raw:
gp.ArraySpec(
interpolatable=True, voxel_size=voxel_size, dtype=np.uint16)
},
),
MouselightSwcFileSource(
filename=str((
filename /
"consensus-neurons-with-machine-centerpoints-labelled-as-swcs/G-002.swc"
).absolute()),
points=(swcs, ),
scale=voxel_size,
transpose=(2, 1, 0),
transform_file=str((filename / "transform.txt").absolute()),
),
) + 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_to_data.iterdir()
if "2018-08-01" in filename.name)
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, 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 create_pipeline_2d(
task, data, predictor, optimizer, batch_size, outdir, snapshot_every
):
raw_channels = data.raw.num_channels
filename = data.raw.train.filename
input_shape = predictor.input_shape
output_shape = predictor.output_shape
dataset_shape = data.raw.train.shape
dataset_roi = data.raw.train.roi
voxel_size = 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")
pred_gradients = gp.ArrayKey("PRED_GRADIENTS")
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 = (
data.raw.train.get_source(raw, overwrite_spec=spec),
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},
gradients={0: pred_gradients},
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",
pred_gradients: "pred_gradients",
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)
request.add(pred_gradients, output_size)
return pipeline, 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((40, 4, 4))
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('sample_' + s + '_padded_20160501.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() for s in ['A', 'B', 'C']) +
# chose a random source (i.e., sample) from the above
gp.RandomProvider() +
# elastically deform the batch
gp.ElasticAugment([4, 40, 40], [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=1) +
# 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=100,
additional_request=snapshot_request,
compression_type='gzip') +
# show a summary of time spend in each node every 10 iterations
gp.PrintProfilingStats(every=10))
#########
# TRAIN #
#########
print("Training for", iterations, "iterations")
with gp.build(pipeline):
for i in range(iterations):
pipeline.request_batch(request)
print("Finished")
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 batch_data_aug_generator(input_path,
batch_size=12,
voxel_shape=[1, 1, 1],
input_shape=[240, 240, 4],
output_shape=[240, 240, 4],
without_background=False,
mix_output=False,
validate=False,
aug=seq):
raw = gp.ArrayKey('raw')
gt = gp.ArrayKey('ground_truth')
files = os.listdir(input_path)
files = [os.path.join(input_path, f) for f in files]
pipeline = (
tuple(
gp.ZarrSource(
files[t], # the zarr container
{
raw: 'raw',
gt: 'ground_truth'
}, # which dataset to associate to the array key
{
raw:
gp.ArraySpec(interpolatable=True,
dtype=np.dtype('float32'),
voxel_size=voxel_shape),
gt:
gp.ArraySpec(interpolatable=True,
dtype=np.dtype('float32'),
voxel_size=voxel_shape)
} # meta-information
) + gp.RandomLocation()
for t in range(len(files))) + gp.RandomProvider()
# +gp.Stack(batch_size)
)
input_size = gp.Coordinate(input_shape)
output_size = gp.Coordinate(output_shape)
request = gp.BatchRequest()
request.add(raw, input_size)
request.add(gt, input_size)
diff = input_shape[1] - output_shape[1]
diff = int(diff / 2)
max_p = input_shape[1] - diff
different_shape = diff > 0
if different_shape:
print('Difference padding: {}'.format(diff))
with gp.build(pipeline):
while 1:
b = 0
imgs = []
masks = []
while b < batch_size:
valid = False
batch = pipeline.request_batch(request)
if validate:
valid = validate_mask(batch[gt].data)
else:
valid = True
while (valid == False):
batch = pipeline.request_batch(request)
valid = validate_mask(batch[gt].data)
im = batch[raw].data
out = batch[gt].data
# im,out = augmentation(im,out,seq)
if different_shape:
out = out[diff:max_p, diff:max_p, :]
if without_background:
out = out[:, :, 1:4]
if mix_output:
out = out.argmax(axis=3).astype(float)
imgs.append(im)
masks.append(out)
b = b + 1
yield augmentation(np.asarray(imgs), np.asarray(masks), seq)
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')
pred_mask = gp.ArrayKey('PRED_MASK')
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)
snapshot_request = gp.BatchRequest()
snapshot_request.add(raw, input_shape)
snapshot_request.add(gt_mask, output_shape)
snapshot_request.add(pred_mask, 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=0.98) +
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']: gt_mask,
},
outputs={
net_names['pred']: pred_mask,
},
gradients={
},
save_every=5000) +
# visualize
gp.Snapshot({
raw: 'volumes/raw',
pred_mask: 'volumes/pred_mask',
gt_mask: 'volumes/gt_mask',
},
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
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_fgbg = gp.ArrayKey('GT_FGBG')
gt_sample_mask = gp.ArrayKey('GT_SAMPLE_MASK')
pred_code = gp.ArrayKey('PRED_CODE')
# pred_code_gradients = gp.ArrayKey('PRED_CODE_GRADIENTS')
pred_numinst = gp.ArrayKey('PRED_NUMINST')
pred_fgbg = gp.ArrayKey('PRED_FGBG')
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
raw_key = kwargs.get('raw_key', 'volumes/raw')
# 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)
else:
request.add(gt_fgbg, 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_code, output_shape_world)
# snapshot_request.add(pred_code_gradients, output_shape_world)
if kwargs['overlapping_inst']:
snapshot_request.add(pred_numinst, output_shape_world)
