#13 Fix eager mode

.travis.yml

@@ -6,6 +6,7 @@ python:
 env:
   - KERAS_BACKEND=tensorflow
   - KERAS_BACKEND=tensorflow TF_KERAS=1
+  - KERAS_BACKEND=tensorflow TF_KERAS=1 TF_EAGER=1
   - KERAS_BACKEND=theano THEANO_FLAGS=optimizer=fast_compile
   # - KERAS_BACKEND=cntk PYTHONWARNINGS=ignore
 install:

keras_transformer/backend.py

@@ -3,11 +3,17 @@
 __all__ = [
     'keras', 'utils', 'activations', 'applications', 'backend', 'datasets', 'engine',
     'layers', 'preprocessing', 'wrappers', 'callbacks', 'constraints', 'initializers',
-    'metrics', 'models', 'losses', 'optimizers', 'regularizers',
+    'metrics', 'models', 'losses', 'optimizers', 'regularizers', 'EAGER_MODE'
 ]
 
+EAGER_MODE = False
+
 if 'TF_KERAS' in os.environ and os.environ['TF_KERAS'] != '0':
     from tensorflow.python import keras
+    if 'TF_EAGER' in os.environ and os.environ['TF_EAGER'] != '0':
+        import tensorflow as tf
+        tf.enable_eager_execution()
+        EAGER_MODE = True
 else:
     import keras
 

keras_transformer/transformer.py

@@ -415,7 +415,14 @@ def _get_max_suffix_repeat_times(tokens, max_len):
     return max_repeat
 
 
-def decode(model, tokens, start_token, end_token, pad_token, max_len=10000, max_repeat=10, max_repeat_block=10):
+def decode(model,
+           tokens,
+           start_token,
+           end_token,
+           pad_token,
+           max_len=10000,
+           max_repeat=10,
+           max_repeat_block=10):
     """Decode with the given model and input tokens.
 
     :param model: The trained model.
@@ -448,7 +455,7 @@ def decode(model, tokens, start_token, end_token, pad_token, max_len=10000, max_
                 max_input_len = max(max_input_len, len(tokens[i]))
         for i in range(len(batch_inputs)):
             batch_inputs[i] += [pad_token] * (max_input_len - len(batch_inputs[i]))
-        predicts = model.predict([np.asarray(batch_inputs), np.asarray(batch_outputs)])
+        predicts = model.predict([np.array(batch_inputs), np.array(batch_outputs)])
         for i in range(len(predicts)):
             last_token = np.argmax(predicts[i][-1])
             decoder_inputs[index_map[i]].append(last_token)

requirements.txt

@@ -4,4 +4,4 @@ keras-pos-embd==0.10.0
 keras-multi-head==0.20.0
 keras-layer-normalization==0.12.0
 keras-position-wise-feed-forward==0.5.0
-keras-embed-sim==0.4.0
+keras-embed-sim==0.5.0

setup.py

@@ -12,7 +12,7 @@
 
 setup(
     name='keras-transformer',
-    version='0.24.0',
+    version='0.25.0',
     packages=find_packages(),
     url='https://github.com/CyberZHG/keras-transformer',
     license='MIT',

tests/test_decode.py

@@ -1,7 +1,7 @@
 import os
 import unittest
 import numpy as np
-from keras_transformer.backend import keras
+from keras_transformer.backend import keras, EAGER_MODE
 from keras_transformer import get_custom_objects, get_model, decode
 
 
@@ -57,8 +57,10 @@ def test_decode(self):
                 epochs=10,
                 batch_size=128,
             )
-            model.save(model_path)
-        model = keras.models.load_model(model_path, custom_objects=get_custom_objects())
+            if not EAGER_MODE:
+                model.save(model_path)
+        if not EAGER_MODE:
+            model = keras.models.load_model(model_path, custom_objects=get_custom_objects())
         decoded = decode(
             model,
             encoder_inputs_no_padding * 2,

tests/test_gelu.py

@@ -9,7 +9,10 @@ def test_sample(self):
         try:
             results = gelu(K.constant([-30.0, -1.0, 0.0, 1.0, 30.0])).eval(session=K.get_session())
         except Exception as e:
-            results = gelu(K.constant([-30.0, -1.0, 0.0, 1.0, 30.0])).eval()
+            try:
+                results = gelu(K.constant([-30.0, -1.0, 0.0, 1.0, 30.0])).eval()
+            except Exception as e:
+                results = gelu(K.constant([-30.0, -1.0, 0.0, 1.0, 30.0])).numpy()
         self.assertEqual(0.0, results[0])
         self.assertGreater(0.0, results[1])
         self.assertLess(-1.0, results[1])