def run_experiment(self, raise_error=False, expected_msg=None):
dummy_exp = DummyExperiment()
dummy_exp = self.messenger_cls(dummy_exp, **self.messenger_kwargs)
dset_train = DummyDataset(10)
dset_test = DummyDataset(10)
dmgr_train = DataManager(dset_train, 2, 1, None)
dmgr_test = DataManager(dset_test, 2, 1, None)
with self.assertLogs(logger, level='INFO') as cm:
if raise_error:
with self.assertRaises(RuntimeError):
dummy_exp.run(dmgr_train,
dmgr_test,
raise_error=True,
**self.run_kwargs)
else:
dummy_exp.run(
dmgr_train,
dmgr_test,
raise_error=False,
**self.run_kwargs,
)
if expected_msg is None or not expected_msg:
logger.info("NoExpectedMessage")
if expected_msg is None or not expected_msg:
self.assertEqual(cm.output,
["INFO:UnitTestMessenger:NoExpectedMessage"])
else:
self.assertEqual(cm.output, expected_msg)
def test_datamanager(self):
batch_size = 16
np.random.seed(1)
dset = DummyDataset(600, [0.5, 0.3, 0.2])
manager = DataManager(dset,
batch_size,
n_process_augmentation=0,
transforms=None)
self.assertIsInstance(manager.get_batchgen(), Augmenter)
# create batch manually
data, labels = [], []
for i in range(batch_size):
data.append(dset[i]["data"])
labels.append(dset[i]["label"])
batch_dict = {"data": np.asarray(data), "label": np.asarray(labels)}
augmenter = manager.get_batchgen()
augmenter_iter = iter(augmenter)
for key, val in next(augmenter_iter).items():
self.assertTrue((val == batch_dict[key]).all())
for key, val in next(augmenter_iter).items():
self.assertEqual(len(val), batch_size)
def run_experiment(experiment_cls, config, network_cls, len_train, len_test,
**kwargs):
assert issubclass(experiment_cls, BaseExperiment)
exp = experiment_cls(config, network_cls, **kwargs)
dset_train = DummyDataset(len_train)
dset_test = DummyDataset(len_test)
dmgr_train = DataManager(dset_train, 16, 4, None)
dmgr_test = DataManager(dset_test, 16, 1, None)
return exp.run(dmgr_train, dmgr_test)
def kfold_experiment(self, raise_error=False, expected_msg=None):
kfold_kwargs = copy.deepcopy(self.run_kwargs)
kfold_kwargs.pop("fold")
dummy_exp = DummyExperiment()
dummy_exp = self.messenger_cls(dummy_exp, **self.messenger_kwargs)
dset = DummyDataset(10)
dmgr = DataManager(dset, 2, 1, None)
with self.assertLogs(logger, level='INFO') as cm:
if raise_error:
with self.assertRaises(RuntimeError):
dummy_exp.kfold(data=dmgr,
metrics={},
num_splits=2,
raise_error=True,
**kfold_kwargs)
else:
dummy_exp.kfold(data=dmgr,
metrics={},
num_splits=2,
raise_error=False,
**kfold_kwargs)
if expected_msg is None:
logger.info("NoExpectedMessage")
if expected_msg is None:
self.assertEqual(cm.output,
["INFO:UnitTestMessenger:NoExpectedMessage"])
else:
self.assertEqual(cm.output, expected_msg)
def kfold_experiment(experiment_cls,
config,
network_cls,
len_data,
shuffle=True,
split_type="random",
num_splits=2,
val_split=None,
**kwargs):
assert issubclass(experiment_cls, BaseExperiment)
metric_keys = kwargs.pop("metric_keys", None)
exp = experiment_cls(config, network_cls, **kwargs)
dset = DummyDataset(len_data)
dmgr = DataManager(dset, 16, 1, None)
return exp.kfold(data=dmgr,
metrics=config.nested_get("metrics"),
shuffle=shuffle,
split_type=split_type,
num_splits=num_splits,
val_split=val_split,
metric_keys=metric_keys)
def test_experiment(experiment_cls, config, network_cls, len_test, **kwargs):
assert issubclass(experiment_cls, BaseExperiment)
exp = experiment_cls(config, network_cls, **kwargs)
dset_test = DummyDataset(len_test)
dmgr_test = DataManager(dset_test, 16, 1, None)
model = network_cls()
return exp.test(model, dmgr_test, config.nested_get("metrics", {}),
kwargs.get("metric_keys", None))
torch.jit.loat(checkpoint_path)
model.eval()
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
model.to(device)
print("Load Data")
dset = WholeLegDataset(root_path=data_path,
include_flipped=True,
img_size=img_size,
contourwidth=25) # , bone_label=1)
dmgr = DataManager(dset, 1, 4, transforms, sampler_cls=SequentialSampler)
print("Start Predictions")
predictor = Predictor(
model,
key_mapping={"x": "data"},
convert_batch_to_npy_fn=convert_torch_to_numpy,
prepare_batch_fn=partial(model.prepare_batch,
input_device=device,
output_device=device),
)
os.makedirs(os.path.join(save_path, "per_class_all_masks"), exist_ok=True)
# os.makedirs(os.path.join(save_path, "per_mask_all_classes"), exist_ok=True)
include_flipped=True,
img_size=config.img_size,
bone_label=bone_label)
# idx_train, idx_val = train_test_split(
# list(range(len(dset))), test_size=config.val_split,
# random_state=config.seed)
# dset_train = dset.get_subset(idx_train)
# dset_val = dset.get_subset(idx_val)
dset_train = dset
dset_val = dset_test
mgr_train = DataManager(dset_train,
config.batchsize,
4,
train_transforms,
sampler_cls=RandomSampler)
mgr_val = DataManager(dset_val,
config.batchsize,
4,
test_transforms,
sampler_cls=SequentialSampler)
mgr_test = DataManager(dset_test,
config.batchsize,
4,
test_transforms,
sampler_cls=SequentialSampler)
experiment = PyTorchExperiment(
config,
def predict_data_mgr(self,
datamgr: DataManager,
batchsize=None,
metrics=None,
metric_keys=None,
verbose=False,
**kwargs):
"""
Defines a routine to predict data obtained from a batchgenerator
without explicitly caching anything
Parameters
----------
datamgr : :class:`DataManager`
Manager producing a generator holding the batches
batchsize : int
Artificial batchsize (sampling will be done with batchsize
1 and sampled data will be stacked to match the artificial
batchsize)(default: None)
metrics : dict
the metrics to calculate
metric_keys : dict
the ``batch_dict`` items to use for metric calculation
verbose : bool
whether to show a progress-bar or not, default: False
kwargs :
keyword arguments passed to :func:`prepare_batch_fn`
Yields
------
dict
a dictionary containing all predictions of the current batch
dict
a dictionary containing all metrics of the current batch
"""
if metrics is None:
metrics = {}
orig_num_aug_processes = datamgr.n_process_augmentation
orig_batch_size = datamgr.batch_size
if batchsize is None:
batchsize = orig_batch_size
datamgr.batch_size = 1
batchgen = datamgr.get_batchgen()
n_batches = datamgr.n_batches
if verbose:
iterable = tqdm(enumerate(batchgen),
unit=' sample',
total=n_batches,
desc=self._tqdm_desc)
else:
iterable = enumerate(batchgen)
batch_list = []
for i, batch in iterable:
Predictor._at_iter_begin(self, iter_num=i)
if not batch_list and (n_batches - i) < batchsize:
batchsize = n_batches - i
logger.debug("Set Batchsize down to %d to avoid cutting "
"of the last batches" % batchsize)
batch_list.append(batch)
# if queue is full process queue:
if batchsize is None or len(batch_list) >= batchsize:
batch_dict = {}
for _batch in batch_list:
for key, val in _batch.items():
if key in batch_dict.keys():
batch_dict[key].append(val)
else:
batch_dict[key] = [val]
for key, val_list in batch_dict.items():
batch_dict[key] = np.concatenate(val_list)
batch_dict = self._prepare_batch(batch_dict)
preds = self.predict(batch_dict,
already_prepared=True,
**kwargs)
# convert batchdict back to numpy (self.predict may convert it
# to backend-specific tensor type) - no-op if already numpy
batch_dict = self._convert_to_npy_fn(**batch_dict)[1]
preds_batch = LookupConfig()
# explicitly free memory of old lookup config
gc.collect()
preds_batch.update(batch_dict)
preds_batch.update(preds)
# calculate metrics for predicted batch
_metric_vals = self.calc_metrics(preds_batch,
metrics=metrics,
metric_keys=metric_keys)
self._at_iter_end(
data_dict={
**batch_dict,
**preds_batch
},
metrics={"val_" + k: v
for k, v in _metric_vals.items()},
iter_num=i)
yield preds, _metric_vals
batch_list = []
datamgr.batch_size = orig_batch_size
datamgr.n_process_augmentation = orig_num_aug_processes
return
def kfold(self,
data: DataManager,
metrics: dict,
num_epochs=None,
num_splits=None,
shuffle=False,
random_seed=None,
split_type="random",
val_split=0.2,
label_key="label",
train_kwargs: dict = None,
metric_keys: dict = None,
test_kwargs: dict = None,
config=None,
verbose=False,
**kwargs):
"""
Performs a k-Fold cross-validation
Parameters
----------
data : :class:`DataManager`
the data to use for training(, validation) and testing. Will be
split based on ``split_type`` and ``val_split``
metrics : dict
dictionary containing the metrics to evaluate during k-fold
num_epochs : int or None
number of epochs to train (if not given, will either be extracted
from ``config``, ``self.config`` or ``self.n_epochs``)
num_splits : int or None
the number of splits to extract from ``data``.