else:
snapshot_request.add(pred_fgbg, 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 = []
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_bf']
elif kwargs['input_format'] == "zarr":
vol = zarr.open(f, 'r')['volumes/raw_bf']
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: 'volumes/raw_bf',
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,
}
outputs = {
net_names['pred_code']: pred_code,
net_names['raw_cropped']: raw_cropped,
}
snapshot = {
raw: '/volumes/raw',
raw_cropped: 'volumes/raw_cropped',
gt_affs: '/volumes/gt_affs',
pred_code: '/volumes/pred_code',
# pred_code_gradients: '/volumes/pred_code_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[pred_numinst] = '/volumes/pred_numinst'
else:
datasets[gt_fgbg] = '/volumes/gt_fgbg'
array_specs[gt_fgbg] = gp.ArraySpec(interpolatable=False)
inputs[net_names['gt_fgbg']] = gt_fgbg
outputs[net_names['pred_fgbg']] = pred_fgbg
snapshot[pred_fgbg] = '/volumes/pred_fgbg'
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()
if kwargs['overlapping_inst']:
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()
source = (
(source_fg, source_overlap) +
# chose a random source (i.e., sample) from the above
gp.RandomProvider(probabilities=[sampling['probability_fg'],
sampling['probability_overlap']]))
else:
source = source_fg
pipeline = (
source +
# 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]) +
gp.Reject(gt_sample_mask, min_masked=0.002, reject_probability=1) +
# 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 if kwargs['overlapping_inst'] else gt_instances,
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_code']: pred_code_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 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 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 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")
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_labels = gp.ArrayKey('GT_LABELS')
gt_affs = gp.ArrayKey('GT_AFFS')
gt_fgbg = gp.ArrayKey('GT_FGBG')
# loss_weights_affs = gp.ArrayKey('LOSS_WEIGHTS_AFFS')
loss_weights_fgbg = gp.ArrayKey('LOSS_WEIGHTS_FGBG')
pred_affs = gp.ArrayKey('PRED_AFFS')
pred_fgbg = gp.ArrayKey('PRED_FGBG')
pred_affs_gradients = gp.ArrayKey('PRED_AFFS_GRADIENTS')
pred_fgbg_gradients = gp.ArrayKey('PRED_FGBG_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_labels, output_shape_world)
request.add(gt_fgbg, output_shape_world)
request.add(anchor, output_shape_world)
request.add(gt_affs, output_shape_world)
# request.add(loss_weights_affs, output_shape_world)
request.add(loss_weights_fgbg, 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)
# snapshot_request.add(pred_affs_gradients, output_shape_world)
snapshot_request.add(gt_fgbg, output_shape_world)
snapshot_request.add(pred_fgbg, output_shape_world)
# snapshot_request.add(pred_fgbg_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 = []
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(
sourceNode(
fls[t] + "." + kwargs['input_format'],
datasets={
raw: 'volumes/raw',
gt_labels: 'volumes/gt_labels',
gt_fgbg: 'volumes/gt_fgbg',
anchor: 'volumes/gt_fgbg',
},
array_specs={
raw: gp.ArraySpec(interpolatable=True),
gt_labels: gp.ArraySpec(interpolatable=False),
gt_fgbg: gp.ArraySpec(interpolatable=False),
anchor: gp.ArraySpec(interpolatable=False)
}
)
+ gp.Pad(raw, None)
+ gp.Pad(gt_labels, None)
+ gp.Pad(gt_fgbg, 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=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
gp.GrowBoundary(
gt_labels,
steps=1,
only_xy=False) +
# 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_affs) +
gp.BalanceLabels(
gt_fgbg,
loss_weights_fgbg) +
# 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_affs']: gt_affs,
net_names['gt_fgbg']: gt_fgbg,
net_names['anchor']: anchor,
net_names['gt_labels']: gt_labels,
# net_names['loss_weights_affs']: loss_weights_affs,
net_names['loss_weights_fgbg']: loss_weights_fgbg
},
outputs={
net_names['pred_affs']: pred_affs,
net_names['pred_fgbg']: pred_fgbg,
net_names['raw_cropped']: raw_cropped,
},
gradients={
net_names['pred_affs']: pred_affs_gradients,
net_names['pred_fgbg']: pred_fgbg_gradients,
},
malis=True,
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_labels: '/volumes/gt_labels',
gt_affs: '/volumes/gt_affs',
gt_fgbg: '/volumes/gt_fgbg',
pred_affs: '/volumes/pred_affs',
pred_affs_gradients: '/volumes/pred_affs_gradients',
pred_fgbg: '/volumes/pred_fgbg',
pred_fgbg_gradients: '/volumes/pred_fgbg_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(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")
voxel_size = gp.Coordinate((10, 3, 3))
input_size = gp.Coordinate(net_config["input_shape"]) * voxel_size * 2
# add request
request = gp.BatchRequest()
request.add(raw, input_size)
request.add(swcs, 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/G-002.swc"
).absolute()),
points=(swcs, ),
scale=voxel_size,
transpose=(2, 1, 0),
transform_file=str((filename / "transform.txt").absolute()),
),
) + gp.MergeProvider() + gp.RandomLocation(ensure_nonempty=swcs,
ensure_centered=True)
for filename in Path(sample_dir).iterdir()
if "2018-08-01" in filename.name)
pipeline = data_sources + gp.RandomProvider()
with gp.build(pipeline):
print("Starting training...")
for i in range(max_iteration - trained_until):
batch = pipeline.request_batch(request)
vis_points_with_array(batch[raw].data,
points_to_graph(batch[swcs].data),
np.array(voxel_size))
