1)))
request.add(labels, OUTPUT_SHAPE, voxel_size=gp.Coordinate((1, 1)))
request.add(point_trees, INPUT_SHAPE)
snapshot_request = gp.BatchRequest()
snapshot_request.add(raw, INPUT_SHAPE)
snapshot_request.add(embedding,
OUTPUT_SHAPE,
voxel_size=gp.Coordinate((1, 1)))
snapshot_request.add(fg, OUTPUT_SHAPE, voxel_size=gp.Coordinate((1, 1)))
snapshot_request.add(gt_fg, OUTPUT_SHAPE, voxel_size=gp.Coordinate((1, 1)))
snapshot_request.add(maxima,
OUTPUT_SHAPE,
voxel_size=gp.Coordinate((1, 1)))
snapshot_request.add(gradient_embedding,
OUTPUT_SHAPE,
voxel_size=gp.Coordinate((1, 1)))
snapshot_request.add(gradient_fg,
OUTPUT_SHAPE,
voxel_size=gp.Coordinate((1, 1)))
snapshot_request[emst] = gp.ArraySpec()
snapshot_request[edges_u] = gp.ArraySpec()
snapshot_request[edges_v] = gp.ArraySpec()
snapshot_request[ratio_pos] = gp.ArraySpec()
snapshot_request[ratio_neg] = gp.ArraySpec()
snapshot_request[dist] = gp.ArraySpec()
snapshot_request[num_pos_pairs] = gp.ArraySpec()
snapshot_request[num_neg_pairs] = gp.ArraySpec()
pipeline = (
nl.SyntheticLightLike(
point_trees,
dims=2,
r=SKEL_GEN_RADIUS,
n_obj=N_OBJS,
thetas=THETAS,
split_ps=SPLIT_PS,
)
# + gp.SimpleAugment()
# + gp.ElasticAugment([10, 10], [0.1, 0.1], [0, 2.0 * math.pi], spatial_dims=2)
+ nl.RasterizeSkeleton(
point_trees,
raw,
gp.ArraySpec(
roi=gp.Roi((None, ) * 2, (None, ) * 2),
voxel_size=gp.Coordinate((1, 1)),
dtype=np.uint64,
),
) + nl.RasterizeSkeleton(
point_trees,
labels,
gp.ArraySpec(
roi=gp.Roi((None, ) * 2, (None, ) * 2),
voxel_size=gp.Coordinate((1, 1)),
dtype=np.uint64,
),
use_component=True,
n_objs=int(setup_config["HIDE_SIGNAL"]),
) + nl.GrowLabels(labels, radii=LABEL_RADII) +
nl.GrowLabels(raw, radii=RAW_RADII) +
LabelToFloat32(raw, intensities=RAW_INTENSITIES) +
gp.NoiseAugment(raw, var=NOISE_VAR) +
gp.PreCache(cache_size=CACHE_SIZE, num_workers=NUM_WORKERS) +
gp.tensorflow.Train(
"train_net",
optimizer=create_custom_loss(mknet_tensor_names, setup_config),
loss=None,
inputs={
mknet_tensor_names["raw"]: raw,
mknet_tensor_names["gt_labels"]: labels
},
outputs={
mknet_tensor_names["embedding"]: embedding,
mknet_tensor_names["fg"]: fg,
"strided_slice_1:0": maxima,
"gt_fg:0": gt_fg,
loss_tensor_names["emst"]: emst,
loss_tensor_names["edges_u"]: edges_u,
loss_tensor_names["edges_v"]: edges_v,
loss_tensor_names["ratio_pos"]: ratio_pos,
loss_tensor_names["ratio_neg"]: ratio_neg,
loss_tensor_names["dist"]: dist,
loss_tensor_names["num_pos_pairs"]: num_pos_pairs,
loss_tensor_names["num_neg_pairs"]: num_neg_pairs,
},
gradients={
mknet_tensor_names["embedding"]: gradient_embedding,
mknet_tensor_names["fg"]: gradient_fg,
},
save_every=CHECKPOINT_EVERY,
summary="Merge/MergeSummary:0",
log_dir="tensorflow_logs",
) + gp.Snapshot(
output_filename="{iteration}.hdf",
dataset_names={
raw: "volumes/raw",
labels: "volumes/labels",
point_trees: "point_trees",
embedding: "volumes/embedding",
fg: "volumes/fg",
maxima: "volumes/maxima",
gt_fg: "volumes/gt_fg",
gradient_embedding: "volumes/gradient_embedding",
gradient_fg: "volumes/gradient_fg",
emst: "emst",
edges_u: "edges_u",
edges_v: "edges_v",
ratio_pos: "ratio_pos",
ratio_neg: "ratio_neg",
dist: "dist",
num_pos_pairs: "num_pos_pairs",
num_neg_pairs: "num_neg_pairs",
},
dataset_dtypes={
maxima: np.float32,
gt_fg: np.float32
},
