AC: update yolact adapter for model with embedded decoder (openvinoto…

…olkit#1673)

tools/accuracy_checker/accuracy_checker/adapters/README.md

@@ -259,8 +259,9 @@ AccuracyChecker supports following set of adapters:
   * `texts_out` - name of output layer with texts.
   * `confidence_threshold` - confidence threshold that is used to filter out detected instances.
 * `yolact` - converting raw outputs of Yolact model to to combination of `DetectionPrediction` and `CoCocInstanceSegmentationPrediction`.
-  * `loc_out` - name of output layer which contains box locations.
-  * `prior_out` - name of output layer which contains prior boxes.
+  * `loc_out` - name of output layer which contains box locations, optional if boxes decoding embedded into model.
+  * `prior_out` - name of output layer which contains prior boxes, optional if boxes decoding embedded into model.
+  * `boxes_out` - name of output layer which contains decoded output boxes, optional if model has `prior` a `loc` outputs for boxes decoding.
   * `conf_out` - name of output layer which contains confidence scores for all classes for each box.
   * `mask_out` - name of output layer which contains instance masks.
   * `proto_out` - name of output layer which contains proto for masks calculation.

tools/accuracy_checker/accuracy_checker/adapters/yolact.py

@@ -17,7 +17,7 @@
 import cv2
 import numpy as np
 
-from ..config import StringField, NumberField
+from ..config import StringField, NumberField, ConfigError
 from ..postprocessor import NMS
 from ..representation import DetectionPrediction, CoCocInstanceSegmentationPrediction, ContainerPrediction
 from ..utils import UnsupportedPackage
@@ -35,9 +35,10 @@ class YolactAdapter(Adapter):
     def parameters(cls):
         params = super().parameters()
         params.update({
-            'loc_out': StringField(description='name of output with box locations'),
+            'boxes_out': StringField(description='name of output with decoded boxes', optional=True),
+            'loc_out': StringField(description='name of output with box locations', optional=True),
             'conf_out': StringField(description='name of output with confidence scores'),
-            'prior_out': StringField(description='name of output with prior boxes'),
+            'prior_out': StringField(description='name of output with prior boxes', optional=True),
             'mask_out': StringField(description='name of output with masks'),
             'proto_out': StringField(description='name of output with proto for masks calculation'),
             'confidence_threshold':  NumberField(
@@ -56,21 +57,25 @@ def configure(self):
         self.loc_out = self.get_value_from_config('loc_out')
         self.conf_out = self.get_value_from_config('conf_out')
         self.prior_out = self.get_value_from_config('prior_out')
+        self.boxes_out = self.get_value_from_config('boxes_out')
         self.mask_out = self.get_value_from_config('mask_out')
         self.proto_out = self.get_value_from_config('proto_out')
         self.conf_thresh = self.get_value_from_config('confidence_threshold')
         self.max_num_detections = self.get_value_from_config('max_detections')
+        if not self.loc_out and not self.prior_out and not self.boxes_out:
+            raise ConfigError('loc_out and prior_out or boxes_out should be provided')
+        if not self.boxes_out and not (self.prior_out and self.loc_out):
+            raise ConfigError('both loc_out and prior_out should be provided')
 
     def process(self, raw, identifiers, frame_meta):
         raw_outputs = self._extract_predictions(raw, frame_meta)
-        prior_boxes = raw_outputs[self.prior_out]
         result = []
-        for identifier, locs, conf, masks, proto, meta in zip(
-                identifiers, raw_outputs[self.loc_out], raw_outputs[self.conf_out],
+        for batch_id, (identifier, conf, masks, proto, meta) in enumerate(zip(
+                identifiers, raw_outputs[self.conf_out],
                 raw_outputs[self.mask_out], raw_outputs[self.proto_out], frame_meta
-        ):
+        )):
             h, w, _ = meta['image_size']
-            boxes = self.decode_boxes(locs, prior_boxes)
+            boxes = self.get_boxes(batch_id, raw_outputs)
             conf = np.transpose(conf)
             cur_scores = conf[1:, :]
             conf_scores = np.max(cur_scores, axis=0)
@@ -179,6 +184,15 @@ def decode_boxes(loc, priors):
 
         return boxes
 
+    def get_boxes(self, batch_id, raw_outputs):
+        if self.boxes_out:
+            boxes = raw_outputs[self.boxes_out][batch_id]
+        else:
+            prior_boxes = raw_outputs[self.prior_out]
+            locs = raw_outputs[self.loc_out][batch_id]
+            boxes = self.decode_boxes(locs, prior_boxes)
+        return boxes
+
     @staticmethod
     def mask_proto_mask_activation(masks):
         return 1 / (1 + np.exp(-masks))