Fixed beamsearch

CapGenerator/eval_model.py

@@ -1,5 +1,5 @@
 from pickle import load
-from numpy import argmax
+import numpy as np
 from keras.preprocessing.sequence import pad_sequences
 from keras.applications.vgg16 import VGG16
 from keras.preprocessing.image import load_img
@@ -53,29 +53,24 @@ def generate_desc(model, tokenizer, photo, index_word, max_length, beam_size=10)
       # pad input
       sequence = pad_sequences([sequence], maxlen=max_length)
       # predict next words
-      yhat = model.predict([photo,sequence], verbose=0)
+      y_pred = model.predict([photo,sequence], verbose=0)[0]
       # convert probability to integer
-      yhats = np.argsort(yhat)[-beam_size:]
+      yhats = np.argsort(y_pred)[-beam_size:]
 
       for j in yhats:
         # map integer to word
-        word = index_word.get(yhat)
+        word = index_word.get(j)
         # stop if we cannot map the word
         if word is None:
           continue
-        # Add word to caption, and generate negative log prob
-        caption = [sentence + ' ' + word, score - np.log(yhat[j])]
+        # Add word to caption, and generate log prob
+        caption = [sentence + ' ' + word, score + np.log(y_pred[j])]
         all_caps.append(caption)
 
     # order all candidates by score
     ordered = sorted(all_caps, key=lambda tup:tup[1], reverse=True)
     captions = ordered[:beam_size]
 
-    # append as input for generating the next word
-    in_text += ' ' + word
-    # stop if we predict the end of the sequence
-
-      break
   return captions
 
 # evaluate the skill of the model
@@ -125,7 +120,10 @@ def evaluate_model(model, descriptions, photos, tokenizer, index_word, max_lengt
     # generate description
     captions = generate_desc(model, tokenizer, photo, index_word, max_length)
     for cap in captions:
-      print('{}. log_prob: {}'.format(cap[0],cap[1]))
+      # remove start and end tokens
+      seq = cap[0].split()[1:-1]
+      desc = ' '.join(seq)
+      print('{} [log prob: {:1.2f}]'.format(desc,cap[1]))
   else:
     # load test set
     test_features, test_descriptions = ld.prepare_dataset('test')[1]