def parse(path, extra=None) -> GenericPipeline:
extraImports = []
if isinstance(path, str):
if not os.path.exists(path) or os.path.isdir(path):
cur_project_pth = context.get_current_project_path()
config_path = os.path.join(path, "config.yaml")
if os.path.exists(config_path):
path = config_path
else:
config_path = os.path.join(cur_project_pth, "experiments",
path, "config.yaml")
if os.path.exists(config_path):
path = config_path
if os.path.exists(cur_project_pth +
"/common.yaml") and extra is None:
extra = cur_project_pth + "/common.yaml"
modules_path = os.path.join(cur_project_pth, 'modules')
if os.path.exists(modules_path):
for m in os.listdir(modules_path):
sys.path.insert(0, modules_path)
if ".py" in m:
extraImports.append(m[0:m.index(".py")])
cfg = configloader.parse("generic", path, extra)
cfg.path = path
for e in extraImports:
cfg.imports.append(e)
return cfg
def perform(self, server, reporter: ProgressMonitor):
ms=ModelSpec(**self.spec)
settings = yaml.load(self.augmentationConfig)
augmenters_lst = configloader.parse("augmenters", settings)
if len(augmenters_lst) > 0:
augmenter = iaa.Sequential(augmenters_lst)
exp:Experiment=server.experiment(self.experimentPath)
cf=exp.parse_config({"augmentation": settings})
ds = _cache.get_dataset(exp, self.datasetName)
wrappedModel = ms.wrap(cf, exp)
targets=_cache.get_targets(exp,self.datasetName)
def augmented_image_visializer(predictionItem:PredictionItem):
cache_path=visualization_ctx().path
path = cache_path + str(predictionItem.id) + ".png"
if os.path.exists(path):
return path
if len(predictionItem.y.shape) > 1: #Should be a segmentation mask in this case, better to refactor this later
batch = imgaug.Batch(images=[predictionItem.x], segmentation_maps=[imgaug.SegmentationMapsOnImage(predictionItem.y, shape=predictionItem.y.shape)])
aug_batch = augmenter.augment_batch(batch)
img = aug_batch.segmentation_maps_aug[0].draw_on_image(aug_batch.images_aug[0], alpha=0.5)[0]
else:
img = augmenter.augment_image(predictionItem.x) # Augment segmantation map also
imageio.imwrite(path,img)
return path
wrapped_dataset=WrappedDS(ds,list(range(len(ds))),"sample",None,None)
sig=inspect.signature(augmented_image_visializer)
visualizerFunc = projects.WrappedVisualizer("augmented", augmented_image_visializer, sig)
wrapped_dataset._visualizer=visualizerFunc.create(wrapped_dataset,tempfile.mkdtemp())
return AnalizeResults([wrapped_dataset],None)
def parse(path, extra=None) -> GenericPipeline:
extraImports = []
if isinstance(path, str):
if not os.path.exists(path) or os.path.isdir(path):
pth = context.get_current_project_path()
if os.path.exists(pth + "/experiments/" + path + "/config.yaml"):
path = pth + "/experiments/" + path + "/config.yaml"
if os.path.exists(pth + "/common.yaml") and extra is None:
extra = pth + "/common.yaml"
if os.path.exists(pth + "/modules"):
for m in os.listdir(pth + "/modules"):
sys.path.insert(0, pth + "/modules")
if ".py" in m:
extraImports.append(m[0:m.index(".py")])
cfg = configloader.parse("generic", path, extra)
cfg.path = path
for e in extraImports:
cfg.imports.append(e)
return cfg
def __init__(self, **args):
super(CallbackModule, self).__init__()
self.relSize = args['relSize'] if 'relSize' in args else None
self.absSize = args['absSize'] if 'absSize' in args else None
self.periodEpochs = args['periodEpochs'] if 'periodEpochs' in args else None
self.periodSteps = args['periodSteps'] if 'periodSteps' in args else None
if self.absSize is None and self.relSize is None:
raise ValueError("'absSize' or 'relSize' must be specified for CallbackModule")
if self.absSize is not None and self.relSize is not None:
raise ValueError("'absSize' and 'relSize' are mutually exclusive for CallbackModule")
if self.periodEpochs is not None and self.periodSteps is not None:
raise ValueError("'periodEpochs' and 'periodSteps' are mutually exclusive for CallbackModule")
self.then = None
if 'then' in args:
thenArg = args['then']
parsedThen = configloader.parse("callbacks", thenArg)
if len(parsedThen) > 0:
self.then = parsedThen[0]
def parse(path) -> PipelineConfig:
cfg = configloader.parse("segmentation", path)
cfg.path = path
return cfg
def execute(self,
kf: datasets.DefaultKFoldedDataSet,
model: keras.Model,
ec: ExecutionConfig,
callbacks=None):
if 'unfreeze_encoder' in self.dict and self.dict['unfreeze_encoder']:
self.unfreeze(model)
if 'unfreeze_encoder' in self.dict and not self.dict[
'unfreeze_encoder']:
self.freeze(model)
if callbacks is None:
cb = [] + self.cfg.callbacks
else:
cb = callbacks
if self.cfg._reporter is not None:
if self.cfg._reporter.isCanceled():
return
cb.append(ReporterCallback(self.cfg._reporter))
pass
prevInfo = None
if self.cfg.resume:
allBest = self.cfg.info()
filtered = list(
filter(lambda x: x.stage == ec.stage and x.fold == ec.fold,
allBest))
if len(filtered) > 0:
prevInfo = filtered[0]
self.lr = prevInfo.lr
if self.loss or self.lr:
self.cfg.compile(model, self.cfg.createOptimizer(self.lr),
self.loss)
if self.initial_weights is not None:
try:
model.load_weights(self.initial_weights)
except:
z = model.layers[-1].name
model.layers[-1].name = "tmpName12312"
model.load_weights(self.initial_weights, by_name=True)
model.layers[-1].name = z
if 'callbacks' in self.dict:
cb = configloader.parse("callbacks", self.dict['callbacks'])
if 'extra_callbacks' in self.dict:
cb = cb + configloader.parse("callbacks",
self.dict['extra_callbacks'])
kepoch = -1
if "logAll" in self.dict and self.dict["logAll"]:
cb = cb + [AllLogger(ec.metricsPath() + "all.csv")]
cb.append(KFoldCallback(kf))
kepoch = self._addLogger(model, ec, cb, kepoch)
md = self.cfg.primary_metric_mode
if self.cfg.gpus == 1:
mcp = keras.callbacks.ModelCheckpoint(
ec.weightsPath(),
save_best_only=True,
monitor=self.cfg.primary_metric,
mode=md,
verbose=1)
if prevInfo != None:
mcp.best = prevInfo.best
cb.append(mcp)
self.add_visualization_callbacks(cb, ec, kf)
if self.epochs - kepoch == 0:
return
self.loadBestWeightsFromPrevStageIfExists(ec, model)
self._doTrain(kf, model, ec, cb, kepoch)
print('saved')
pass
def parse(path) -> ClassificationPipeline:
cfg = configloader.parse("classification", path)
cfg.path = path
return cfg