did some more tests

insurance_qa_eval.py

@@ -253,7 +253,7 @@ if __name__ == '__main__':
             'conv_activation': 'relu',
 
             # recurrent
-            'n_lstm_dims': 141,
+            'n_lstm_dims': 141, # * 2
 
             'initial_embed_weights': np.load('word2vec_100_dim.embeddings'),
         },
@@ -269,7 +269,7 @@ if __name__ == '__main__':
     evaluator = Evaluator(conf)
 
     ##### Define model ######
-    model = ConvolutionModel(conf)
+    model = AttentionModel(conf)
     optimizer = conf.get('training_params', dict()).get('optimizer', 'adam')
     model.compile(optimizer=optimizer)
 
@@ -288,5 +288,5 @@ if __name__ == '__main__':
     evaluator.train(model)
 
     # evaluate mrr for a particular epoch
-    # evaluator.load_epoch(model, 22)
+    # evaluator.load_epoch(model, 32)
     # evaluator.get_mrr(model, evaluate_all=True)

keras_models.py

@@ -257,8 +257,8 @@ class AttentionModel(LanguageModel):
         answer_embedding = embedding(answer)
 
         # turn off layer updating
-        embedding.params = []
-        embedding.updates = []
+        # embedding.params = []
+        # embedding.updates = []
 
         # dropout
         dropout = Dropout(0.25)

results.notes

@@ -13,12 +13,12 @@ Embedding + Max Pooling:
 Dense + CNN + Max Pooling:
  - Top 1 precision:
    - 0.507 on test 1
-   - 0.427 on test 2
-   - 0.470 on dev
+   - 0.458 on test 2
+   - 0.515 on dev
  - MRR:
    - 0.635 on test 1
-   - 0.557 on test 2
-   - 0.596 on dev
+   - 0.593 on test 2
+   - 0.642 on dev
 
 Attentional LSTM + Max Pooling:
  - Top 1 precision: