class nnUNetTrainerV2(nnUNetTrainer):
"""
Info for Fabian: same as internal nnUNetTrainerV2_2
"""
def __init__(self,
plans_file,
fold,
output_folder=None,
dataset_directory=None,
batch_dice=True,
stage=None,
unpack_data=True,
deterministic=True,
fp16=False):
super().__init__(plans_file, fold, output_folder, dataset_directory,
batch_dice, stage, unpack_data, deterministic, fp16)
self.max_num_epochs = 1000
self.initial_lr = 1e-2
self.deep_supervision_scales = None
self.ds_loss_weights = None
self.pin_memory = True
def initialize(self, training=True, force_load_plans=False):
"""
- replaced get_default_augmentation with get_moreDA_augmentation
- enforce to only run this code once
- loss function wrapper for deep supervision
:param training:
:param force_load_plans:
:return:
"""
if not self.was_initialized:
maybe_mkdir_p(self.output_folder)
if force_load_plans or (self.plans is None):
self.load_plans_file()
self.process_plans(self.plans)
self.setup_DA_params()
################# Here we wrap the loss for deep supervision ############
# we need to know the number of outputs of the network
net_numpool = len(self.net_num_pool_op_kernel_sizes)
# we give each output a weight which decreases exponentially (division by 2) as the resolution decreases
# this gives higher resolution outputs more weight in the loss
weights = np.array([1 / (2**i) for i in range(net_numpool)])
# we don't use the lowest 2 outputs. Normalize weights so that they sum to 1
mask = np.array([True] + [
True if i < net_numpool - 1 else False
for i in range(1, net_numpool)
])
weights[~mask] = 0
weights = weights / weights.sum()
self.ds_loss_weights = weights
# now wrap the loss
self.loss = MultipleOutputLoss2(self.loss, self.ds_loss_weights)
################# END ###################
self.folder_with_preprocessed_data = join(
self.dataset_directory,
self.plans['data_identifier'] + "_stage%d" % self.stage)
if training:
self.dl_tr, self.dl_val = self.get_basic_generators()
if self.unpack_data:
print("unpacking dataset")
unpack_dataset(self.folder_with_preprocessed_data)
print("done")
else:
print(
"INFO: Not unpacking data! Training may be slow due to that. Pray you are not using 2d or you "
"will wait all winter for your model to finish!")
self.tr_gen, self.val_gen = get_moreDA_augmentation(
self.dl_tr,
self.dl_val,
self.data_aug_params['patch_size_for_spatialtransform'],
self.data_aug_params,
deep_supervision_scales=self.deep_supervision_scales,
pin_memory=self.pin_memory)
self.print_to_log_file("TRAINING KEYS:\n %s" %
(str(self.dataset_tr.keys())),
also_print_to_console=False)
self.print_to_log_file("VALIDATION KEYS:\n %s" %
(str(self.dataset_val.keys())),
also_print_to_console=False)
else:
pass
self.initialize_network()
self.initialize_optimizer_and_scheduler()
assert isinstance(self.network,
(SegmentationNetwork, nn.DataParallel))
else:
self.print_to_log_file(
'self.was_initialized is True, not running self.initialize again'
)
self.was_initialized = True
def initialize_network(self):
"""
- momentum 0.99
- SGD instead of Adam
- self.lr_scheduler = None because we do poly_lr
- deep supervision = True
- i am sure I forgot something here
Known issue: forgot to set neg_slope=0 in InitWeights_He; should not make a difference though
:return:
"""
if self.threeD:
conv_op = nn.Conv3d
dropout_op = nn.Dropout3d
norm_op = nn.InstanceNorm3d
else:
conv_op = nn.Conv2d
dropout_op = nn.Dropout2d
norm_op = nn.InstanceNorm2d
norm_op_kwargs = {'eps': 1e-5, 'affine': True}
dropout_op_kwargs = {'p': 0, 'inplace': True}
net_nonlin = nn.LeakyReLU
net_nonlin_kwargs = {'negative_slope': 1e-2, 'inplace': True}
self.network = Generic_UNet(
self.num_input_channels, self.base_num_features, self.num_classes,
len(self.net_num_pool_op_kernel_sizes), self.conv_per_stage, 2,
conv_op, norm_op, norm_op_kwargs, dropout_op, dropout_op_kwargs,
net_nonlin, net_nonlin_kwargs, True, False, lambda x: x,
InitWeights_He(1e-2), self.net_num_pool_op_kernel_sizes,
self.net_conv_kernel_sizes, False, True, True)
if torch.cuda.is_available():
self.network.cuda()
self.network.inference_apply_nonlin = softmax_helper
def initialize_optimizer_and_scheduler(self):
assert self.network is not None, "self.initialize_network must be called first"
self.optimizer = torch.optim.SGD(self.network.parameters(),
self.initial_lr,
weight_decay=self.weight_decay,
momentum=0.99,
nesterov=True)
self.lr_scheduler = None
def run_online_evaluation(self, output, target):
"""
due to deep supervision the return value and the reference are now lists of tensors. We only need the full
resolution output because this is what we are interested in in the end. The others are ignored
:param output:
:param target:
:return:
"""
target = target[0]
output = output[0]
return super().run_online_evaluation(output, target)
def validate(self,
do_mirroring: bool = True,
use_sliding_window: bool = True,
step_size: float = 0.5,
save_softmax: bool = True,
use_gaussian: bool = True,
overwrite: bool = True,
validation_folder_name: str = 'validation_raw',
debug: bool = False,
all_in_gpu: bool = False,
segmentation_export_kwargs: dict = None):
"""
We need to wrap this because we need to enforce self.network.do_ds = False for prediction
"""
ds = self.network.do_ds
self.network.do_ds = False
ret = super().validate(do_mirroring, use_sliding_window, step_size,
save_softmax, use_gaussian, overwrite,
validation_folder_name, debug, all_in_gpu,
segmentation_export_kwargs)
self.network.do_ds = ds
return ret
def predict_preprocessed_data_return_seg_and_softmax(
self,
data: np.ndarray,
do_mirroring: bool = True,
mirror_axes: Tuple[int] = None,
use_sliding_window: bool = True,
step_size: float = 0.5,
use_gaussian: bool = True,
pad_border_mode: str = 'constant',
pad_kwargs: dict = None,
all_in_gpu: bool = True,
verbose: bool = True) -> Tuple[np.ndarray, np.ndarray]:
"""
We need to wrap this because we need to enforce self.network.do_ds = False for prediction
"""
ds = self.network.do_ds
self.network.do_ds = False
ret = super().predict_preprocessed_data_return_seg_and_softmax(
data, do_mirroring, mirror_axes, use_sliding_window, step_size,
use_gaussian, pad_border_mode, pad_kwargs, all_in_gpu, verbose)
self.network.do_ds = ds
return ret
def run_iteration(self,
data_generator,
do_backprop=True,
run_online_evaluation=False):
"""
gradient clipping improves training stability
:param data_generator:
:param do_backprop:
:param run_online_evaluation:
:return:
"""
data_dict = next(data_generator)
data = data_dict['data']
target = data_dict['target']
data = maybe_to_torch(data)
target = maybe_to_torch(target)
if torch.cuda.is_available():
data = to_cuda(data)
target = to_cuda(target)
self.optimizer.zero_grad()
output = self.network(data)
del data
loss = self.loss(output, target)
if run_online_evaluation:
self.run_online_evaluation(output, target)
del target
if do_backprop:
if not self.fp16 or amp is None or not torch.cuda.is_available():
loss.backward()
else:
with amp.scale_loss(loss, self.optimizer) as scaled_loss:
scaled_loss.backward()
_ = clip_grad_norm_(self.network.parameters(), 12)
self.optimizer.step()
return loss.detach().cpu().numpy()
def do_split(self):
"""
we now allow more than 5 splits. IMPORTANT: and fold > 4 will not be a real split but just another random
80:20 split of the data. You cannot run X-fold cross-validation with this code. It will always be a 5-fold CV.
Folds > 4 will be independent from each other
:return:
"""
if self.fold == "all":
# if fold==all then we use all images for training and validation
tr_keys = val_keys = list(self.dataset.keys())
else:
splits_file = join(self.dataset_directory, "splits_final.pkl")
# if the split file does not exist we need to create it
if not isfile(splits_file):
self.print_to_log_file("Creating new split...")
splits = []
all_keys_sorted = np.sort(list(self.dataset.keys()))
kfold = KFold(n_splits=5, shuffle=True, random_state=12345)
for i, (train_idx,
test_idx) in enumerate(kfold.split(all_keys_sorted)):
train_keys = np.array(all_keys_sorted)[train_idx]
test_keys = np.array(all_keys_sorted)[test_idx]
splits.append(OrderedDict())
splits[-1]['train'] = train_keys
splits[-1]['val'] = test_keys
save_pickle(splits, splits_file)
splits = load_pickle(splits_file)
if self.fold < len(splits):
tr_keys = splits[self.fold]['train']
val_keys = splits[self.fold]['val']
else:
self.print_to_log_file(
"INFO: Requested fold %d but split file only has %d folds. I am now creating a "
"random 80:20 split!" % (self.fold, len(splits)))
# if we request a fold that is not in the split file, create a random 80:20 split
rnd = np.random.RandomState(seed=12345 + self.fold)
keys = np.sort(list(self.dataset.keys()))
idx_tr = rnd.choice(len(keys),
int(len(keys) * 0.8),
replace=False)
idx_val = [i for i in range(len(keys)) if i not in idx_tr]
tr_keys = [keys[i] for i in idx_tr]
val_keys = [keys[i] for i in idx_val]
tr_keys.sort()
val_keys.sort()
self.dataset_tr = OrderedDict()
for i in tr_keys:
self.dataset_tr[i] = self.dataset[i]
self.dataset_val = OrderedDict()
for i in val_keys:
self.dataset_val[i] = self.dataset[i]
def setup_DA_params(self):
"""
- we increase roation angle from [-15, 15] to [-30, 30]
- scale range is now (0.7, 1.4), was (0.85, 1.25)
- we don't do elastic deformation anymore
:return:
"""
self.deep_supervision_scales = [[1, 1, 1]] + list(
list(i) for i in 1 / np.cumprod(
np.vstack(self.net_num_pool_op_kernel_sizes), axis=0))[:-1]
if self.threeD:
self.data_aug_params = default_3D_augmentation_params
self.data_aug_params['rotation_x'] = (-30. / 360 * 2. * np.pi,
30. / 360 * 2. * np.pi)
self.data_aug_params['rotation_y'] = (-30. / 360 * 2. * np.pi,
30. / 360 * 2. * np.pi)
self.data_aug_params['rotation_z'] = (-30. / 360 * 2. * np.pi,
30. / 360 * 2. * np.pi)
if self.do_dummy_2D_aug:
self.data_aug_params["dummy_2D"] = True
self.print_to_log_file("Using dummy2d data augmentation")
self.data_aug_params["elastic_deform_alpha"] = \
default_2D_augmentation_params["elastic_deform_alpha"]
self.data_aug_params["elastic_deform_sigma"] = \
default_2D_augmentation_params["elastic_deform_sigma"]
self.data_aug_params[
"rotation_x"] = default_2D_augmentation_params[
"rotation_x"]
else:
self.do_dummy_2D_aug = False
if max(self.patch_size) / min(self.patch_size) > 1.5:
default_2D_augmentation_params['rotation_x'] = (-15. / 360 *
2. * np.pi,
15. / 360 *
2. * np.pi)
self.data_aug_params = default_2D_augmentation_params
self.data_aug_params[
"mask_was_used_for_normalization"] = self.use_mask_for_norm
if self.do_dummy_2D_aug:
self.basic_generator_patch_size = get_patch_size(
self.patch_size[1:], self.data_aug_params['rotation_x'],
self.data_aug_params['rotation_y'],
self.data_aug_params['rotation_z'],
self.data_aug_params['scale_range'])
self.basic_generator_patch_size = np.array(
[self.patch_size[0]] + list(self.basic_generator_patch_size))
patch_size_for_spatialtransform = self.patch_size[1:]
else:
self.basic_generator_patch_size = get_patch_size(
self.patch_size, self.data_aug_params['rotation_x'],
self.data_aug_params['rotation_y'],
self.data_aug_params['rotation_z'],
self.data_aug_params['scale_range'])
patch_size_for_spatialtransform = self.patch_size
self.data_aug_params["scale_range"] = (0.7, 1.4)
self.data_aug_params["do_elastic"] = False
self.data_aug_params['selected_seg_channels'] = [0]
self.data_aug_params[
'patch_size_for_spatialtransform'] = patch_size_for_spatialtransform
self.data_aug_params["num_cached_per_thread"] = 2
def maybe_update_lr(self, epoch=None):
"""
if epoch is not None we overwrite epoch. Else we use epoch = self.epoch + 1
(maybe_update_lr is called in on_epoch_end which is called before epoch is incremented.
