Bug fixes

PiperOrigin-RevId: 175611828

Author: Ryan Sepassi

CONTRIBUTING.md

@@ -1,5 +1,15 @@
 # How to Contribute
 
+# Issues
+
+* Please tag your issue with `bug`, `feature request`, or `question` to help us
+  effectively respond.
+* Please include the versions of TensorFlow and Tensor2Tensor you are running
+  (run `pip list | grep tensor`)
+* Please provide the command line you ran as well as the log output.
+
+# Pull Requests
+
 We'd love to accept your patches and contributions to this project. There are
 just a few small guidelines you need to follow.
 

tensor2tensor/bin/t2t-datagen

@@ -43,7 +43,6 @@ from tensor2tensor.data_generators import all_problems  # pylint: disable=unused
 from tensor2tensor.data_generators import audio
 from tensor2tensor.data_generators import generator_utils
 from tensor2tensor.data_generators import snli
-from tensor2tensor.data_generators import translate
 from tensor2tensor.data_generators import wsj_parsing
 from tensor2tensor.utils import registry
 from tensor2tensor.utils import usr_dir
@@ -82,10 +81,10 @@ _SUPPORTED_PROBLEM_GENERATORS = {
         lambda: algorithmic_math.algebra_inverse(26, 0, 2, 100000),
         lambda: algorithmic_math.algebra_inverse(26, 3, 3, 10000)),
     "parsing_english_ptb8k": (
-        lambda: translate.parsing_token_generator(
-            FLAGS.data_dir, FLAGS.tmp_dir, True, 2**13),
-        lambda: translate.parsing_token_generator(
-            FLAGS.data_dir, FLAGS.tmp_dir, False, 2**13)),
+        lambda: wsj_parsing.parsing_token_generator(
+            FLAGS.data_dir, FLAGS.tmp_dir, True, 2**13, 2**9),
+        lambda: wsj_parsing.parsing_token_generator(
+            FLAGS.data_dir, FLAGS.tmp_dir, False, 2**13, 2**9)),
     "parsing_english_ptb16k": (
         lambda: wsj_parsing.parsing_token_generator(
             FLAGS.data_dir, FLAGS.tmp_dir, True, 2**14, 2**9),

tensor2tensor/models/shake_shake.py

@@ -132,6 +132,8 @@ def model_fn_body(self, features):
 @registry.register_hparams
 def shakeshake_cifar10():
   """Parameters for CIFAR-10."""
+  tf.logging.warning("shakeshake_cifar10 hparams have not been verified to "
+                     "achieve good performance.")
   hparams = common_hparams.basic_params1()
   # This leads to effective batch size 128 when number of GPUs is 1
   hparams.batch_size = 4096 * 8

tensor2tensor/utils/expert_utils.py

@@ -200,8 +200,8 @@ def daisy_chain_getter(getter, name, *args, **kwargs):
         else:
           var = getter(name, *args, **kwargs)
           v = tf.identity(var._ref())  # pylint: disable=protected-access
+          _add_variable_proxy_methods(var, v)
         # update the cache
-        _add_variable_proxy_methods(var, v)
         cache[name] = v
         cache[device_var_key] = v
         return v
@@ -210,10 +210,12 @@ def daisy_chain_getter(getter, name, *args, **kwargs):
       # so we make a custom getter that uses identity to cache the variable.
       # pylint: disable=cell-var-from-loop
       def caching_getter(getter, name, *args, **kwargs):
-        v = getter(name, *args, **kwargs)
+        """Cache variables on device."""
         key = (self._caching_devices[i], name)
         if key in cache:
           return cache[key]
+
+        v = getter(name, *args, **kwargs)
         with tf.device(self._caching_devices[i]):
           ret = tf.identity(v._ref())  # pylint: disable=protected-access
         _add_variable_proxy_methods(v, ret)