change to beam search strategy

app.py

@@ -9,7 +9,7 @@ from inference import Inference
 import utils
 
 encoder_model_name = 'google/vit-large-patch32-224-in21k'
-decoder_model_name = 'gpt2'
+decoder_model_name = 'gpt2-large'
 frame_step = 300
 
 inference = Inference(

inference.py

@@ -23,7 +23,12 @@ class Inference:
       self.tokenizer.add_special_tokens({'pad_token': '[PAD]'})
       self.encoder_decoder_model.decoder.resize_token_embeddings(len(self.tokenizer))
 
-    generated_ids = self.encoder_decoder_model.generate(pixel_values.unsqueeze(0).to(self.device), max_length=self.max_length)
+    generated_ids = self.encoder_decoder_model.generate(
+      pixel_values.unsqueeze(0).to(self.device),
+      max_length=self.max_length,
+      num_beams=4, 
+      no_repeat_ngram_size=2,
+    )
     generated_text = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
 
     return generated_text