AC: fix custom evaluator for custom asr model (#3423)

* AC: fix custom evaluator for custom asr model

* more fixes

Author: eaidova

tools/accuracy_checker/openvino/tools/accuracy_checker/evaluators/custom_evaluators/asr_custom_encoder_decoder_joint.py

@@ -64,10 +64,10 @@ def predict(self, identifiers, input_data):
             outputs, raw_outputs = outputs
         else:
             raw_outputs = outputs
-        encoder_output = np.array(outputs[self.encoder_out]).squeeze()
+        encoder_output = outputs[self.encoder_out]
         self.h0 = outputs[self.h0_out]
         self.c0 = outputs[self.c0_out]
-        return encoder_output, raw_outputs
+        return encoder_output.squeeze(), raw_outputs
 
     def fit_to_input(self, input_data):
         return {self.input: input_data, self.h0_input: self.h0, self.c0_input: self.c0}
@@ -127,7 +127,7 @@ def predict(self, identifiers, input_data, hidden=None):
             raw_outputs = outputs
         self.h0 = outputs[self.h0_out]
         self.c0 = outputs[self.c0_out]
-        return np.array(outputs[self.decoder_out]).squeeze(), (self.h0, self.c0), raw_outputs
+        return outputs[self.decoder_out].squeeze(), (self.h0, self.c0), raw_outputs
 
     def fit_to_input(self, token_id, hidden):
         if hidden is None:
@@ -189,7 +189,7 @@ def predict(self, identifiers, input_data):
         else:
             raw_outputs = outputs
         joint_out = outputs[self.output]
-        return log_softmax(np.array(joint_out).squeeze()), raw_outputs
+        return log_softmax(joint_out), raw_outputs
 
     def fit_to_input(self, encoder_out, predictor_out):
         return {self.input1: encoder_out, self.input2: predictor_out}
@@ -339,7 +339,7 @@ def __init__(self, network_info, launcher, suffix=None, delayed_model_loading=Fa
 class OVJoint(Joint, CommonOpenVINOModel):
     def __init__(self, network_info, launcher, suffix=None, delayed_model_loading=False):
         self.default_inputs = ['0', '1']
-        self.default_outputs = ['8/sink_port']
+        self.default_outputs = ['8/sink_port_0']
         super().__init__(network_info, launcher, suffix, delayed_model_loading)
 
 
@@ -353,25 +353,22 @@ def infer(self, input_data):
         results = self.inference_session.run(self.output_names, input_data)
         return dict(zip(self.output_names, results))
 
-    def select_inputs_outputs(self, network_info):
-        pass
-
 
-class ONNXEncoder(CommonONNXModel, Encoder):
+class ONNXEncoder(Encoder, CommonONNXModel):
     def __init__(self, network_info, launcher, suffix=None, delayed_model_loading=False):
         self.default_inputs = ['input_0', 'input_1', 'input_2']
         self.default_outputs = ['output_0', 'output_1', 'output_2']
         super().__init__(network_info, launcher, suffix, delayed_model_loading)
 
 
-class ONNXDecoder(CommonONNXModel, Decoder):
+class ONNXDecoder(Decoder, CommonONNXModel):
     def __init__(self, network_info, launcher, suffix=None, delayed_model_loading=False):
         self.default_inputs = ['input_0', 'input_1', 'input_2']
         self.default_outputs = ['output_0', 'output_1', 'output_2']
         super().__init__(network_info, launcher, suffix, delayed_model_loading)
 
 
-class ONNXJoint(CommonONNXModel, Joint):
+class ONNXJoint(Joint, CommonONNXModel):
     def __init__(self, network_info, launcher, suffix=None, delayed_model_loading=False):
         self.default_inputs = ['0', '1']
         self.default_outputs = ['8']
@@ -454,7 +451,7 @@ def predict(self, identifiers, input_data, encoder_callback=None):
                 if len(B) >= self.beam_width and yb.log_prob >= y_hat.log_prob:
                     break
             B = heapq.nlargest(self.beam_width, B)
-            return self.adapter.process([B[0].sequence], identifiers, [{}]), {}
+        return [{}], self.adapter.process([B[0].sequence], identifiers, [{}])
 
     @staticmethod
     def prepare_records(features):

tools/accuracy_checker/openvino/tools/accuracy_checker/evaluators/custom_evaluators/base_models.py

@@ -169,13 +169,15 @@ def automatic_model_search(self, network_info):
                 if not model_list:
                     model_list = list(model.glob('*.blob'))
             else:
-                model_list = list(model.glob('*{}.xml'.format(self.default_model_suffix)))
-                blob_list = list(model.glob('*{}.blob'.format(self.default_model_suffix)))
-                if not model_list and not blob_list:
+                model_list = list(model.glob('*{}*.xml'.format(self.default_model_suffix)))
+                blob_list = list(model.glob('*{}*.blob'.format(self.default_model_suffix)))
+                onnx_list = list(model.glob('*{}*.onnx'.format(self.default_model_suffix)))
+                if not model_list and not blob_list and not onnx_list:
                     model_list = list(model.glob('*.xml'))
                     blob_list = list(model.glob('*.blob'))
-                    if not model_list:
-                        model_list = blob_list
+                    onnx_list = list(model.glob('*.onnx'))
+                if not model_list:
+                    model_list = blob_list if blob_list else onnx_list
             if not model_list:
                 raise ConfigError('Suitable model for {} not found'.format(self.default_model_suffix))
             if len(model_list) > 1:
@@ -391,7 +393,7 @@ def release(self):
     def automatic_model_search(self, network_info):
         model = Path(network_info['model'])
         if model.is_dir():
-            model_list = list(model.glob('*{}.onnx'.format(self.default_model_suffix)))
+            model_list = list(model.glob('*{}*.onnx'.format(self.default_model_suffix)))
             if not model_list:
                 model_list = list(model.glob('*.onnx'))
             if not model_list: