update

inferbeddings/lm/model.py

@@ -77,6 +77,23 @@ def loop(prev, _):
 
         self.random_state = np.random.RandomState(seed)
 
+    def score_sequence(self, session, sequence):
+        x = np.zeros((1, 1))
+        state = session.run(self.cell.zero_state(1, tf.float32))
+        res = 0.0
+        for i, idx in enumerate(sequence):
+            x[0, 0] = idx
+            feed = {
+                self.input_data: x,
+                self.initial_state: state
+            }
+            probabilities, state = session.run([self.probabilities, self.final_state], feed)
+            if i < len(sequence) - 1:
+                next_idx = sequence[i + 1]
+                res += np.log(probabilities[0, next_idx])
+        print(res)
+        return res
+
     def sample(self, session, words, vocab, num=200, prime='first all', sampling_type=1, pick=0, width=4):
         def weighted_pick(weights):
             t = np.cumsum(weights)

tests/inferbeddings/lm/test_sample.py

@@ -65,7 +65,7 @@ def test_lm_sample():
         ckpt = tf.train.get_checkpoint_state(lm_path)
         saver.restore(session, ckpt.model_checkpoint_path)
 
-        for _ in range(32):
+        for _ in range(4):
             sample_value = imodel.sample(session=session,
                                          words=index_to_token,
                                          vocab=token_to_index,
@@ -76,6 +76,10 @@ def test_lm_sample():
                                          width=4)
             print(sample_value)
 
+        sequence = [token_to_index[w] for w in ['A', 'girl', 'runs']]
+        imodel.score_sequence(session=session, sequence=sequence)
+
+
 if __name__ == '__main__':
     #pytest.main([__file__])
     test_lm_sample()