Fixed opennmt evaluator in case running with not default beams (#3266)

Author: Anna Grebneva

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

@@ -118,7 +118,7 @@ def predict(self, identifiers, input_data, encoder_callback=None):
                 if encoder_callback:
                     encoder_callback(raw_outputs)
 
-                log_probs, raw_outputs = self.generator.predict(identifiers, {'input': decoder_output.squeeze()})
+                log_probs, raw_outputs = self.generator.predict(identifiers, {'input': decoder_output.squeeze(axis=0)})
                 if encoder_callback:
                     encoder_callback(raw_outputs)
 
@@ -313,7 +313,7 @@ def _pick(self, log_probs):
         curr_scores = log_probs.reshape(-1, self.beam_size * vocab_size)
         topk_ids = np.argsort(curr_scores)[..., range(self.beam_size * vocab_size - 1,
                                                       self.beam_size * (vocab_size - 1) - 1, -1)]
-        topk_scores = curr_scores[..., topk_ids.squeeze()]
+        topk_scores = curr_scores[..., topk_ids.squeeze(axis=0)]
         return topk_scores, topk_ids
 
     def update_finished(self):
@@ -373,7 +373,7 @@ def advance(self, log_probs):
         self.topk_ids = np.fmod(self.topk_ids, vocab_size)  # resolve true word ids
 
         self.alive_seq = np.concatenate(
-            [np.take(self.alive_seq, self.select_indices.squeeze(), 0),
+            [np.take(self.alive_seq, self.select_indices.squeeze(axis=-1), 0),
              self.topk_ids.view().reshape((_B * self.beam_size, 1))], axis=-1)
 
         self.is_finished = np.equal(self.topk_ids, self.eos)