fixed return value

handler.py

@@ -146,6 +146,51 @@ def get_model():
     return AurelioRNN
 
 
+def calculate_perplexity_on_text(model, text, seq_length, tokenizer):
+    loss_fn = nn.CrossEntropyLoss()
+
+    model.eval()
+    total_loss = 0
+    total_words = 0
+
+    # Tokenize the text
+    encoded = tokenizer.encode(text)
+    ids = encoded.ids
+    if len(ids) <= seq_length:
+        print(
+            "Input text is too short to calculate perplexity. length:",
+            len(ids),
+            "seq_length:",
+            seq_length,
+        )
+        return float(
+            "inf"
+        )
+
+    inputs = [ids[i : i + seq_length] for i in range(len(ids) - seq_length)]
+    targets = [ids[i + 1 : i + seq_length + 1] for i in range(len(ids) - seq_length)]
+
+    state_h, state_c = model.init_state(1)
+
+    with torch.no_grad():
+        for i in range(len(inputs)):
+            input_tensor = torch.tensor(inputs[i]).unsqueeze(0).to(device)
+            target_tensor = torch.tensor(targets[i]).unsqueeze(0).to(device)
+
+            output, (state_h, state_c) = model(
+                input_tensor, (state_h.detach(), state_c.detach())
+            )
+            loss = loss_fn(output.transpose(1, 2), target_tensor)
+            total_loss += loss.item()
+            total_words += seq_length
+
+    average_loss = total_loss / total_words
+
+    perplexity = np.exp(average_loss)
+    return perplexity
+
+
+
 class EndpointHandler:
     def __init__(self, path=""):
         # load the optimized model
@@ -183,7 +228,6 @@ class EndpointHandler:
             "embedding_dim": embedding_dim,
             "rnn_units": rnn_units,
         }
-        # load_from_hub
         lstm = get_model()
         model = lstm.from_pretrained("jed-tiotuico/aurelio-rnn", config=config)
         model.eval()  # Set the model to evaluation mode
@@ -191,7 +235,6 @@ class EndpointHandler:
         dir_path = os.path.abspath(os.path.dirname(__file__))
         # Load the Kapampangan vocabulary
         kapampangan_vocabulary = load_vocabulary_from_file(os.path.join(dir_path, "kapampangan.txt"))
-        # Define the source and destination paths
         seq_length = 64
 
         tokenizer = ByteLevelBPETokenizer(
@@ -210,7 +253,16 @@ class EndpointHandler:
                 vocabulary=kapampangan_vocabulary,
                 tokenizer=tokenizer,
             )
-            predictions.append(generated_text)
+            perplexity = calculate_perplexity_on_text(
+                model, generated_text, seq_length=seq_length - 1, tokenizer=tokenizer
+            )
+            predictions.append(
+                {
+                    "label": error_rate,
+                    "score": 1 - error_rate,
+                    "generated_text": generated_text,
+                    "perplexity": perplexity
+                }
+            )
 
-        # return preductions as concated string
         return predictions