herefore we need to do +1 here)
:param epoch:
:return:
"""
if epoch is None:
ep = self.epoch + 1
else:
ep = epoch
self.optimizer.param_groups[0]['lr'] = poly_lr(ep, self.max_num_epochs,
self.initial_lr, 0.9)
self.print_to_log_file(
"lr:", np.round(self.optimizer.param_groups[0]['lr'], decimals=6))
def on_epoch_end(self):
"""
overwrite patient-based early stopping. Always run to 1000 epochs
:return:
"""
super().on_epoch_end()
continue_training = self.epoch < self.max_num_epochs
# it can rarely happen that the momentum of nnUNetTrainerV2 is too high for some dataset. If at epoch 100 the
# estimated validation Dice is still 0 then we reduce the momentum from 0.99 to 0.95
if self.epoch == 100:
if self.all_val_eval_metrics[-1] == 0:
self.optimizer.param_groups[0]["momentum"] = 0.95
self.network.apply(InitWeights_He(1e-2))
self.print_to_log_file(
"At epoch 100, the mean foreground Dice was 0. This can be caused by a too "
"high momentum. High momentum (0.99) is good for datasets where it works, but "
"sometimes causes issues such as this one. Momentum has now been reduced to "
"0.95 and network weights have been reinitialized")
return continue_training
def run_training(self):
"""
if we run with -c then we need to set the correct lr for the first epoch, otherwise it will run the first
continued epoch with self.initial_lr
we also need to make sure deep supervision in the network is enabled for training, thus the wrapper
:return:
"""
self.maybe_update_lr(
self.epoch
) # if we dont overwrite epoch then self.epoch+1 is used which is not what we
# want at the start of the training
ds = self.network.do_ds
self.network.do_ds = True
ret = super().run_training()
self.network.do_ds = ds
return ret
class nnUNetMultiTrainerV2(nnUNetTrainer):
"""
Info for Fabian: same as internal nnUNetTrainerV2_2
"""
def __init__(self,
plans_file,
fold,
tasks,
tags,
output_folder_dict=None,
dataset_directory_dict=None,
batch_dice=True,
stage=None,
unpack_data=True,
deterministic=True,
fp16=False):
"""
:param deterministic:
:param fold: can be either [0 ... 5) for cross-validation, 'all' to train on all available training data or
None if you wish to load some checkpoint and do inference only
:param plans_file: the pkl file generated by preprocessing. This file will determine all design choices
:param subfolder_with_preprocessed_data: must be a subfolder of dataset_directory (just the name of the folder,
not the entire path). This is where the preprocessed data lies that will be used for network training. We made
this explicitly available so that differently preprocessed data can coexist and the user can choose what to use.
Can be None if you are doing inference only.
:param output_folder: where to store parameters, plot progress and to the validation
:param dataset_directory: the parent directory in which the preprocessed Task data is stored. This is required
because the split information is stored in this directory. For running prediction only this input is not
required and may be set to None
:param batch_dice: compute dice loss for each sample and average over all samples in the batch or pretend the
batch is a pseudo volume?
:param stage: The plans file may contain several stages (used for lowres / highres / pyramid). Stage must be
specified for training:
if stage 1 exists then stage 1 is the high resolution stage, otherwise it's 0
:param unpack_data: if False, npz preprocessed data will not be unpacked to npy. This consumes less space but
is considerably slower! Running unpack_data=False with 2d should never be done!
IMPORTANT: If you inherit from nnUNetTrainer and the init args change then you need to redefine self.init_args
in your init accordingly. Otherwise checkpoints won't load properly!
"""
self.fp16 = fp16
self.amp_initialized = False
self.x_tags = None
if deterministic:
np.random.seed(12345)
torch.manual_seed(12345)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(12345)
cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
else:
cudnn.deterministic = False
torch.backends.cudnn.benchmark = True
################# SET THESE IN self.initialize() ###################################
self.network: Tuple[SegmentationNetwork, nn.DataParallel] = None
self.optimizer = None
self.lr_scheduler = None
self.tr_gen = self.val_gen = None
self.was_initialized = False
################# SET THESE IN INIT ################################################
self.output_folder = None
self.fold = None
self.dataset_directory = None
################# SET THESE IN LOAD_DATASET OR DO_SPLIT ############################
self.dataset = None # these can be None for inference mode
# do not need to be used, they just appear if you are using the suggested load_dataset_and_do_split
self.dataset_tr = self.dataset_val = None
################# THESE DO NOT NECESSARILY NEED TO BE MODIFIED #####################
self.patience = 50
self.val_eval_criterion_alpha = 0.9 # alpha * old + (1-alpha) * new
# if this is too low then the moving average will be too noisy and the training may terminate early. If it is
# too high the training will take forever
self.train_loss_MA_alpha = 0.93 # alpha * old + (1-alpha) * new
# new MA must be at least this much better (smaller)
self.train_loss_MA_eps = 5e-4
self.save_every = 1
self.save_latest_only = True
self.max_num_epochs = 500
self.stage_2_start_epoch = 120
self.num_batches_per_epoch = 250
self.num_val_batches_per_epoch = 50
self.also_val_in_tr_mode = False
# the network will not terminate training if the lr is still above this threshold
self.lr_threshold = 1e-6
################# LEAVE THESE ALONE ################################################
self.val_eval_criterion_MA = None
self.train_loss_MA = None
self.best_val_eval_criterion_MA = None
self.best_MA_tr_loss_for_patience = None
self.best_epoch_based_on_MA_tr_loss = None
