best version 1.2: four model

hw5/best.py

@@ -124,16 +124,30 @@ def main():
  predict3[predict3 >= threshold] = 1
 
 
+ ### RNN 2
+ model = load_model(model_name2, custom_objects={'fmeasure': fmeasure})
+ predict4 = model.predict(sequences_test)
+ predict4[predict4 < threshold] = 0
+ predict4[predict4 >= threshold] = 1
+
+
  ### Voting
- pred = predict + predict2 + predict3
- pred[pred < 1.5] = 0
- pred[pred >= 1.5] = 1
+ pred = predict + predict2 + predict3 + predict4
+ check = predict3 + predict4
+ pred[pred < 2] = 0
+ pred[pred > 2] = 1
+ x = np.argwhere(pred == 2)
+ for i, j in x:
+ if check[i, j] == 2:
+ pred[i, j] = 0
+ else:
+ pred[i, j] = 1
 
 
  # Test data
  ensure_dir(output_path)
  result = []
- mlb_backup = mlb.inverse_transform(predict3)
+ mlb_backup = mlb.inverse_transform(predict4)
  for i, categories in enumerate(mlb.inverse_transform(pred)):
  ret = []
  if len(categories) == 0:
@@ -169,7 +183,8 @@ def main():
  vectorizer_name2 = os.path.join(base_dir, 'vec2')
  linear_svc_name2 = os.path.join(base_dir, 'linSVC2')
  tokenizer_name = os.path.join(base_dir, 'word_index')
- model_name = os.path.join(base_dir, 'model-5.h5')
+ model_name = os.path.join(base_dir, 'model-6.h5')
+ model_name2 = os.path.join(base_dir, 'model-5.h5')
  threshold = 0.4
 
  main()