Working version

bin/train.py

@@ -32,7 +32,6 @@
 from tensorflow import gfile
 
 from seq2seq import models
-from seq2seq.contrib.experiment import Experiment as PatchedExperiment
 from seq2seq.configurable import _maybe_load_yaml, _create_from_dict
 from seq2seq.configurable import _deep_merge_dict
 from seq2seq.data import input_pipeline
@@ -79,7 +78,7 @@
                        to. If None, a local temporary directory is created.""")
 
 # Training parameters
-tf.flags.DEFINE_string("schedule", "continuous_train_and_eval",
+tf.flags.DEFINE_string("schedule", None,
                        """Estimator function to call, defaults to
                        continuous_train_and_eval for local run""")
 tf.flags.DEFINE_integer("train_steps", None,
@@ -106,13 +105,13 @@
                         """In addition to keeping the most recent checkpoint
                         files, keep one checkpoint file for every N hours of
                         training.""")
-tf.flags.DEFINE_float("gpu_memory_fraction", 1.0,
+tf.flags.DEFINE_float("gpu_memory_fraction", 0.6,
                       """Fraction of GPU memory used by the process on
                       each GPU uniformly on the same machine.""")
-tf.flags.DEFINE_boolean("gpu_allow_growth", False,
+tf.flags.DEFINE_boolean("gpu_allow_growth", True,
                         """Allow GPU memory allocation to grow
                         dynamically.""")
-tf.flags.DEFINE_boolean("log_device_placement", False,
+tf.flags.DEFINE_boolean("log_device_placement", True,
                         """Log the op placement to devices""")
 
 
@@ -126,14 +125,16 @@ def create_experiment(output_dir):
     output_dir: Output directory for model checkpoints and summaries.
   """
 
+  session_config = tf.ConfigProto()
+  session_config.gpu_options.allow_growth = FLAGS.gpu_allow_growth
+  session_config.gpu_options.per_process_gpu_memory_fraction = FLAGS.gpu_memory_fraction
   config = run_config.RunConfig(
       tf_random_seed=FLAGS.tf_random_seed,
       save_checkpoints_secs=FLAGS.save_checkpoints_secs,
       save_checkpoints_steps=FLAGS.save_checkpoints_steps,
       keep_checkpoint_max=FLAGS.keep_checkpoint_max,
       keep_checkpoint_every_n_hours=FLAGS.keep_checkpoint_every_n_hours,
-      gpu_memory_fraction=FLAGS.gpu_memory_fraction)
-  config.tf_config.gpu_options.allow_growth = FLAGS.gpu_allow_growth
+      session_config=session_config)
   config.tf_config.log_device_placement = FLAGS.log_device_placement
 
   train_options = training_utils.TrainOptions(
@@ -203,7 +204,7 @@ def model_fn(features, labels, params, mode):
     metric = _create_from_dict(dict_, metric_specs)
     eval_metrics[metric.name] = metric
 
-  experiment = PatchedExperiment(
+  experiment = tf.contrib.learn.Experiment(
       estimator=estimator,
       train_input_fn=train_input_fn,
       eval_input_fn=eval_input_fn,

example_configs/nmt_large.yml

@@ -26,7 +26,7 @@ model_params:
   optimizer.name: Adam
   optimizer.params:
     epsilon: 0.0000008
-  optimizer.learning_rate: 0.0001
+  optimizer.learning_rate: 0.001
   source.max_seq_len: 50
   source.reverse: false
   target.max_seq_len: 50

example_configs/nmt_medium.yml

@@ -24,9 +24,7 @@ model_params:
       dropout_output_keep_prob: 1.0
       num_layers: 2
   optimizer.name: Adam
-  optimizer.params:
-    epsilon: 0.0000008
-  optimizer.learning_rate: 0.0001
+  optimizer.learning_rate: 0.001
   source.max_seq_len: 50
   source.reverse: false
   target.max_seq_len: 50

example_configs/nmt_small.yml

@@ -26,7 +26,7 @@ model_params:
   optimizer.name: Adam
   optimizer.params:
     epsilon: 0.0000008
-  optimizer.learning_rate: 0.0001
+  optimizer.learning_rate: 0.001
   source.max_seq_len: 50
   source.reverse: false
   target.max_seq_len: 50