self.all_tr_losses = []
self.all_val_losses = []
self.all_val_losses_tr_mode = []
self.all_val_eval_metrics = [] # does not have to be used
self.epoch = 0
# self.need_updateGT = False #update p
self.log_file = None
self.deterministic = deterministic
self.use_progress_bar = True
if 'nnunet_use_progress_bar' in os.environ.keys():
self.use_progress_bar = bool(
int(os.environ['nnunet_use_progress_bar']))
#################################################################
self.unpack_data = unpack_data
self.init_args = (plans_file, fold, output_folder_dict,
dataset_directory_dict, batch_dice, stage,
unpack_data, deterministic, fp16)
# set through arguments from init
self.stage = stage
self.experiment_name = self.__class__.__name__
self.plans_file = plans_file
self.output_folder_dict = output_folder_dict
self.output_folder = output_folder_dict[tasks[0]]
self.dataset_directory_dict = dataset_directory_dict
self.output_folder_base = self.output_folder
self.fold = fold
self.tasks = tasks
self.tags = tags
self.plans = None
# if we are running inference only then the self.dataset_directory is set (due to checkpoint loading) but it
# irrelevant
self.gt_niftis_folder_dict = {}
for task in tasks:
dataset_directory = self.dataset_directory_dict[task]
if dataset_directory is not None and isdir(dataset_directory):
self.gt_niftis_folder_dict[task] = join(
dataset_directory, "gt_segmentations")
else:
self.gt_niftis_folder_dict[task] = None
self.gt_niftis_folder = self.gt_niftis_folder_dict[
self.tasks[0]] # 0 gt_nii
self.folder_with_preprocessed_data = None
# set in self.initialize()
self.dl_tr = self.dl_val = None
self.num_input_channels = self.num_classes = self.net_pool_per_axis = self.patch_size = self.batch_size = \
self.threeD = self.base_num_features = self.intensity_properties = self.normalization_schemes = \
self.net_num_pool_op_kernel_sizes = self.net_conv_kernel_sizes = None # loaded automatically from plans_file
self.basic_generator_patch_size = self.data_aug_params = self.transpose_forward = self.transpose_backward = None
self.batch_dice = batch_dice
# self.loss = None
self.online_eval_foreground_dc = []
self.online_eval_tp = []
self.online_eval_fp = []
self.online_eval_fn = []
self.classes = self.do_dummy_2D_aug = self.use_mask_for_norm = self.only_keep_largest_connected_component = \
self.min_region_size_per_class = self.min_size_per_class = None
self.inference_pad_border_mode = "constant"
self.inference_pad_kwargs = {'constant_values': 0}
self.update_fold(fold)
self.pad_all_sides = None
self.lr_scheduler_eps = 1e-3
self.lr_scheduler_patience = 30
self.initial_lr = 3e-4
self.weight_decay = 3e-5
self.oversample_foreground_percent = 0.33
self.conv_per_stage = None
self.regions_class_order = None
self.initial_lr = 1e-2
self.deep_supervision_scales = None
self.ds_loss_weights = None
self.pin_memory = True
self.loss = DC_CE_Marginal_Exclusion_loss(
{
'batch_dice': self.batch_dice,
'smooth': 1e-5,
'do_bg': False
}, {})
# self.loss = pann_loss({'batch_dice': self.batch_dice, 'smooth': 1e-5, 'do_bg': False}, {})
# self.loss = DC_and_CE_loss({'batch_dice': self.batch_dice, 'smooth': 1e-5, 'do_bg': False}, {})
def load_plans_file(self):
"""
This is what actually configures the entire experiment. The plans file is generated by experiment planning
:return:
"""
self.plans = load_pickle(self.plans_file[self.tasks[0]])
def initialize(self, training=True, force_load_plans=False):
"""
- replaced get_default_augmentation with get_moreDA_augmentation
- enforce to only run this code once
- loss function wrapper for deep supervision
:param training:
:param force_load_plans:
:return:
"""
if not self.was_initialized:
maybe_mkdir_p(self.output_folder)
if force_load_plans or (self.plans is None):
self.load_plans_file()
self.process_plans(self.plans)
self.setup_DA_params()
################# Here we wrap the loss for deep supervision ############
# we need to know the number of outputs of the network
net_numpool = len(self.net_num_pool_op_kernel_sizes)
# print("---net_numpool:", net_numpool) #Task_100 MAB 5 class
# we give each output a weight which decreases exponentially (division by 2) as the resolution decreases
# this gives higher resolution outputs more weight in the loss
weights = np.array([1 / (2**i) for i in range(net_numpool)])
# we don't use the lowest 2 outputs. Normalize weights so that they sum to 1
mask = np.array([True] + [
True if i < net_numpool - 1 else False
for i in range(1, net_numpool)
])
weights[~mask] = 0
weights = weights / weights.sum()
self.ds_loss_weights = weights
# now wrap the loss
# self.loss = MultipleOutputLoss2withTags_pann(self.loss, self.ds_loss_weights)
self.loss = MultipleOutputLoss2withTags(self.loss,
self.ds_loss_weights)
################# END ###################
self.folder_with_preprocessed_data = {}
self.dl_tr = []
self.dl_val = []
self.tr_gens = []
self.val_gens = []
for task in self.tasks:
self.folder_with_preprocessed_data[task] = join(
self.dataset_directory_dict[task],
self.plans['data_identifier'] + "_stage%d" % self.stage)
if training:
dl_tr, dl_val = self.get_basic_generators(task)
self.dl_tr.append(dl_tr)
self.dl_val.append(dl_val)
# print('%s.dl_tr raw data size:%d'%(task, len(dl_tr)))
if self.unpack_data:
print("unpacking dataset")
unpack_dataset(
self.folder_with_preprocessed_data[task])
print("done")
else:
print(
"INFO: Not unpacking data! Training may be slow due to that. Pray you are not using 2d or you "
"will wait all winter for your model to finish!")
tr_gen, val_gen = get_moreDA_augmentation( # data augmentation
dl_tr,
dl_val,
self.
data_aug_params['patch_size_for_spatialtransform'],
self.data_aug_params,
deep_supervision_scales=self.deep_supervision_scales,
pin_memory=self.pin_memory)
self.tr_gens.append(tr_gen) #tr_gen: multithreadaug..