every=SNAPSHOT_EVERY,
additional_request=snapshot_request,
)
# + gp.PrintProfilingStats(every=100)
)
with gp.build(pipeline):
for i in range(n_iterations + 1):
pipeline.request_batch(request)
request._update_random_seed()
def train_until(max_iteration, name='train_net', output_folder='.', clip_max=2000):
# get the latest checkpoint
if tf.train.latest_checkpoint(output_folder):
trained_until = int(tf.train.latest_checkpoint(output_folder).split('_')[-1])
else:
trained_until = 0
if trained_until >= max_iteration:
return
with open(os.path.join(output_folder, name + '_config.json'), 'r') as f:
net_config = json.load(f)
with open(os.path.join(output_folder, name + '_names.json'), 'r') as f:
net_names = json.load(f)
# array keys
raw = gp.ArrayKey('RAW')
gt_mask = gp.ArrayKey('GT_MASK')
gt_dt = gp.ArrayKey('GT_DT')
pred_dt = gp.ArrayKey('PRED_DT')
loss_gradient = gp.ArrayKey('LOSS_GRADIENT')
voxel_size = gp.Coordinate((1, 1, 1))
input_shape = gp.Coordinate(net_config['input_shape'])
output_shape = gp.Coordinate(net_config['output_shape'])
context = gp.Coordinate(input_shape - output_shape) / 2
request = gp.BatchRequest()
request.add(raw, input_shape)
request.add(gt_mask, output_shape)
request.add(gt_dt, output_shape)
snapshot_request = gp.BatchRequest()
snapshot_request.add(raw, input_shape)
snapshot_request.add(gt_mask, output_shape)
snapshot_request.add(gt_dt, output_shape)
snapshot_request.add(pred_dt, output_shape)
snapshot_request.add(loss_gradient, output_shape)
# specify data source
data_sources = tuple()
for data_file in data_files:
current_path = os.path.join(data_dir, data_file)
with h5py.File(current_path, 'r') as f:
data_sources += tuple(
gp.Hdf5Source(
current_path,
datasets={
raw: sample + '/raw',
gt_mask: sample + '/fg'
},
array_specs={
raw: gp.ArraySpec(interpolatable=True, dtype=np.uint16, voxel_size=voxel_size),
gt_mask: gp.ArraySpec(interpolatable=False, dtype=np.bool, voxel_size=voxel_size),
}
) +
Convert(gt_mask, np.uint8) +
gp.Pad(raw, context) +
gp.Pad(gt_mask, context) +
gp.RandomLocation()
for sample in f)
pipeline = (
data_sources +
gp.RandomProvider() +
gp.Reject(gt_mask, min_masked=0.005, reject_probability=1.) +
DistanceTransform(gt_mask, gt_dt, 3) +
nl.Clip(raw, 0, clip_max) +
gp.Normalize(raw, factor=1.0/clip_max) +
gp.ElasticAugment(
control_point_spacing=[20, 20, 20],
jitter_sigma=[1, 1, 1],
rotation_interval=[0, math.pi/2.0],
subsample=4) +
gp.SimpleAugment(mirror_only=[1,2], transpose_only=[1,2]) +
gp.IntensityAugment(raw, 0.9, 1.1, -0.1, 0.1) +
gp.IntensityScaleShift(raw, 2,-1) +
# train
gp.PreCache(
cache_size=40,
num_workers=5) +
gp.tensorflow.Train(
os.path.join(output_folder, name),
optimizer=net_names['optimizer'],
loss=net_names['loss'],
inputs={
net_names['raw']: raw,
net_names['gt_dt']: gt_dt,
},
outputs={
net_names['pred_dt']: pred_dt,
},
gradients={
net_names['pred_dt']: loss_gradient,
},
save_every=5000) +
# visualize
gp.Snapshot({
raw: 'volumes/raw',
gt_mask: 'volumes/gt_mask',
gt_dt: 'volumes/gt_dt',
pred_dt: 'volumes/pred_dt',
loss_gradient: 'volumes/gradient',
},
output_filename=os.path.join(output_folder, 'snapshots', 'batch_{iteration}.hdf'),
additional_request=snapshot_request,
every=2000) +
gp.PrintProfilingStats(every=500)
)
with gp.build(pipeline):
print("Starting training...")
