def _construct_samples_dct(self):
logger.info('Will collect samples (img/ann pairs).')
self.name_to_tag = self.config['dataset_tags']
project_fs = ProjectFS.from_disk_dir_project(self.helper.paths.project_dir)
logger.info('Project structure has been read. Samples: {}.'.format(project_fs.pr_structure.image_cnt))
self.samples_dct = nn_data.samples_by_tags(tags=list(self.name_to_tag.values()), project_fs=project_fs,
project_meta=self.helper.in_project_meta)
def _load_train_config(self):
train_config_rw = JsonConfigRW(self.paths.model_config_fpath)
if not train_config_rw.config_exists:
raise RuntimeError(
'Unable to run inference, config from training wasn\'t found.')
self.train_config = train_config_rw.load()
self.class_title_to_idx = self.train_config[
self.class_title_to_idx_key]
self.train_classes = FigClasses(
self.train_config[self.train_classes_key])
logger.info('Read model internal class mapping',
extra={'class_mapping': self.class_title_to_idx})
logger.info('Read model out classes',
extra={'classes': self.train_classes.py_container})
# Make a separate [class title] --> [index] map that excludes the 'special' classes that should not be in the`
# final output.
train_class_titles = set(train_class['title']
for train_class in self.train_classes)
self.out_class_mapping = {
title: idx
for title, idx in self.class_title_to_idx.items()
if title in train_class_titles
}
def load_settings(self):
self.settings = deepcopy(ModelDeploy.settings)
new_settings = json_utils.json_load(
task_paths.TaskPaths(determine_in_project=False).settings_path)
logger.info('Input settings', extra={'settings': new_settings})
config_readers.update_recursively(self.settings, new_settings)
logger.info('Full settings', extra={'settings': self.settings})
def _determine_settings(self):
input_config = self.helper.task_settings
logger.info('Input config', extra={'config': input_config})
config = deepcopy(self.default_settings)
config_readers.update_recursively(config, input_config)
logger.info('Full config', extra={'config': config})
SettingsValidator.validate_train_cfg(config)
self.config = config
def __init__(self, default_settings={}):
logger.info('Will init all required to evaluation.')
self.helper = task_helpers.TaskHelperMetrics()
self.default_settings = deepcopy(default_settings)
self._determine_settings()
self._determine_input_data()
self._check_project_meta()
def __init__(self, settings=None):
logger.info('Starting base single image inference applier init.')
settings = settings or {}
self.settings = config_readers.update_recursively(
deepcopy(self.get_default_settings()), settings)
self.paths = TaskPaths(determine_in_project=False)
self._load_train_config()
self._construct_and_fill_model()
logger.info('Base single image inference applier init done.')
def _determine_model_input_size(self):
src_size = self.train_config['settings']['input_size']
self.input_size_wh = (src_size['width'], src_size['height'])
logger.info('Model input size is read (for auto-rescale).',
extra={
'input_size': {
'width': self.input_size_wh[0],
'height': self.input_size_wh[1]
}
})
def main():
task_helpers.task_verification(check_in_graph)
logger.info('DTL started')
helper = DtlHelper()
try:
net = Net(helper.graph, helper.in_project_metas,
helper.paths.results_dir)
helper.save_res_meta(net.get_result_project_meta())
datasets_conflict_map = calculate_datasets_conflict_map(helper)
except Exception as e:
logger.error("Error occurred on DTL-graph initialization step!")
raise e
# is_archive = net.is_archive()
results_counter = 0
for pr_name, pr_dir in helper.in_project_dirs.items():
project = sly.Project(directory=pr_dir, mode=sly.OpenMode.READ)
progress = progress_counter.progress_counter_dtl(
pr_name, project.total_items)
for dataset in project:
for item_name in dataset:
try:
img_desc = ImageDescriptor(
make_legacy_project_item(project, dataset, item_name),
datasets_conflict_map[project.name][dataset.name])
ann = json_utils.json_load(dataset.get_ann_path(item_name))
data_el = (img_desc, ann)
export_output_generator = net.start(data_el)
for res_export in export_output_generator:
logger.trace("image processed",
extra={
'img_name':
res_export[0][0].get_img_name()
})
results_counter += 1
except Exception as e:
extra = {
'project_name': project.name,
'ds_name': dataset.name,
'image_name': item_name,
'exc_str': str(e),
}
logger.warn(
'Image was skipped because some error occurred',
exc_info=True,
extra=extra)
progress.iter_done_report()
logger.info('DTL finished',
extra={
'event_type': EventType.DTL_APPLIED,
'new_proj_size': results_counter
})
def _determine_model_classes_detection(self):
in_project_classes = self.helper.in_project_meta.classes
self.out_classes = nn_data.make_out_classes(in_project_classes, shape='rectangle')
logger.info('Determined model out classes', extra={'classes': self.out_classes.py_container})
in_project_class_to_idx = nn_data.make_new_class_to_idx_map(in_project_classes,
start_class_id=self.get_start_class_id())
self.class_title_to_idx = nn_data.infer_training_class_to_idx_map(
self.config['weights_init_type'],
in_project_class_to_idx,
self.helper.paths.model_config_fpath,
class_to_idx_config_key=self.class_title_to_idx_key)
logger.info('Determined class mapping.', extra={'class_mapping': self.class_title_to_idx})
def _determine_settings(self):
input_config = self.helper.task_settings
logger.info('Input config', extra={'config': input_config})
config = deepcopy(self.default_settings)
config_readers.update_recursively(config, input_config)
logger.info('Full config', extra={'config': config})
if len(config['classes_mapping']) < 1:
raise RuntimeError('At least one classes pair should be defined')
self.classes_mapping = config['classes_mapping']
self.config = config
def __init__(self, default_settings):
logger.info('Will init all required to train.')
self.helper = task_helpers.TaskHelperTrain()
self.default_settings = deepcopy(default_settings)
self._determine_settings()
self._determine_model_classes()
self._determine_out_config()
self._construct_samples_dct()
self._construct_data_loaders()
self._construct_and_fill_model()
self._construct_loss()
self.epoch_flt = 0
def main():
task_helpers.task_verification(check_in_graph)
logger.info('DTL started')
helper = DtlHelper()
net = Net(helper.graph, helper.in_project_metas, helper.paths.results_dir)
helper.save_res_meta(net.get_result_project_meta())
datasets_conflict_map = calculate_datasets_conflict_map(helper)
# is_archive = net.is_archive()
results_counter = 0
for pr_name, pr_dir in helper.in_project_dirs.items():
root_path, project_name = ProjectFS.split_dir_project(pr_dir)
project_fs = ProjectFS.from_disk(root_path,
project_name,
by_annotations=True)
progress = progress_counter.progress_counter_dtl(
pr_name, project_fs.image_cnt)
for sample in project_fs:
try:
img_desc = ImageDescriptor(
sample, datasets_conflict_map[pr_name][sample.ds_name])
ann = json_utils.json_load(sample.ann_path)
data_el = (img_desc, ann)
export_output_generator = net.start(data_el)
for res_export in export_output_generator:
logger.trace(
"image processed",
extra={'img_name': res_export[0][0].get_img_name()})
results_counter += 1
except Exception as e:
extra = {
'project_name': sample.project_name,
'ds_name': sample.ds_name,
'image_name': sample.image_name,
'exc_str': str(e),
}
logger.warn('Image was skipped because some error occurred',
exc_info=True,
extra=extra)
progress.iter_done_report()
logger.info('DTL finished',
extra={
'event_type': EventType.DTL_APPLIED,
'new_proj_size': results_counter
})
def __init__(self,
single_image_inference: SingleImageInferenceApplier,
default_mode_settings,
default_settings_for_modes=None,
settings_validator_cls=AlwaysPassingValidator):
self.single_image_inference = single_image_inference
self.default_settings = config_readers.update_recursively(
deepcopy(default_mode_settings),
single_image_inference.get_default_settings())
self.default_settings_for_modes = deepcopy(
default_settings_for_modes or get_default_settings_for_modes())
self.settings_validator_cls = settings_validator_cls
self.helper = task_helpers.TaskHelperInference()
self._determine_settings()
self._determine_input_data()
logger.info('Dataset inference preparation done.')
def _determine_settings(self):
input_config = self.helper.task_settings
logger.info('Input config', extra={'config': input_config})
self.config = deepcopy(self.default_settings)
if 'mode' in input_config and 'source' in input_config['mode']:
mode_name = input_config['mode']['source']
default_mode_settings = self.default_settings_for_modes.get(
mode_name, None)
if default_mode_settings is not None:
config_readers.update_recursively(
self.config['mode'], deepcopy(default_mode_settings))
config_readers.update_recursively(self.config, input_config)
logger.info('Full config', extra={'config': self.config})
self.settings_validator_cls.validate_inference_cfg(self.config)
self.debug_copy_images = os.getenv('DEBUG_COPY_IMAGES') is not None
def _determine_input_data(self):
project_1 = self.config['project_1']
project_2 = self.config.get('project_2')
if project_2 is None:
project_2 = project_1
data_dir = os.path.join(self.helper.paths.task_dir, 'data')
for pr in [project_1, project_2]:
if not os.path.exists(os.path.join(data_dir, pr)):
raise RuntimeError('Project {} does not exist.'.format(pr))
self.in_project_fs_1 = ProjectFS.from_disk(data_dir, project_1)
self.in_project_fs_2 = ProjectFS.from_disk(data_dir, project_2)
if self.in_project_fs_1.image_cnt != self.in_project_fs_2.image_cnt:
raise RuntimeError(
'Projects should contain same number of samples.')
logger.info('Projects structure has been read. Samples: {}.'.format(
self.in_project_fs_1.image_cnt))
self.in_pr_meta_1 = ProjectMeta.from_dir(
os.path.join(data_dir, project_1))
self.in_pr_meta_2 = ProjectMeta.from_dir(
os.path.join(data_dir, project_2))
def _determine_input_data(self):
project_fs = ProjectFS.from_disk_dir_project(
self.helper.paths.project_dir)
logger.info('Project structure has been read. Samples: {}.'.format(
project_fs.pr_structure.image_cnt))
self.in_project_fs = project_fs