self.val_gens.append(val_gen)
self.print_to_log_file("TRAINING KEYS:\n %s" %
(str(self.dataset_tr.keys())),
also_print_to_console=False)
print("TRAINING KEYS:\n %s" %
(str(self.dataset_tr.keys())))
self.print_to_log_file("VALIDATION KEYS:\n %s" %
(str(self.dataset_val.keys())),
also_print_to_console=False)
else:
pass
################## dataset all in ###########################
if training:
self.tr_gen = switchable_generator(self.tr_gens)
self.val_gen = switchable_generator(self.val_gens)
self.initialize_network()
self.initialize_optimizer_and_scheduler()
assert isinstance(self.network,
(SegmentationNetwork, nn.DataParallel))
else:
self.print_to_log_file(
'self.was_initialized is True, not running self.initialize again'
)
self.was_initialized = True
def initialize_network(self):
"""
- momentum 0.99
- SGD instead of Adam
- self.lr_scheduler = None because we do poly_lr
- deep supervision = True
- i am sure I forgot something here
Known issue: forgot to set neg_slope=0 in InitWeights_He; should not make a difference though
:return:
"""
if self.threeD:
conv_op = nn.Conv3d
dropout_op = nn.Dropout3d
norm_op = nn.InstanceNorm3d
else:
conv_op = nn.Conv2d
dropout_op = nn.Dropout2d
norm_op = nn.InstanceNorm2d
norm_op_kwargs = {'eps': 1e-5, 'affine': True}
dropout_op_kwargs = {'p': 0, 'inplace': True}
net_nonlin = nn.LeakyReLU
net_nonlin_kwargs = {'negative_slope': 1e-2, 'inplace': True}
self.network = Generic_UNet(
self.num_input_channels, self.base_num_features, self.num_classes,
len(self.net_num_pool_op_kernel_sizes), self.conv_per_stage, 2,
conv_op, norm_op, norm_op_kwargs, dropout_op, dropout_op_kwargs,
net_nonlin, net_nonlin_kwargs, True, False, lambda x: x,
InitWeights_He(1e-2), self.net_num_pool_op_kernel_sizes,
self.net_conv_kernel_sizes, False, True, True)
if torch.cuda.is_available():
self.network.cuda()
self.network.inference_apply_nonlin = softmax_helper
def initialize_optimizer_and_scheduler(self):
assert self.network is not None, "self.initialize_network must be called first"
self.optimizer = torch.optim.SGD(self.network.parameters(),
self.initial_lr,
weight_decay=self.weight_decay,
momentum=0.99,
nesterov=True)
self.lr_scheduler = None
def get_basic_generators(self, task):
self.load_dataset(task)
self.do_split(task)
if self.threeD:
dl_tr = DataLoader3D(self.dataset_tr,
self.basic_generator_patch_size,
self.patch_size,
self.batch_size,
self.tags[task],
False,
oversample_foreground_percent=self.
oversample_foreground_percent,
pad_mode="constant",
pad_sides=self.pad_all_sides)
dl_val = DataLoader3D(self.dataset_val,
self.patch_size,
self.patch_size,
self.batch_size,
self.tags[task],
False,
oversample_foreground_percent=self.
oversample_foreground_percent,
pad_mode="constant",
pad_sides=self.pad_all_sides)
else:
dl_tr = DataLoader2D(
self.dataset_tr,
self.basic_generator_patch_size,
self.patch_size,
self.batch_size,
self.tags[task],
# self.plans.get('transpose_forward'),
transpose=None,
oversample_foreground_percent=self.
oversample_foreground_percent,
pad_mode="constant",
pad_sides=self.pad_all_sides)
dl_val = DataLoader2D(
self.dataset_val,
self.patch_size,
self.patch_size,
self.batch_size,
self.tags[task],
# self.plans.get('transpose_forward'),
transpose=None,
oversample_foreground_percent=self.
oversample_foreground_percent,
pad_mode="constant",
pad_sides=self.pad_all_sides)
return dl_tr, dl_val
def load_dataset(self, task=None):
if task is None:
self.dataset = load_dataset(
self.folder_with_preprocessed_data[self.tasks[0]])
else:
self.dataset = load_dataset(
self.folder_with_preprocessed_data[task])
def run_online_evaluation(self, output, target):
"""
due to deep supervision the return value and the reference are now lists of tensors. We only need the full
resolution output because this is what we are interested in in the end. The others are ignored
:param output:
:param target:
:return:
"""
target = target[0]
output = output[0]
return super().run_online_evaluation(output, target)
def validate(self,
do_mirroring: bool = True,
use_sliding_window: bool = True,
step_size: float = 0.5,
save_softmax: bool = True,
use_gaussian: bool = True,
overwrite: bool = True,
validation_folder_name: str = 'validation_raw',
debug: bool = False,
all_in_gpu: bool = False,
force_separate_z: bool = None,
interpolation_order: int = 3,
interpolation_order_z=0):
"""
We need to wrap this because we need to enforce self.network.do_ds = False for prediction
"""
ds = self.network.do_ds
self.network.do_ds = False
ret = super().validate(do_mirroring,
use_sliding_window,
step_size,
save_softmax,
use_gaussian,
overwrite,
validation_folder_name,
debug,
all_in_gpu,
force_separate_z=force_separate_z,
interpolation_order=interpolation_order,
interpolation_order_z=interpolation_order_z)
self.network.do_ds = ds
return ret
def validate_specific_data(self,
task,
do_mirroring: bool = True,
use_sliding_window: bool = True,
step_size: float = 0.5,
save_softmax: bool = True,
use_gaussian: bool = True,
overwrite: bool = False,
validation_folder_name: str = 'validation_raw',
debug: bool = False,
all_in_gpu: bool = False,
force_separate_z: bool = None,
interpolation_order: int = 3,
interpolation_order_z=0):
ds = self.network.do_ds #?
self.network.do_ds = False
###########################################
current_mode = self.network.training
self.network.eval()
assert self.was_initialized, "must initialize, ideally with checkpoint (or train first)"
self.load_dataset(task)
self.do_split(task)
# predictions as they come from the network go here
self.output_folder = self.output_folder_dict[task]
self.gt_niftis_folder = self.gt_niftis_folder_dict[task]
output_folder = join(self.output_folder, 'fold_4',
validation_folder_name)
maybe_mkdir_p(output_folder)
# this is for debug purposes
my_input_args = {
'do_mirroring': do_mirroring,
'use_sliding_window': use_sliding_window,
'step_size': step_size,
'save_softmax': save_softmax,
'use_gaussian': use_gaussian,
'overwrite': overwrite,
'validation_folder_name': validation_folder_name,
'debug': debug,
'all_in_gpu': all_in_gpu, #? why not use
'force_separate_z': force_separate_z,
'interpolation_order': interpolation_order,
'interpolation_order_z': interpolation_order_z,
}
save_json(my_input_args, join(output_folder, "validation_args.json"))
if do_mirroring:
if not self.data_aug_params['do_mirror']:
raise RuntimeError(
"We did not train with mirroring so you cannot do inference with mirroring enabled"
)
mirror_axes = self.data_aug_params['mirror_axes']
else:
mirror_axes = ()
pred_gt_tuples = []
export_pool = Pool(default_num_threads)
results = []
if '104' in task: #kidney
temp_transpose = [0, 2, 3, 1]
else:
temp_transpose = [0, 1, 2, 3]
for k in self.dataset_val.keys():
properties = self.dataset[k]['properties']
fname = properties['list_of_data_files'][0].split("/")[-1][:-12]
if overwrite or (not isfile(join(output_folder, fname + ".nii.gz"))) or \
(save_softmax and not isfile(join(output_folder, fname + ".npz"))):
data = np.load(self.dataset[k]['data_file'])['data']
print(k, data.shape)
data[-1][data[-1] == -1] = 0
softmax_pred = self.predict_preprocessed_data_return_seg_and_softmax(
data[:-1],
do_mirroring,
mirror_axes,
use_sliding_window,
step_size,
use_gaussian,
all_in_gpu=all_in_gpu)[1]
softmax_pred = softmax_pred.transpose(temp_transpose)
# softmax_pred = softmax_pred.transpose([0] + [i + 1 for i in self.transpose_backward])
if save_softmax:
softmax_fname = join(output_folder, fname + ".npz")
else:
softmax_fname = None
"""There is a problem with python process communication that prevents us from communicating obejcts
larger than 2 GB between processes (basically when the length of the pickle string that will be sent is
communicated by the multiprocessing.Pipe object then the placeholder (\%i I think) does not allow for long
enough strings (lol). This could be fixed by changing i to l (for long) but that would require manually
patching system python code. We circumvent that problem here by saving softmax_pred to a npy file that will
then be read (and finally deleted) by the Process. save_segmentation_nifti_from_softmax can take either
filename or np.ndarray and will handle this automatically"""
if np.prod(softmax_pred.shape) > (
2e9 / 4 * 0.85): # *0.85 just to be save
np.save(join(output_folder, fname + ".npy"), softmax_pred)
softmax_pred = join(output_folder, fname + ".npy")
# save_segmentation_nifti_from_softmax(softmax_pred, join(output_folder, fname + ".nii.gz"),
# properties, interpolation_order, None, None, None,
# softmax_fname, None, force_separate_z,
# interpolation_order_z, task)
results.append(
export_pool.starmap_async(
save_segmentation_nifti_from_softmax,
((softmax_pred, join(output_folder, fname + ".nii.gz"),
properties, interpolation_order, None, None, None,
softmax_fname, None, force_separate_z,
interpolation_order_z, task), )))
pred_gt_tuples.append([
join(output_folder, fname + ".nii.gz"),
join(self.gt_niftis_folder, fname + ".nii.gz")
])
_ = [i.get() for i in results]
self.print_to_log_file("finished prediction")
# evaluate raw predictions
self.print_to_log_file("evaluation of raw predictions")
# task = self.dataset_directory.split("/")[-1]
job_name = self.experiment_name
# x_tags = ['rightkidney','leftkidney']
x_tags = ['liver', 'spleen', 'pancreas', 'rightkidney', 'leftkidney']
if "100" in task:
y_tags = x_tags
elif "101" in task or "105" in task:
y_tags = ['liver']
elif "102" in task:
y_tags = ['spleen']
elif "103" in task:
y_tags = ['pancreas']
elif "104" in task:
y_tags = ['rightkidney', 'leftkidney']
elif "105" in task: #PrivateLiver
y_tags = ['liver']
else:
exit()
all_score = aggregate_scores_withtags(
pred_gt_tuples,
labels=list(range(self.num_classes)),
x_tags=x_tags,
y_tags=y_tags,
json_output_file=join(output_folder, "summary.json"),
json_name=job_name + " val tiled %s" % (str(use_sliding_window)),
json_author="Fabian",
json_task=task,
num_threads=default_num_threads)
# in the old nnunet we would stop here. Now we add a postprocessing. This postprocessing can remove everything
# except the largest connected component for each class. To see if this improves results, we do this for all
# classes and then rerun the evaluation. Those classes for which this resulted in an improved dice score will
# have this applied during inference as well
# after this the final predictions for the vlaidation set can be found in validation_folder_name_base + "_postprocessed"
# They are always in that folder, even if no postprocessing as applied!
# detemining postprocesing on a per-fold basis may be OK for this fold but what if another fold finds another
# postprocesing to be better? In this case we need to consolidate. At the time the consolidation is going to be
# done we won't know what self.gt_niftis_folder was, so now we copy all the niftis into a separate folder to
# be used later
self.output_folder_base = self.output_folder
gt_nifti_folder = join(self.output_folder_base, "gt_niftis")
maybe_mkdir_p(gt_nifti_folder)
for f in subfiles(self.gt_niftis_folder, suffix=".nii.gz"):
success = False
attempts = 0
e = None
while not success and attempts < 10:
try:
shutil.copy(f, gt_nifti_folder)
success = True
except OSError as e:
attempts += 1
sleep(1)
if not success:
print("Could not copy gt nifti file %s into folder %s" %
(f, gt_nifti_folder))
if e is not None:
raise e
self.network.train(current_mode)
###########################################
self.network.do_ds = ds
# return ret
def predict_preprocessed_data_return_seg_and_softmax(
self,
data: np.ndarray,
do_mirroring: bool = True,
mirror_axes: Tuple[int] = None,
use_sliding_window: bool = True,
step_size: float = 0.5,
use_gaussian: bool = True,
pad_border_mode: str = 'constant',
pad_kwargs: dict = None,
all_in_gpu: bool = True,
verbose: bool = True) -> Tuple[np.ndarray, np.ndarray]:
"""
We need to wrap this because we need to enforce self.network.do_ds = False for prediction
"""
ds = self.network.do_ds
self.network.do_ds = False
ret = super().predict_preprocessed_data_return_seg_and_softmax(
data, do_mirroring, mirror_axes, use_sliding_window, step_size,
use_gaussian, pad_border_mode, pad_kwargs, all_in_gpu, verbose)
self.network.do_ds = ds
return ret
def run_iteration(self,
data_generator,
do_backprop=True,
run_online_evaluation=False):
"""
gradient clipping improves training stability
:param data_generator:
:param do_backprop:
:param run_online_evaluation:
:return:
"""
data_dict = next(data_generator)
# length = get_length(data_generator)
data = data_dict['data']
target = data_dict['target']
# self.x_tags = ['liver','spleen','pancreas','rightkidney','leftkidney'] #test-mk
if self.x_tags is None:
self.x_tags = [tag.lower() for tag in data_dict['tags']]
y_tags = [tag.lower() for tag in data_dict['tags']]
# print("------------------x_tags:",self.x_tags)
# print("------------------y_tags:",y_tags)
data = maybe_to_torch(data)
target = maybe_to_torch(target)
if torch.cuda.is_available():
data = to_cuda(data)
target = to_cuda(target)
self.optimizer.zero_grad()
output = self.network(data)
del data
# loss = self.loss(output, target,self.x_tags,y_tags,need_updateGT=need_updateGT)
loss = self.loss(output, target, self.x_tags, y_tags)
if run_online_evaluation:
self.run_online_evaluation(output, target)
del target
if do_backprop:
if not self.fp16 or amp is None or not torch.cuda.is_available():
loss.backward()
else:
with amp.scale_loss(loss, self.optimizer) as scaled_loss:
scaled_loss.backward()
_ = clip_grad_norm_(self.network.parameters(), 12)
self.optimizer.step()
return loss.detach().cpu().numpy()
def do_split(self, task=None):
"""
we now allow more than 5 splits. IMPORTANT: and fold > 4 will not be a real split but just another random
80:20 split of the data. You cannot run X-fold cross-validation with this code. It will always be a 5-fold CV.
Folds > 4 will be independent from each other
:return:
"""
if task is None:
task = self.tasks[0]
if self.fold == 'all' or self.fold < 5: #fold = 4
self.dataset_directory = self.dataset_directory_dict[task]
return super().do_split()
else:
print("---------------!!!!!!!!--------------")
rnd = np.random.RandomState(seed=12345 + self.fold)
keys = np.sort(list(self.dataset.keys()))
idx_tr = rnd.choice(len(keys), int(len(keys) * 0.8), replace=False)
idx_val = [i for i in range(len(keys)) if i not in idx_tr]
self.dataset_tr = OrderedDict()
for i in idx_tr:
self.dataset_tr[keys[i]] = self.dataset[keys[i]]
self.dataset_val = OrderedDict()
for i in idx_val:
self.dataset_val[keys[i]] = self.dataset[keys[i]]
def setup_DA_params(self):
"""
- we increase roation angle from [-15, 15] to [-30, 30]
- scale range is now (0.7, 1.4), was (0.85, 1.25)
- we don't do elastic deformation anymore
:return:
"""
self.deep_supervision_scales = [[1, 1, 1]] + list(
list(i) for i in 1 / np.cumprod(
np.vstack(self.net_num_pool_op_kernel_sizes), axis=0))[:-1]
if self.threeD:
self.data_aug_params = default_3D_augmentation_params
self.data_aug_params['rotation_x'] = (-30. / 360 * 2. * np.pi,
30. / 360 * 2. * np.pi)
self.data_aug_params['rotation_y'] = (-30. / 360 * 2. * np.pi,
30. / 360 * 2. * np.pi)
self.data_aug_params['rotation_z'] = (-30. / 360 * 2. * np.pi,
30. / 360 * 2. * np.pi)
if self.do_dummy_2D_aug:
self.data_aug_params["dummy_2D"] = True
self.print_to_log_file("Using dummy2d data augmentation")
self.data_aug_params["elastic_deform_alpha"] = \
default_2D_augmentation_params["elastic_deform_alpha"]
self.data_aug_params["elastic_deform_sigma"] = \
default_2D_augmentation_params["elastic_deform_sigma"]
self.data_aug_params[
"rotation_x"] = default_2D_augmentation_params[
"rotation_x"]
else:
self.do_dummy_2D_aug = False
if max(self.patch_size) / min(self.patch_size) > 1.5:
default_2D_augmentation_params['rotation_x'] = (-15. / 360 *
2. * np.pi,
15. / 360 *
2. * np.pi)
self.data_aug_params = default_2D_augmentation_params
self.data_aug_params[
"mask_was_used_for_normalization"] = self.use_mask_for_norm
if self.do_dummy_2D_aug:
self.basic_generator_patch_size = get_patch_size(
self.patch_size[1:], self.data_aug_params['rotation_x'],
self.data_aug_params['rotation_y'],
self.data_aug_params['rotation_z'],
self.data_aug_params['scale_range'])
self.basic_generator_patch_size = np.array(
[self.patch_size[0]] + list(self.basic_generator_patch_size))
patch_size_for_spatialtransform = self.patch_size[1:]
else:
self.basic_generator_patch_size = get_patch_size(
self.patch_size, self.data_aug_params['rotation_x'],
self.data_aug_params['rotation_y'],
self.data_aug_params['rotation_z'],
self.data_aug_params['scale_range'])
patch_size_for_spatialtransform = self.patch_size
self.data_aug_params["scale_range"] = (0.7, 1.4)
self.data_aug_params["do_elastic"] = False
self.data_aug_params['selected_seg_channels'] = [0]
self.data_aug_params[
'patch_size_for_spatialtransform'] = patch_size_for_spatialtransform
self.data_aug_params["num_cached_per_thread"] = 2
def maybe_update_lr(self, epoch=None):
"""
if epoch is not None we overwrite epoch. Else we use epoch = self.epoch + 1
(maybe_update_lr is called in on_epoch_end which is called before epoch is incremented.
