AC: fix text classification truncation by max len (#3015)

* AC: fix text classification truncation by max len

* update data reading for annotation

Author: eaidova

tools/accuracy_checker/openvino/tools/accuracy_checker/annotation_converters/text_classification.py

@@ -20,6 +20,7 @@
 
 
 from ..config import PathField, StringField, NumberField, BoolField, ListField, ConfigError
+from ..data_readers import AnnotationDataIdentifier
 from ..representation import TextClassificationAnnotation
 from ..utils import string_to_list, UnsupportedPackage, read_json
 from .format_converter import BaseFormatConverter, ConverterReturn, verify_label_map
@@ -112,11 +113,7 @@ def read_annotation(self):
         return lines
 
     def convert_single_example(self, example): # pylint:disable=R0912
-        identifier = [
-            'input_ids_{}'.format(example.guid),
-            'input_mask_{}'.format(example.guid),
-            'segment_ids_{}'.format(example.guid)
-        ]
+        identifier = AnnotationDataIdentifier(example.guid, [])
         if not self.external_tok:
             tokens_a = self.tokenizer.tokenize(example.text_a)
             tokens_b = None
@@ -165,6 +162,7 @@ def convert_single_example(self, example): # pylint:disable=R0912
 
             if len(tokens) > self.max_seq_length:
                 tokens = tokens[:self.max_seq_length]
+                segment_ids = segment_ids[:self.max_seq_length]
 
         input_ids = self.tokenizer.convert_tokens_to_ids(tokens) if self.support_vocab or self.external_tok else tokens
         input_mask = [0 if not self.class_token_first else 1] * len(input_ids)

tools/accuracy_checker/openvino/tools/accuracy_checker/data_readers/__init__.py

@@ -19,6 +19,7 @@
     DataReaderField,
     ReaderCombiner,
     DataRepresentation,
+    AnnotationDataIdentifier,
     ClipIdentifier,
     MultiFramesInputIdentifier,
     ImagePairIdentifier,
@@ -32,6 +33,7 @@
     serialize_identifier,
     deserialize_identifier,
     create_identifier_key,
+    create_ann_identifier_key,
 
     create_reader,
     REQUIRES_ANNOTATIONS
@@ -67,6 +69,7 @@
     'KaldiFrameIdentifier',
     'ParametricImageIdentifier',
     'VideoFrameIdentifier',
+    'AnnotationDataIdentifier',
 
     'OpenCVFrameReader',
     'OpenCVImageReader',
@@ -94,5 +97,6 @@
 
     'serialize_identifier',
     'deserialize_identifier',
-    'create_identifier_key'
+    'create_identifier_key',
+    'create_ann_identifier_key'
 ]

tools/accuracy_checker/openvino/tools/accuracy_checker/data_readers/annotation_readers.py

@@ -15,7 +15,7 @@
 """
 
 from ..config import ListField, ConfigError
-from .data_reader import BaseReader, create_identifier_key
+from .data_reader import BaseReader, create_ann_identifier_key, AnnotationDataIdentifier
 from ..utils import contains_all
 
 
@@ -47,7 +47,10 @@ def configure(self):
         self.multi_infer = self.get_value_from_config('multi_infer')
 
     def read(self, data_id):
-        relevant_annotation = self.data_source[create_identifier_key(data_id)]
+        if isinstance(data_id, AnnotationDataIdentifier):
+            ordered_data_id = ['{}_{}'.format(feat, data_id.annotation_id) for feat in self.feature_list]
+            data_id.data_id = ordered_data_id if not self.single else ordered_data_id[0]
+        relevant_annotation = self.data_source[create_ann_identifier_key(data_id)]
         if not contains_all(relevant_annotation.__dict__, self.feature_list):
             raise ConfigError(
                 'annotation_class prototype does not contain provided features {}'.format(', '.join(self.feature_list))

tools/accuracy_checker/openvino/tools/accuracy_checker/data_readers/data_reader.py

@@ -49,6 +49,11 @@ def __init__(self, data, meta=None, identifier=''):
             self.metadata['image_size'] = data.shape if not isinstance(data, list) else np.shape(data[0])
 