example_configs/text_metrics_bpe.yml

@@ -7,31 +7,3 @@ metrics:
   - class: BleuMetricSpec
     params:
       <<: *default_params
-  - class: RougeMetricSpec
-    params:
-      <<: *default_params
-      rouge_type: rouge_1/f_score
-  - class: RougeMetricSpec
-    params:
-      <<: *default_params
-      rouge_type: rouge_1/r_score
-  - class: RougeMetricSpec
-    params:
-      <<: *default_params
-      rouge_type: rouge_1/p_score
-  - class: RougeMetricSpec
-    params:
-      <<: *default_params
-      rouge_type: rouge_2/f_score
-  - class: RougeMetricSpec
-    params:
-      <<: *default_params
-      rouge_type: rouge_2/r_score
-  - class: RougeMetricSpec
-    params:
-      <<: *default_params
-      rouge_type: rouge_2/p_score
-  - class: RougeMetricSpec
-    params:
-      <<: *default_params
-      rouge_type: rouge_l/f_score

example_configs/train_seq2seq.yml

@@ -1,8 +1,8 @@
-buckets: 10,20,30,40
-hooks:
-  - class: PrintModelAnalysisHook
-  - class: MetadataCaptureHook
-  - class: SyncReplicasOptimizerHook
-  - class: TrainSampleHook
-    params:
-      every_n_steps: 1000
+buckets: 10,20,30,40
+hooks:
+  - class: PrintModelAnalysisHook
+  - class: MetadataCaptureHook
+  - class: SyncReplicasOptimizerHook
+  - class: TrainSampleHook
+    params:
+      every_n_steps: 1000

seq2seq/contrib/seq2seq/helper.py

@@ -32,8 +32,8 @@
 
 import six
 
-from tensorflow.contrib.distributions.python.ops import bernoulli
-from tensorflow.contrib.distributions.python.ops import categorical
+from tensorflow.contrib.distributions import Bernoulli
+from tensorflow.contrib.distributions import Categorical
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import ops
 from tensorflow.python.layers import base as layers_base
@@ -264,7 +264,7 @@ def sample(self, time, outputs, state, name=None):
       select_sample_noise = random_ops.random_uniform(
           [self.batch_size], seed=self._scheduling_seed)
       select_sample = (self._sampling_probability > select_sample_noise)
-      sample_id_sampler = categorical.Categorical(logits=outputs)
+      sample_id_sampler = Categorical(logits=outputs)
       return array_ops.where(
           select_sample,
           sample_id_sampler.sample(seed=self._seed),
@@ -384,8 +384,8 @@ def initialize(self, name=None):
   def sample(self, time, outputs, state, name=None):
     with ops.name_scope(name, "ScheduledOutputTrainingHelperSample",
                         [time, outputs, state]):
-      sampler = bernoulli.Bernoulli(probs=self._sampling_probability)
-      return math_ops.cast(
+        sampler = Bernoulli(probs=self._sampling_probability)
+    return math_ops.cast(
           sampler.sample(sample_shape=self.batch_size, seed=self._seed),
           dtypes.bool)
 

seq2seq/test/pipeline_test.py

@@ -39,10 +39,12 @@
 
 
 def _clear_flags():
-  """Resets Tensorflow's FLAG values"""
-  #pylint: disable=W0212
-  tf.app.flags.FLAGS = tf.app.flags._FlagValues()
-  tf.app.flags._global_parser = argparse.ArgumentParser()
+    """Resets Tensorflow's FLAG values"""
+    #pylint: disable=W0212
+    for flag_key in dir(tf.app.flags.FLAGS):
+        delattr(tf.app.flags.FLAGS, flag_key)
+    #tf.app.flags.FLAGS = tf.app.flags._FlagValues()
+    tf.app.flags._global_parser = argparse.ArgumentParser()
 
 
 class PipelineTest(tf.test.TestCase):

seq2seq/training/utils.py

@@ -115,7 +115,7 @@ def cell_from_spec(cell_classname, cell_params):
   cell_class = locate(cell_classname) or getattr(rnn_cell, cell_classname)
 
   # Make sure additional arguments are valid
-  cell_args = set(inspect.getargspec(cell_class.__init__).args[1:])
+  cell_args = set(inspect.signature(cell_class.__init__).parameters)
   for key in cell_params.keys():
     if key not in cell_args:
       raise ValueError(