herefore we need to do +1 here)
:param epoch:
:return:
"""
if epoch is None:
ep = self.epoch + 1
else:
ep = epoch
self.optimizer.param_groups[0]['lr'] = poly_lr(ep, self.max_num_epochs,
self.initial_lr, 0.9)
self.print_to_log_file(
"lr:", np.round(self.optimizer.param_groups[0]['lr'], decimals=6))
def on_epoch_end(self):
"""
overwrite patient-based early stopping. Always run to 1000 epochs
:return:
"""
super().on_epoch_end()
continue_training = self.epoch < self.max_num_epochs
# it can rarely happen that the momentum of nnUNetTrainerV2 is too high for some dataset. If at epoch 100 the
# estimated validation Dice is still 0 then we reduce the momentum from 0.99 to 0.95
if self.epoch == 100:
if self.all_val_eval_metrics[-1] == 0:
self.optimizer.param_groups[0]["momentum"] = 0.95
self.network.apply(InitWeights_He(1e-2))
self.print_to_log_file(
"At epoch 100, the mean foreground Dice was 0. This can be caused by a too "
"high momentum. High momentum (0.99) is good for datasets where it works, but "
"sometimes causes issues such as this one. Momentum has now been reduced to "
"0.95 and network weights have been reinitialized")
return continue_training
def run_training(self):
"""
if we run with -c then we need to set the correct lr for the first epoch, otherwise it will run the first
continued epoch with self.initial_lr
we also need to make sure deep supervision in the network is enabled for training, thus the wrapper
:return:
"""
self.maybe_update_lr(
self.epoch
) # if we dont overwrite epoch then self.epoch+1 is used which is not what we
# want at the start of the training
ds = self.network.do_ds
self.network.do_ds = True
# ret = super().run_training()
########################################################################################
dct = OrderedDict()
for k in self.__dir__():
if not k.startswith("__"):
if not callable(getattr(self, k)):
dct[k] = str(getattr(self, k))
del dct['plans']
del dct['intensity_properties']
del dct['dataset']
del dct['dataset_tr']
del dct['dataset_val']
save_json(dct, join(self.output_folder, "debug.json"))
import shutil
shutil.copy(self.plans_file[self.tasks[0]],
join(self.output_folder_base, "plans.pkl"))
for i in range(len(self.tasks)):
self.tr_gen.setPart(i)
_ = self.tr_gen.next()
_ = self.val_gen.next()
self.tr_gen.setPart(0)
if torch.cuda.is_available():
torch.cuda.empty_cache()
# self._maybe_init_amp()
self.plot_network_architecture()
if cudnn.benchmark and cudnn.deterministic:
warn(
"torch.backends.cudnn.deterministic is True indicating a deterministic training is desired. "
"But torch.backends.cudnn.benchmark is True as well and this will prevent deterministic training! "
"If you want deterministic then set benchmark=False")
maybe_mkdir_p(self.output_folder)
if not self.was_initialized:
self.initialize(True)
flag = True
while self.epoch < self.max_num_epochs:
# self.need_updateGT=True
self.print_to_log_file("\nepoch: ", self.epoch)
epoch_start_time = time()
train_losses_epoch = []
# train one epoch
self.network.train()
if self.epoch >= self.stage_2_start_epoch and flag:
self.num_batches_per_epoch = 50
flag = False
if self.use_progress_bar:
with trange(self.num_batches_per_epoch) as tbar:
for b in tbar:
tbar.set_description("Epoch {}/{}".format(
self.epoch + 1, self.max_num_epochs))
l = self.run_iteration(self.tr_gen, True)
tbar.set_postfix(loss=l)
train_losses_epoch.append(l)
if self.epoch >= self.stage_2_start_epoch:
for i in range(1, len(self.tasks)):
self.tr_gen.setPart(i)
_ = self.run_iteration(self.tr_gen, True)
self.tr_gen.setPart(0)
else:
for _ in range(self.num_batches_per_epoch):
l = self.run_iteration(self.tr_gen, True)
train_losses_epoch.append(l)
if self.epoch > self.stage_2_start_epoch:
for i in range(1, len(self.tasks)):
self.tr_gen.setPart(i)
_ = self.run_iteration(self.tr_gen, True)
self.tr_gen.setPart(0)
self.all_tr_losses.append(np.mean(train_losses_epoch))
self.print_to_log_file("train loss : %.4f" %
self.all_tr_losses[-1])
with torch.no_grad():
# validation with train=False
self.network.eval()
val_losses = []
for b in range(self.num_val_batches_per_epoch):
l = self.run_iteration(self.val_gen, False, True)
val_losses.append(l)
self.all_val_losses.append(np.mean(val_losses))
self.print_to_log_file("validation loss: %.4f" %
self.all_val_losses[-1])
if self.also_val_in_tr_mode:
self.network.train()
# validation with train=True
val_losses = []
for b in range(self.num_val_batches_per_epoch):
l = self.run_iteration(self.val_gen, False)
val_losses.append(l)
self.all_val_losses_tr_mode.append(np.mean(val_losses))
self.print_to_log_file(
"validation loss (train=True): %.4f" %
self.all_val_losses_tr_mode[-1])
self.update_train_loss_MA(
) # needed for lr scheduler and stopping of training
continue_training = self.on_epoch_end()
epoch_end_time = time()
if not continue_training:
# allows for early stopping
break
self.epoch += 1
self.print_to_log_file("This epoch took %f s\n" %
(epoch_end_time - epoch_start_time))
self.epoch -= 1 # if we don't do this we can get a problem with loading model_final_checkpoint.
self.save_checkpoint(
join(self.output_folder, "model_final_checkpoint.model"))
# now we can delete latest as it will be identical with final
if isfile(join(self.output_folder, "model_latest.model")):
os.remove(join(self.output_folder, "model_latest.model"))
if isfile(join(self.output_folder, "model_latest.model.pkl")):
os.remove(join(self.output_folder, "model_latest.model.pkl"))
#########################################################################################
self.network.do_ds = ds