 
+class AnnotationDataIdentifier:
+    def __init__(self, ann_id, data_id):
+        self.annotation_id = ann_id
+        self.data_id = data_id
+
 ClipIdentifier = namedtuple('ClipIdentifier', ['video', 'clip_id', 'frames'])
 MultiFramesInputIdentifier = namedtuple('MultiFramesInputIdentifier', ['input_id', 'frames'])
 ImagePairIdentifier = namedtuple('ImagePairIdentifier', ['first', 'second'])
@@ -64,6 +69,10 @@ def __init__(self, data, meta=None, identifier=''):
 )
 
 identifier_serialization = {
+    'AnnotationDataIdentifier': IdentifierSerializationOptions(
+        'annotation_data_identifier', ['annotation_id', 'data_id'],
+        AnnotationDataIdentifier, [False, True], [False, True]
+    ),
     'ClipIdentifier': IdentifierSerializationOptions(
         'clip_identifier', ['video', 'clip_id', 'frames'], ClipIdentifier, [False, False, False], [False, False, True]),
     'MultiFramesInputIdentifier': IdentifierSerializationOptions(
@@ -120,15 +129,31 @@ def deserialize_identifier(identifier):
     return identifier
 
 
+def create_ann_identifier_key(identifier):
+    if isinstance(identifier, list):
+        return ListIdentifier(tuple(create_ann_identifier_key(elem) for elem in identifier))
+    if isinstance(identifier, ClipIdentifier):
+        return ClipIdentifier(identifier.video, identifier.clip_id, tuple(identifier.frames))
+    if isinstance(identifier, MultiFramesInputIdentifier):
+        return MultiFramesInputIdentifier(tuple(identifier.input_id), tuple(identifier.frames))
+    if isinstance(identifier, ParametricImageIdentifier):
+        return ParametricImageIdentifier(identifier.identifier, tuple(identifier.parameters))
+    if isinstance(identifier, AnnotationDataIdentifier):
+        return identifier.annotation_id
+    return identifier
+
+
 def create_identifier_key(identifier):
     if isinstance(identifier, list):
-        return ListIdentifier(tuple(create_identifier_key(elem) for elem in identifier))
+        return ListIdentifier(tuple(create_ann_identifier_key(elem) for elem in identifier))
     if isinstance(identifier, ClipIdentifier):
         return ClipIdentifier(identifier.video, identifier.clip_id, tuple(identifier.frames))
     if isinstance(identifier, MultiFramesInputIdentifier):
         return MultiFramesInputIdentifier(tuple(identifier.input_id), tuple(identifier.frames))
     if isinstance(identifier, ParametricImageIdentifier):
         return ParametricImageIdentifier(identifier.identifier, tuple(identifier.parameters))
+    if isinstance(identifier, AnnotationDataIdentifier):
+        return AnnotationDataIdentifier(identifier.annotation_id, tuple(identifier.data_id))
     return identifier
 
 

tools/accuracy_checker/openvino/tools/accuracy_checker/dataset.py

@@ -53,7 +53,7 @@
 )
 from .data_readers import (
     DataReaderField, REQUIRES_ANNOTATIONS, BaseReader,
-    serialize_identifier, deserialize_identifier, create_identifier_key
+    serialize_identifier, deserialize_identifier, create_ann_identifier_key
 )
 from .logging import print_info
 
@@ -460,15 +460,15 @@ def __init__(self, annotations, meta, name='', config=None):
         self._data_buffer = OrderedDict()
         self._meta = meta
         for ann in annotations:
-            idx = create_identifier_key(ann.identifier)
+            idx = create_ann_identifier_key(ann.identifier)
             self._data_buffer[idx] = ann
 
     def __getitem__(self, item):
-        return self._data_buffer[item]
+        return self._data_buffer[create_ann_identifier_key(item)]
 
     @property
     def identifiers(self):
-        return list(self._data_buffer)
+        return list(map(lambda ann: ann.identifier, self._data_buffer.values()))
 
     def __len__(self):
         return len(self._data_buffer)

tools/accuracy_checker/openvino/tools/accuracy_checker/launcher/input_feeder.py

@@ -21,7 +21,11 @@
 from ..config import ConfigError
 from ..utils import extract_image_representations
 from ..data_readers import (
-    MultiFramesInputIdentifier, KaldiFrameIdentifier, KaldiMatrixIdentifier, ParametricImageIdentifier
+    MultiFramesInputIdentifier,
+    KaldiFrameIdentifier,
+    KaldiMatrixIdentifier,
+    ParametricImageIdentifier,
+    AnnotationDataIdentifier
 )
 
 LAYER_LAYOUT_TO_IMAGE_LAYOUT = {
@@ -192,6 +196,9 @@ def match_by_regex(data, identifiers, input_regex):
             for data_representation in data_representation_batch:
                 identifiers = data_representation.identifier
                 data = data_representation.data
+                if isinstance(identifiers, AnnotationDataIdentifier):
+                    identifiers = identifiers.data_id
+
                 if isinstance(identifiers, ParametricImageIdentifier):
                     input_batch.append(data[idx])
                     continue