diff options
Diffstat (limited to 'model-optimizer/mo/utils/pipeline_config.py')
| -rw-r--r-- | model-optimizer/mo/utils/pipeline_config.py | 142 |
1 files changed, 97 insertions, 45 deletions
diff --git a/model-optimizer/mo/utils/pipeline_config.py b/model-optimizer/mo/utils/pipeline_config.py index c5ae47f89..901bf453c 100644 --- a/model-optimizer/mo/utils/pipeline_config.py +++ b/model-optimizer/mo/utils/pipeline_config.py @@ -13,11 +13,65 @@ See the License for the specific language governing permissions and limitations under the License. """ +import logging as log +import re from mo.utils.error import Error from mo.utils.simple_proto_parser import SimpleProtoParser +# The list of rules how to map the value from the pipeline.config file to the dictionary with attributes. +# The rule is either a string or a tuple with two elements. In the first case the rule string is used as a key to +# search in the parsed pipeline.config file attributes dictionary and a key to save found value. In the second case the +# first element of the tuple is the key to save found value; the second element of the tuple is a string defining the +# path to the value of the attribute in the pipeline.config file. The path consists of the regular expression strings +# defining the dictionary key to look for separated with a '/' character. +mapping_rules = [ + 'num_classes', + # preprocessing block attributes + ('resizer_image_height', 'image_resizer/fixed_shape_resizer/height'), + ('resizer_image_width', 'image_resizer/fixed_shape_resizer/width'), + ('resizer_min_dimension', 'image_resizer/keep_aspect_ratio_resizer/min_dimension'), + ('resizer_max_dimension', 'image_resizer/keep_aspect_ratio_resizer/max_dimension'), + # anchor generator attributes + ('anchor_generator_height', 'first_stage_anchor_generator/grid_anchor_generator/height$', 256), + ('anchor_generator_width', 'first_stage_anchor_generator/grid_anchor_generator/width$', 256), + ('anchor_generator_height_stride', 'first_stage_anchor_generator/grid_anchor_generator/height_stride', 16), + ('anchor_generator_width_stride', 'first_stage_anchor_generator/grid_anchor_generator/width_stride', 16), + ('anchor_generator_scales', 'first_stage_anchor_generator/grid_anchor_generator/scales'), + ('anchor_generator_aspect_ratios', 'first_stage_anchor_generator/grid_anchor_generator/aspect_ratios'), + ('multiscale_anchor_generator_min_level', 'anchor_generator/multiscale_anchor_generator/min_level'), + ('multiscale_anchor_generator_max_level', 'anchor_generator/multiscale_anchor_generator/max_level'), + ('multiscale_anchor_generator_anchor_scale', 'anchor_generator/multiscale_anchor_generator/anchor_scale'), + ('multiscale_anchor_generator_aspect_ratios', 'anchor_generator/multiscale_anchor_generator/aspect_ratios'), + ('multiscale_anchor_generator_scales_per_octave', 'anchor_generator/multiscale_anchor_generator/scales_per_octave'), + # SSD anchor generator attributes + ('ssd_anchor_generator_min_scale', 'anchor_generator/ssd_anchor_generator/min_scale'), + ('ssd_anchor_generator_max_scale', 'anchor_generator/ssd_anchor_generator/max_scale'), + ('ssd_anchor_generator_num_layers', 'anchor_generator/ssd_anchor_generator/num_layers'), + ('ssd_anchor_generator_aspect_ratios', 'anchor_generator/ssd_anchor_generator/aspect_ratios'), + ('ssd_anchor_generator_reduce_lowest', 'anchor_generator/ssd_anchor_generator/reduce_boxes_in_lowest_layer'), + ('ssd_anchor_generator_base_anchor_height', 'anchor_generator/ssd_anchor_generator/base_anchor_height', 1.0), + ('ssd_anchor_generator_base_anchor_width', 'anchor_generator/ssd_anchor_generator/base_anchor_width', 1.0), + # Proposal and ROI Pooling layers attributes + ('first_stage_nms_score_threshold', '.*_nms_score_threshold'), + ('first_stage_nms_iou_threshold', '.*_nms_iou_threshold'), + ('first_stage_max_proposals', '.*_max_proposals'), + 'initial_crop_size', + # Detection Output layer attributes + ('postprocessing_score_converter', '.*/score_converter'), + ('postprocessing_score_threshold', '.*/batch_non_max_suppression/score_threshold'), + ('postprocessing_iou_threshold', '.*/batch_non_max_suppression/iou_threshold'), + ('postprocessing_max_detections_per_class', '.*/batch_non_max_suppression/max_detections_per_class'), + ('postprocessing_max_total_detections', '.