[HN] it works but why is it gibberish? ugh

app.py

@@ -6,47 +6,64 @@
 
 @app.route('/')
 def home():
-    return render_template('home.html')
+    return render_template('index.html')
 
 
 @app.route('/predict', methods=['POST'])
 def predict():
 
-    def inference(question):
-        import tensorflow as tf
-        from lstm import EncoderDecoder, Encoder, Decoder, inference
-        import pickle
-        import numpy as np 
+    import tensorflow as tf
+    from lstm import EncoderDecoder, Encoder, Decoder, inference
+    import pickle
+    import numpy as np
+
+    char2idx = pickle.load(open('char2idx.pkl', 'rb'))
+    idx2char = pickle.load(open('idx2char.pkl', 'rb'))
+
+    def output_to_tensor(tokens):
+        tensor_tokens = tf.squeeze(tf.convert_to_tensor(tokens), axis=2)
+        return tf.transpose(tensor_tokens)
+
+    def token_to_text(batch_tensor):
+        batch_array = batch_tensor.numpy()
+        text_outputs = []
+        for sequence_pred in batch_array:
+            text = ''.join([idx2char[pred] for pred in sequence_pred])
+            text_outputs.append(text)
+        return text_outputs
 
-        char2idx = pickle.load(open('char2idx.pkl', 'rb'))
-        idx2char = pickle.load(open('idx2char.pkl', 'rb'))
+    def inference(question):
 
-        VOCAB_SIZE = len(char2idx) + 2
+        VOCAB_SIZE = len(char2idx)
 
-        ANSWER_MAX_LENGTH=30
-        BATCH_SIZE = 128
+        ANSWER_MAX_LENGTH = 30
         EMBEDDING_SIZE = 512
-        LSTM_HIDDEN_SIZE = 1024
-        NUM_EPOCHS = 50
-        NUM_EXAMPLES = 666666 * 3
-        p_test = .2
+        LSTM_HIDDEN_SIZE = 512
 
         questions_encoded = [char2idx[q] for q in question]
         answers_encoded = [char2idx[q] for q in question]
 
+        questions_encoded = np.expand_dims(np.array(questions_encoded), axis=0)
+
         # TODO preprocess into single arrays
-        dataset = tf.data.Dataset.from_tensor_slices((questions_encoded, answers_encoded))
-        input_data = dataset.take(1).batch(1)
+        #dataset = tf.data.Dataset.from_tensor_slices((questions_encoded, answers_encoded))
+        #input_data = dataset.take(1).batch(1)
+
+        #data = input_data[0]
 
         model = EncoderDecoder(input_dim=VOCAB_SIZE, embedding_dim=EMBEDDING_SIZE, hidden_dim=LSTM_HIDDEN_SIZE,
                                output_dim=VOCAB_SIZE, max_len=ANSWER_MAX_LENGTH)
         # model.save_weights('experiment_results/test')
         model.load_weights('model_weights')
 
         print(questions_encoded)
-        outputs, output_tokens = model(np.expand_dims(np.array(questions_encoded), axis=1))
+        outputs, output_tokens = model(questions_encoded)
+
+        predicted_text = token_to_text(output_to_tensor(output_tokens))[0]
+        first_stop_token = predicted_text.index('~')
+        predicted_text = predicted_text[0:first_stop_token]
 
-        return outputs, output_tokens
+        return predicted_text
 
     if request.method == 'POST':
         message = request.form['message']

config.py

@@ -5,7 +5,7 @@
 
 BATCH_SIZE = 128
 EMBEDDING_SIZE = 512
-LSTM_HIDDEN_SIZE = 1024
+LSTM_HIDDEN_SIZE = 512
 NUM_EPOCHS = 50
 NUM_EXAMPLES = 666666*3
 p_test = .2