def _determine_model_configuration(self):
self.model_configuration = ObjectDetectionTrainer._determine_architecture_model_configuration(
sly.TaskPaths.MODEL_CONFIG_PATH)
# Check for possible model_configuration field in old-style config. If exists, make sure it is consistent with
# the actual model config and clear model_configuration from training config before writing new-style model
# config.
training_model_configuration = self.config.get('model_configuration',
None)
if training_model_configuration is not None:
if training_model_configuration != self.model_configuration:
error_msg = (
'Unable to start training. model_confguration in the training config is not consistent with '
+
'selected model architecture. Make sure you have selected the right model plugin and remove '
+
'model_confguration from the training config as it is not required anymore.'
)
logger.critical(error_msg,
extra={
'training_model_configuration':
self.config['model_configuration'],
'model_configuration':
self.model_configuration
})
raise RuntimeError(error_msg)
del self.config['model_configuration']
def infer_training_class_to_idx_map(weights_init_type, in_project_class_to_idx, model_config_fpath,
class_to_idx_config_key, special_class_ids=None):
if weights_init_type == TRANSFER_LEARNING:
logger.info('Transfer learning mode, using a class mapping created from scratch.')
class_title_to_idx = in_project_class_to_idx
elif weights_init_type == CONTINUE_TRAINING:
logger.info('Continued training mode, reusing the existing class mapping from the model.')
class_title_to_idx = read_validate_model_class_to_idx_map(
model_config_fpath=model_config_fpath,
in_project_classes_set=set(in_project_class_to_idx.keys()),
class_to_idx_config_key=class_to_idx_config_key)
else:
raise RuntimeError('Unknown weights init type: {}'.format(weights_init_type))
if special_class_ids is not None:
for class_title, requested_class_id in special_class_ids.items():
effective_class_id = class_title_to_idx[class_title]
if requested_class_id != effective_class_id:
error_msg = ('Unable to start training. Effective integer id for class {} does not match the ' +
'requested value in the training config ({} vs {}).'.format(
class_title, effective_class_id, requested_class_id))
logger.critical(error_msg, extra={'class_title_to_idx': class_title_to_idx,
'special_class_ids': special_class_ids})
raise RuntimeError(error_msg)
return class_title_to_idx
def create_task(task_msg, docker_api):
task_cls = _task_class_mapping.get(task_msg['task_type'], None)
if task_cls is None:
logger.critical('unknown task type', extra={'task_msg': task_msg})
raise RuntimeError('unknown task type')
task_obj = task_cls(task_msg)
if issubclass(task_cls, TaskDockerized):
task_obj.docker_api = docker_api
return task_obj
def read_validate_model_class_to_idx_map(model_config_fpath, in_project_classes_set, class_to_idx_config_key):
"""Reads class id --> int index mapping from the model config; checks that the set of classes matches the input."""
if not fs.file_exists(model_config_fpath):
raise RuntimeError('Unable to continue_training, config for previous training wasn\'t found.')
with open(model_config_fpath) as fin:
model_config = json.load(fin)
model_class_mapping = model_config.get(class_to_idx_config_key, None)
if model_class_mapping is None:
raise RuntimeError('Unable to continue_training, model does not have class mapping information.')
model_classes_set = set(model_class_mapping.keys())
if model_classes_set != in_project_classes_set:
error_message_text = 'Unable to continue_training, sets of classes for model and dataset do not match.'
logger.critical(
error_message_text, extra={'model_classes': model_classes_set, 'dataset_classes': in_project_classes_set})
raise RuntimeError(error_message_text)
return model_class_mapping.copy()