*/batch_non_max_suppression/max_total_detections'), + # Variances for predicted bounding box deltas (tx, ty, tw, th) + ('frcnn_variance_x', 'box_coder/faster_rcnn_box_coder/x_scale', 10.0), + ('frcnn_variance_y', 'box_coder/faster_rcnn_box_coder/y_scale', 10.0), + ('frcnn_variance_width', 'box_coder/faster_rcnn_box_coder/width_scale', 5.0), + ('frcnn_variance_height', 'box_coder/faster_rcnn_box_coder/height_scale', 5.0) +] + + class PipelineConfig: """ The class that parses pipeline.config files used to generate TF models generated using Object Detection API. @@ -33,57 +87,55 @@ class PipelineConfig: self._initialize_model_params() + @staticmethod + def _get_value_by_path(params: dict, path: list): + if not path or len(path) == 0: + return None + if not isinstance(params, dict): + return None + compiled_regexp = re.compile(path[0]) + for key in params.keys(): + if re.match(compiled_regexp, key): + if len(path) == 1: + return params[key] + else: + value = __class__._get_value_by_path(params[key], path[1:]) + if value is not None: + return value + return None + + def _update_param_using_rule(self, params: dict, rule: [str, tuple]): + if isinstance(rule, str): + if rule in params: + self._model_params[rule] = params[rule] + log.debug('Found value "{}" for path "{}"'.format(params[rule], rule)) + elif isinstance(rule, tuple): + if len(rule) != 2 and len(rule) != 3: + raise Error('Invalid rule length. Rule must be a tuple with two elements: key and path, or three ' + 'elements: key, path, default_value.') + value = __class__._get_value_by_path(params, rule[1].split('/')) + if value is not None: + log.debug('Found value "{}" for path "{}"'.format(value, rule[1])) + self._model_params[rule[0]] = value + elif len(rule) == 3: + self._model_params[rule[0]] = rule[2] + log.debug('There is no value path "{}". Set default value "{}"'.format(value, rule[2])) + + else: + raise Error('Invalid rule type. Rule can be either string or tuple') + def _initialize_model_params(self): """ Store global params in the dedicated dictionary self._model_params for easier use. :return: None """ - params = list(self._raw_data_dict['model'].values())[0] - # global topology parameters - self._model_params['num_classes'] = params['num_classes'] - - # pre-processing of the image - self._model_params['image_resizer'] = list(params['image_resizer'].keys())[0] - image_resize_params = list(params['image_resizer'].values())[0] - if self._model_params['image_resizer'] == 'keep_aspect_ratio_resizer': - self._model_params['preprocessed_image_height'] = image_resize_params['min_dimension'] - self._model_params['preprocessed_image_width'] = self._model_params['preprocessed_image_height'] - elif self._model_params['image_resizer'] == 'fixed_shape_resizer': - self._model_params['preprocessed_image_height'] = image_resize_params['height'] - self._model_params['preprocessed_image_width'] = image_resize_params['width'] - else: - raise Error('Unknown image resizer type "{}"'.format(self._model_params['image_resizer'])) - - # grid anchors generator - if 'first_stage_anchor_generator' in params: - grid_params = params['first_stage_anchor_generator']['grid_anchor_generator'] - self._model_params['anchor_generator_base_size'] = 256 - self._model_params['anchor_generator_stride'] = grid_params['height_stride'] - self._model_params['anchor_generator_scales'] = grid_params['scales'] - self._model_params['anchor_generator_aspect_ratios'] = grid_params['aspect_ratios'] - - if 'feature_extractor' in params: - if 'first_stage_features_stride' in params['feature_extractor']: - self._model_params['features_extractor_stride'] = params['feature_extractor']['first_stage_features_stride'] - else: # the value is not specified in the configuration file for NASNet so use default value here - self._model_params['features_extractor_stride'] = 16 - - # Proposal and ROI Pooling layers - for param in ['first_stage_nms_score_threshold', 'first_stage_nms_iou_threshold', 'first_stage_max_proposals', - 'initial_crop_size']: - if param in params: - self._model_params[param] = params[param] - - # post-processing parameters - postprocessing_params = None - for postpocessing_type in ['post_processing', 'second_stage_post_processing']: - if postpocessing_type in params: - postprocessing_params = params[postpocessing_type]['batch_non_max_suppression'] - self._model_params['postprocessing_score_converter'] = params[postpocessing_type]['score_converter'] - if postprocessing_params is not None: - for param in ['score_threshold', 'iou_threshold', 'max_detections_per_class', 'max_total_detections']: - self._model_params['postprocessing_' + param] = postprocessing_params[param] + if 'model' not in self._raw_data_dict: + raise Error('The "model" key is not found in the configuration file. Looks like the parsed file is not ' + 'Object Detection API model configuration file.') + params = list(self._raw_data_dict['model'].values())[0] + for rule in mapping_rules: + self._update_param_using_rule(params, rule) def get_param(self, param: str): if param not in self._model_params: |