Refactoring: now config weights are explicitly initialized. Prodeeded…

… to #199.

arekit/contrib/networks/core/data_handling/data.py

@@ -24,6 +24,7 @@ def __init__(self, labeled_collections, bags_collection):
         assert(isinstance(bags_collection, dict))
         self.__labeled_collections = labeled_collections
         self.__bags_collection = bags_collection
+        self.__train_stat_uint_labeled_sample_row_ids = None
 
     # region Properties
 
@@ -35,6 +36,10 @@ def BagsCollections(self):
     def SamplesLabelingCollection(self):
         return self.__labeled_collections
 
+    @property
+    def HasNormalizedWeights(self):
+        return self.__train_stat_uint_labeled_sample_row_ids is not None
+
     # endregion
 
     @staticmethod
@@ -53,7 +58,9 @@ def create_empty(cls):
                    bags_collection={})
 
     def initialize(self, dtypes, create_samples_reader_func, has_model_predefined_state,
-                   vocab, labels_count, bags_collection_type, config):
+                   vocab, labels_count, bags_collection_type,
+                   # TODO: 199. Remove config.
+                   config):
         """
         Perform reading information from the serialized experiment inputs.
         Initializing core configuration.
@@ -63,8 +70,6 @@ def initialize(self, dtypes, create_samples_reader_func, has_model_predefined_st
         assert(isinstance(has_model_predefined_state, bool))
         assert(isinstance(labels_count, int) and labels_count > 0)
 
-        stat_uint_labeled_sample_row_ids = None
-
         # Reading from serialized information
         for data_type in dtypes:
 
@@ -77,6 +82,7 @@ def initialize(self, dtypes, create_samples_reader_func, has_model_predefined_st
                 is_external_vocab=has_model_predefined_state,
                 bags_collection_type=bags_collection_type,
                 vocab=vocab,
+                # TODO: 199. Remove config.
                 config=config,
                 desc="Filling bags collection [{}]".format(data_type))
 
@@ -86,14 +92,19 @@ def initialize(self, dtypes, create_samples_reader_func, has_model_predefined_st
                 uint_labeled_sample_row_ids=uint_labeled_sample_row_ids)
 
             if data_type == DataType.Train:
-                stat_uint_labeled_sample_row_ids = uint_labeled_sample_row_ids
+                self.__train_stat_uint_labeled_sample_row_ids = uint_labeled_sample_row_ids
+
+    def calc_normalized_weigts(self, labels_count):
+        assert(isinstance(labels_count, int) and labels_count > 0)
+
+        if self.__train_stat_uint_labeled_sample_row_ids is None:
+            return
+
+        normalized_label_stat, _ = calculate_labels_distribution_stat(
+            uint_labeled_sample_row_ids=self.__train_stat_uint_labeled_sample_row_ids,
+            classes_count=labels_count)
 
-        # Calculate class weights.
-        if stat_uint_labeled_sample_row_ids is not None:
-            normalized_label_stat, _ = calculate_labels_distribution_stat(
-                uint_labeled_sample_row_ids=stat_uint_labeled_sample_row_ids,
-                classes_count=labels_count)
-            config.set_class_weights(normalized_label_stat)
+        return normalized_label_stat
 
     # region writing methods
 
@@ -156,9 +167,12 @@ def __iter_parsed_news_func(doc_ops, data_type):
 
     @staticmethod
     def __read_for_data_type(samples_reader, is_external_vocab,
-                             bags_collection_type, vocab, config, desc=""):
+                             bags_collection_type, vocab,
+                             # TODO: 199. Remove config.
+                             config, desc=""):
         assert(isinstance(samples_reader, BaseInputSampleReader))
 
+        # TODO: 199. Use shapes.
         terms_per_context = config.TermsPerContext
         frames_per_context = config.FramesPerContext
         synonyms_per_context = config.SynonymsPerContext
@@ -169,6 +183,7 @@ def __read_for_data_type(samples_reader, is_external_vocab,
             bag_size=config.BagSize,
             shuffle=True,
             create_empty_sample_func=lambda: InputSample.create_empty(
+                # TODO: 199. Use shapes.
                 terms_per_context=terms_per_context,
                 frames_per_context=frames_per_context,
                 synonyms_per_context=synonyms_per_context),
@@ -184,6 +199,7 @@ def __read_for_data_type(samples_reader, is_external_vocab,
                 frame_sent_roles=row.TextFrameVariantRoles,
                 syn_obj_inds=row.SynonymObjectInds,
                 syn_subj_inds=row.SynonymSubjectInds,
+                # TODO: 199. Use shapes.
                 terms_per_context=terms_per_context,
                 frames_per_context=frames_per_context,
                 synonyms_per_context=synonyms_per_context,

arekit/contrib/networks/run_training.py

@@ -83,8 +83,13 @@ def _handle_iteration(self, it_index):
             labels_count=self._experiment.DataIO.LabelsCount,
             vocab=vocab,
             bags_collection_type=self.__bags_collection_type,
+            # TODO: 199. Remove config.
             config=self.__config)
 
+        if handled_data.HasNormalizedWeights:
+            weights = handled_data.calc_normalized_weigts(labels_count=self._experiment.DataIO.LabelsCount)
+            self.__config.set_class_weights(weights)
+
         # Update parameters after iteration preparation has been completed.
         self.__config.reinit_config_dependent_parameters()
 

arekit/contrib/networks/sample.py

@@ -121,6 +121,7 @@ def create_empty(cls, terms_per_context, frames_per_context, synonyms_per_contex
                    frame_indices=blank_frames,
                    input_sample_id="1")
 
+    # TODO. Refactoring #199.
     @classmethod
     def _generate_test(cls, config):
         assert(isinstance(config, DefaultNetworkConfig))
@@ -164,6 +165,7 @@ def create_from_parameters(cls,
                                subj_ind,
                                obj_ind,
                                words_vocab,  # for indexing input (all the vocabulary, obtained from offsets.py)
+                               # TODO: 199. Use shapes.
                                terms_per_context,  # for terms_per_context, frames_per_context.
                                frames_per_context,
                                synonyms_per_context,

examples/pipeline.py

@@ -137,7 +137,8 @@ def extract(text):
         vocab=None,
         labels_count=3,
         bags_collection_type=SingleBagsCollection,
-        config=None,                                  # TODO. Конфигурация сети.
+        # TODO: 199. Remove config.
+        config=None,
     )
 
     ############################