Issue 452: Added the assertion for consistency check and evaluation frequency check

algorithmic_efficiency/random_utils.py

@@ -18,30 +18,30 @@
 
 # Annoyingly, RandomState(seed) requires seed to be in [0, 2 ** 32 - 1] (an
 # unsigned int), while RandomState.randint only accepts and returns signed ints.
-MAX_INT32 = 2**31
-MIN_INT32 = -MAX_INT32
+MAX_UINT32 = 2**32-1
+MIN_UINT32 = 0
 
 SeedType = Union[int, list, np.ndarray]
 
 
 def _signed_to_unsigned(seed: SeedType) -> SeedType:
   if isinstance(seed, int):
-    return seed % 2**32
+    return seed % MAX_UINT32
   if isinstance(seed, list):
-    return [s % 2**32 for s in seed]
+    return [s % MAX_UINT32 for s in seed]
   if isinstance(seed, np.ndarray):
-    return np.array([s % 2**32 for s in seed.tolist()])
+    return np.array([s % MAX_UINT32 for s in seed.tolist()])
 
 
 def _fold_in(seed: SeedType, data: Any) -> List[Union[SeedType, Any]]:
   rng = np.random.RandomState(seed=_signed_to_unsigned(seed))
-  new_seed = rng.randint(MIN_INT32, MAX_INT32, dtype=np.int32)
+  new_seed = rng.randint(MIN_UINT32, MAX_UINT32, dtype=np.uint32)
   return [new_seed, data]
 
 
 def _split(seed: SeedType, num: int = 2) -> SeedType:
   rng = np.random.RandomState(seed=_signed_to_unsigned(seed))
-  return rng.randint(MIN_INT32, MAX_INT32, dtype=np.int32, size=[num, 2])
+  return rng.randint(MIN_UINT32, MAX_UINT32, dtype=np.uint32, size=[num, 2])
 
 
 def _PRNGKey(seed: SeedType) -> SeedType:  # pylint: disable=invalid-name
@@ -75,5 +75,5 @@ def split(seed: SeedType, num: int = 2) -> SeedType:
 def PRNGKey(seed: SeedType) -> SeedType:  # pylint: disable=invalid-name
   if FLAGS.framework == 'jax':
     _check_jax_install()
-    return jax_rng.PRNGKey(seed)
+    return jax_rng.key(seed)
   return _PRNGKey(seed)

submission_runner.py

@@ -12,6 +12,7 @@
     --num_tuning_trials=3 \
     --experiment_dir=/home/znado/experiment_dir \
     --experiment_name=baseline
+    --skip_eval = True/False
 """
 
 import datetime
@@ -88,6 +89,10 @@
 flags.DEFINE_string('librispeech_tokenizer_vocab_path',
                     '',
                     'Location to librispeech tokenizer.')
+flags.DEFINE_boolean(
+    'skip_eval',
+    True,
+    help='True to skip eval on the datasets and false otherwise')
 
 flags.DEFINE_enum(
     'framework',
@@ -327,7 +332,10 @@ def train_once(
   train_state['last_step_end_time'] = global_start_time
 
   logging.info('Starting training loop.')
-  goals_reached = (
+  if FLAGS.skip_eval == True:
+    goals_reached = (train_state['validation_goal_reached'])
+  else:
+    goals_reached = (
       train_state['validation_goal_reached'] and
       train_state['test_goal_reached'])
   while train_state['is_time_remaining'] and \
@@ -402,9 +410,12 @@ def train_once(
           train_state['test_goal_reached'] = (
               workload.has_reached_test_target(latest_eval_result) or
               train_state['test_goal_reached'])
-          goals_reached = (
-              train_state['validation_goal_reached'] and
-              train_state['test_goal_reached'])
+          if FLAGS.skip_eval == True:
+              goals_reached = (train_state['validation_goal_reached'])
+          else:
+              goals_reached = (
+                train_state['validation_goal_reached'] and
+                train_state['test_goal_reached'])
           # Save last eval time.
           eval_end_time = get_time()
           train_state['last_eval_time'] = eval_end_time
@@ -487,7 +498,11 @@ def train_once(
           preemption_count=preemption_count,
           checkpoint_dir=log_dir,
           save_intermediate_checkpoints=FLAGS.save_intermediate_checkpoints)
-
+  assert(abs(metrics['eval_results'][-1][1]['total_duration'] - 
+  (train_state['accumulated_submission_time'] + 
+  train_state['accumulated_logging_time'] + 
+  train_state['accumulated_eval_time']) <= 10))
+  assert(int(train_state['accumulated_submission_time'] // workload.eval_period_time_sec) <= len(metrics['eval_results']) + 2)
   return train_state['accumulated_submission_time'], metrics