If None: uses a default of 10
shuffle : bool
whether to shuffle the data before splitting or not (implemented by
index-shuffling rather than actual data-shuffling to retain
potentially lazy-behavior of datasets)
random_seed : None
seed to seed numpy, the splitting functions and the used
backend-framework
split_type : str
must be one of ['random', 'stratified']
if 'random': uses random data splitting
if 'stratified': uses stratified data splitting. Stratification
will be based on ``label_key``
val_split : float or None
the fraction of the train data to use as validation set. If None:
No validation will be done during training; only testing for each
fold after the training is complete
label_key : str
the label to use for stratification. Will be ignored unless
``split_type`` is 'stratified'. Default: 'label'
train_kwargs : dict or None
kwargs to update the behavior of the :class:`DataManager`
containing the train data. If None: empty dict will be passed
metric_keys : dict of tuples
the batch_dict keys to use for each metric to calculate.
Should contain a value for each key in ``metrics``.
If no values are given for a key, per default ``pred`` and
``label`` will be used for metric calculation
test_kwargs : dict or None
kwargs to update the behavior of the :class:`DataManager`
containing the test and validation data.
If None: empty dict will be passed
config : :class:`DeliraConfig`or None
the training and model parameters
(will be merged with ``self.config``)
verbose : bool
verbosity
**kwargs :
additional keyword arguments
Returns
-------
dict
all predictions from all folds
dict
all metric values from all folds
Raises
------
ValueError
if ``split_type`` is neither 'random', nor 'stratified'
See Also
--------
* :class:`sklearn.model_selection.KFold`
and :class:`sklearn.model_selection.ShuffleSplit`
for random data-splitting
* :class:`sklearn.model_selection.StratifiedKFold`
and :class:`sklearn.model_selection.StratifiedShuffleSplit`
for stratified data-splitting
* :meth:`DataManager.update_from_state_dict` for updating the
data managers by kwargs
* :meth:`BaseExperiment.run` for the training
* :meth:`BaseExperiment.test` for the testing
Notes
-----
using stratified splits may be slow during split-calculation, since
each item must be loaded once to obtain the labels necessary for
stratification.
"""
# set number of splits if not specified
if num_splits is None:
num_splits = 10
logger.warning("num_splits not defined, using default value of \
10 splits instead ")
metrics_test, outputs = {}, {}
split_idxs = list(range(len(data.dataset)))
if train_kwargs is None:
train_kwargs = {}
if test_kwargs is None:
test_kwargs = {}
# switch between differnt kfold types
if split_type == "random":
split_cls = KFold
val_split_cls = ShuffleSplit
# split_labels are ignored for random splitting, set them to
# split_idxs just ensures same length
split_labels = split_idxs
elif split_type == "stratified":
split_cls = StratifiedKFold
val_split_cls = StratifiedShuffleSplit
# iterate over dataset to get labels for stratified splitting
split_labels = [
data.dataset[_idx][label_key] for _idx in split_idxs
]
else:
raise ValueError("split_type must be one of "
"['random', 'stratified'], but got: %s" %
str(split_type))
fold = split_cls(n_splits=num_splits,
shuffle=shuffle,
random_state=random_seed)
if random_seed is not None:
np.random.seed(random_seed)
# iterate over folds
for idx, (train_idxs,
test_idxs) in enumerate(fold.split(split_idxs,
split_labels)):
# extract data from single manager
train_data = data.get_subset(train_idxs)
test_data = data.get_subset(test_idxs)
train_data.update_state_from_dict(copy.deepcopy(train_kwargs))
test_data.update_state_from_dict(copy.deepcopy(test_kwargs))
val_data = None
if val_split is not None:
if split_type == "random":
# split_labels are ignored for random splitting, set them
# to split_idxs just ensures same length
train_labels = train_idxs
elif split_type == "stratified":
# iterate over dataset to get labels for stratified
# splitting
train_labels = [
train_data.dataset[_idx][label_key]
for _idx in train_idxs
]
else:
raise ValueError("split_type must be one of "
"['random', 'stratified'], but got: %s" %
str(split_type))
_val_split = val_split_cls(n_splits=1,
test_size=val_split,
random_state=random_seed)
for _train_idxs, _val_idxs in _val_split.split(
train_idxs, train_labels):
val_data = train_data.get_subset(_val_idxs)
val_data.update_state_from_dict(copy.deepcopy(test_kwargs))
train_data = train_data.get_subset(_train_idxs)
model = self.run(train_data=train_data,
val_data=val_data,
config=config,
num_epochs=num_epochs,
fold=idx,
**kwargs)
_outputs, _metrics_test = self.test(model,
test_data,
metrics=metrics,
metric_keys=metric_keys,
verbose=verbose)
outputs[str(idx)] = _outputs
metrics_test[str(idx)] = _metrics_test
return outputs, metrics_test