for i in range(max_iteration - trained_until):
pipeline.request_batch(request)
def train_until(**kwargs):
print("cuda visibile devices", os.environ["CUDA_VISIBLE_DEVICES"])
if tf.train.latest_checkpoint(kwargs['output_folder']):
trained_until = int(
tf.train.latest_checkpoint(kwargs['output_folder']).split('_')[-1])
else:
trained_until = 0
if trained_until >= kwargs['max_iteration']:
return
anchor = gp.ArrayKey('ANCHOR')
raw = gp.ArrayKey('RAW')
raw_cropped = gp.ArrayKey('RAW_CROPPED')
gt_labels = gp.ArrayKey('GT_LABELS')
gt_affs = gp.ArrayKey('GT_AFFS')
pred_affs = gp.ArrayKey('PRED_AFFS')
pred_affs_gradients = gp.ArrayKey('PRED_AFFS_GRADIENTS')
with open(
os.path.join(kwargs['output_folder'],
kwargs['name'] + '_config.json'), 'r') as f:
net_config = json.load(f)
with open(
os.path.join(kwargs['output_folder'],
kwargs['name'] + '_names.json'), 'r') as f:
net_names = json.load(f)
voxel_size = gp.Coordinate(kwargs['voxel_size'])
input_shape_world = gp.Coordinate(net_config['input_shape']) * voxel_size
output_shape_world = gp.Coordinate(net_config['output_shape']) * voxel_size
# formulate the request for what a batch should (at least) contain
request = gp.BatchRequest()
# when we make a snapshot for inspection (see below), we also want to
# request the predicted affinities and gradients of the loss wrt the
# affinities
snapshot_request = gp.BatchRequest()
snapshot_request.add(raw_cropped, output_shape_world)
snapshot_request.add(pred_affs, output_shape_world)
snapshot_request.add(gt_affs, output_shape_world)
if kwargs['input_format'] != "hdf" and kwargs['input_format'] != "zarr":
raise NotImplementedError("train node for %s not implemented yet",
kwargs['input_format'])
fls = []
for f in kwargs['data_files']:
fls.append(os.path.splitext(f)[0])
ln = len(fls)
print("first 5 files: ", fls[0:4])
if kwargs['input_format'] == "hdf":
sourceNode = gp.Hdf5Source
elif kwargs['input_format'] == "zarr":
sourceNode = gp.ZarrSource
neighborhood = []
psH = np.array(kwargs['patchshape']) // 2
for i in range(-psH[0], psH[0] + 1, kwargs['patchstride'][0]):
for j in range(-psH[1], psH[1] + 1, kwargs['patchstride'][1]):
for k in range(-psH[2], psH[2] + 1, kwargs['patchstride'][2]):
neighborhood.append([i, j, k])
datasets = {
raw: 'volumes/raw',
gt_labels: 'volumes/gt_labels',
anchor: 'volumes/gt_fgbg',
}
input_specs = {
raw:
gp.ArraySpec(roi=gp.Roi((0, ) * len(input_shape_world),
input_shape_world),
interpolatable=True,
dtype=np.float32),
gt_labels:
gp.ArraySpec(roi=gp.Roi((0, ) * len(output_shape_world),
output_shape_world),
interpolatable=False,
dtype=np.uint16),
anchor:
gp.ArraySpec(roi=gp.Roi((0, ) * len(output_shape_world),
output_shape_world),
interpolatable=False,
dtype=np.uint8),
gt_affs:
gp.ArraySpec(roi=gp.Roi((0, ) * len(output_shape_world),
output_shape_world),
interpolatable=False,
dtype=np.uint8)
}
inputs = {
net_names['raw']: raw,
net_names['gt_affs']: gt_affs,
net_names['anchor']: anchor,
}
outputs = {
net_names['pred_affs']: pred_affs,
net_names['raw_cropped']: raw_cropped,
}
snapshot = {
raw_cropped: 'volumes/raw_cropped',
gt_affs: '/volumes/gt_affs',
pred_affs: '/volumes/pred_affs',
}
optimizer_args = None
if kwargs['auto_mixed_precision']:
optimizer_args = (kwargs['optimizer'], {
'args': kwargs['args'],
'kwargs': kwargs['kwargs']
})
augmentation = kwargs['augmentation']
pipeline = (
tuple(
sourceNode(
fls[t] + "." + kwargs['input_format'],
datasets=datasets,
# array_specs=array_specs
)
+ gp.Pad(raw, None)
+ gp.Pad(gt_labels, None)
# chose a random location for each requested batch
+ gp.RandomLocation()
for t in range(ln)
) +
# chose a random source (i.e., sample) from the above
gp.RandomProvider() +
# elastically deform the batch
gp.ElasticAugment(
augmentation['elastic']['control_point_spacing'],
augmentation['elastic']['jitter_sigma'],
[augmentation['elastic']['rotation_min']*np.pi/180.0,
augmentation['elastic']['rotation_max']*np.pi/180.0],
subsample=4) +
# apply transpose and mirror augmentations
gp.SimpleAugment(mirror_only=augmentation['simple'].get("mirror"),
transpose_only=augmentation['simple'].get("transpose")) +
# scale and shift the intensity of the raw array
gp.IntensityAugment(
raw,
scale_min=augmentation['intensity']['scale'][0],
scale_max=augmentation['intensity']['scale'][1],
shift_min=augmentation['intensity']['shift'][0],
shift_max=augmentation['intensity']['shift'][1],
z_section_wise=False) +
# grow a boundary between labels
gp.GrowBoundary(
gt_labels,
steps=1,
only_xy=False) +
# convert labels into affinities between voxels
gp.AddAffinities(
neighborhood,
gt_labels,
gt_affs) +
# create a weight array that balances positive and negative samples in
# the affinity array
# gp.BalanceLabels(
# gt_affs,
# loss_weights_affs) +
# pre-cache batches from the point upstream
gp.PreCache(
cache_size=kwargs['cache_size'],
num_workers=kwargs['num_workers']) +
# pre-fetch batches from the point upstream
(gp.tensorflow.TFData() \
if kwargs.get('use_tf_data') else NoOp()) +
# perform one training iteration for each passing batch (here we use
# the tensor names earlier stored in train_net.config)
gp.tensorflow.Train(
os.path.join(kwargs['output_folder'], kwargs['name']),
optimizer=net_names['optimizer'],
summary=net_names['summaries'],
log_dir=kwargs['output_folder'],
loss=net_names['loss'],
inputs=inputs,
outputs=outputs,
array_specs=input_specs,
gradients={
net_names['pred_affs']: pred_affs_gradients,
},
auto_mixed_precision=kwargs['auto_mixed_precision'],
optimizer_args=optimizer_args,
use_tf_data=kwargs['use_tf_data'],
save_every=kwargs['checkpoints'],
snapshot_every=kwargs['snapshots']) +
# save the passing batch as an HDF5 file for inspection
gp.Snapshot(
snapshot,
output_dir=os.path.join(kwargs['output_folder'], 'snapshots'),
output_filename='batch_{iteration}.hdf',
every=kwargs['snapshots'],
additional_request=snapshot_request,
compression_type='gzip') +
# show a summary of time spend in each node every 10 iterations
gp.PrintProfilingStats(every=kwargs['profiling'])
)
#########
# TRAIN #
#########
print("Starting training...")
try:
with gp.build(pipeline):
print(pipeline)
for i in range(trained_until, kwargs['max_iteration']):
start = time.time()
pipeline.request_batch(request)
time_of_iteration = time.time() - start
logger.info("Batch: iteration=%d, time=%f", i,
time_of_iteration)
# exit()
except KeyboardInterrupt:
sys.exit()
print("Training finished")
