Registry refactor (tensorflow#1410)

* registry refactor and deprecated call-site updates

* added on_problem_set callback, simplified name

* changed optimizer registration names to snake_case, documentation

* removed create_registry

Author: jackd

tensor2tensor/bin/t2t_attack.py

@@ -73,7 +73,7 @@ def create_attack_params():
 
 
 def create_attack(attack):
-  return registry.attacks(attack)
+  return registry.attack(attack)
 
 
 def create_surrogate_hparams():

tensor2tensor/bin/t2t_datagen.py

@@ -147,7 +147,7 @@ def main(_):
 
   # Calculate the list of problems to generate.
   problems = sorted(
-      list(_SUPPORTED_PROBLEM_GENERATORS) + registry.list_problems())
+      list(_SUPPORTED_PROBLEM_GENERATORS) + registry.list_base_problems())
   for exclude in FLAGS.exclude_problems.split(","):
     if exclude:
       problems = [p for p in problems if exclude not in p]
@@ -169,7 +169,8 @@ def main(_):
 
   if not problems:
     problems_str = "\n  * ".join(
-        sorted(list(_SUPPORTED_PROBLEM_GENERATORS) + registry.list_problems()))
+        sorted(list(_SUPPORTED_PROBLEM_GENERATORS) +
+               registry.list_base_problems()))
     error_msg = ("You must specify one of the supported problems to "
                  "generate data for:\n  * " + problems_str + "\n")
     error_msg += ("TIMIT and parsing need data_sets specified with "

tensor2tensor/bin/t2t_prune.py

@@ -54,7 +54,7 @@ def create_pruning_params():
 
 
 def create_pruning_strategy(name):
-  return registry.pruning_strategies(name)
+  return registry.pruning_strategy(name)
 
 
 def main(argv):

tensor2tensor/layers/common_hparams.py

@@ -55,7 +55,7 @@ def basic_params1():
       initializer="orthogonal",
       initializer_gain=1.5,
       label_smoothing=0.1,
-      optimizer="Adam",
+      optimizer="adam",
       optimizer_adam_epsilon=1e-6,
       optimizer_adam_beta1=0.85,
       optimizer_adam_beta2=0.997,
@@ -466,7 +466,7 @@ def basic_range1(ranged_hparams):
   rhp.set_float("optimizer_adam_beta2", 0.995, 0.999)
   rhp.set_categorical(
       "optimizer",
-      ["Adam", "Adagrad", "Momentum", "RMSProp", "SGD", "YellowFin"])
+      ["adam", "adagrad", "momentum", "rms_prop", "sgd", "yellow_fin"])
 
 
 @registry.register_ranged_hparams

tensor2tensor/models/mtf_transformer2.py

@@ -269,7 +269,7 @@ def sample(self, features, mesh):
     return self.combine_batch_dims(ret)
 
 
-layers_registry = registry.create_registry("layers")
+layers_registry = registry.Registries.layers
 
 
 # The following functions construct layers based on hyperparmeters

tensor2tensor/models/research/adafactor_experiments.py

@@ -30,16 +30,16 @@ def mimic_adam_with_adafactor(hparams):
   Some minor things may be different, like epsilon and beta1 correction.
 
   Args:
-    hparams: model hyperparameters where "Adam" in hparams.optimizer
+    hparams: model hyperparameters where "adam" in hparams.optimizer
   """
-  assert "Adam" in hparams.optimizer
-  hparams.optimizer = "Adafactor"
+  assert "adam" in hparams.optimizer
+  hparams.optimizer = "adafactor"
   hparams.optimizer_adafactor_beta1 = hparams.optimizer_adam_beta1
   hparams.optimizer_adafactor_beta2 = hparams.optimizer_adam_beta2
   hparams.optimizer_adafactor_multiply_by_parameter_scale = False
   hparams.optimizer_adafactor_factored = False
   hparams.optimizer_adafactor_clipping_threshold = None
-  hparams.optimizer_adafactor_decay_type = "Adam"
+  hparams.optimizer_adafactor_decay_type = "adam"
 
 
 @registry.register_hparams
@@ -50,7 +50,7 @@ def afx_adam():
   hparams.optimizer_adam_beta2 = 0.999
   hparams.symbol_modality_num_shards = 1
   hparams.batch_size = 2048
-  hparams.optimizer = "Adam"
+  hparams.optimizer = "adam"
   hparams.learning_rate_schedule = (
       "constant*rsqrt_decay*linear_warmup*rsqrt_hidden_size")
   hparams.learning_rate_constant = 2.0

tensor2tensor/models/research/autoencoders.py

@@ -1020,7 +1020,7 @@ def body(self, features):
 def autoencoder_basic():
   """Basic autoencoder model."""
   hparams = common_hparams.basic_params1()
-  hparams.optimizer = "Adam"
+  hparams.optimizer = "adam"
   hparams.learning_rate_constant = 0.0002
   hparams.learning_rate_warmup_steps = 500
   hparams.learning_rate_schedule = "constant * linear_warmup"

tensor2tensor/models/research/transformer_nat.py

@@ -392,7 +392,7 @@ def transformer_nat_small():
   hparams.filter_size = 2048
   hparams.label_smoothing = 0.0
   hparams.force_full_predict = True
-  hparams.optimizer = "Adam"
+  hparams.optimizer = "adam"
   hparams.optimizer_adam_epsilon = 1e-9
   hparams.optimizer_adam_beta1 = 0.9
   hparams.optimizer_adam_beta2 = 0.997

tensor2tensor/models/research/transformer_vae.py

@@ -767,7 +767,7 @@ def transformer_ae_small():
   hparams.filter_size = 2048
   hparams.add_hparam("compress_filter_size", 2048 * 2)
   hparams.label_smoothing = 0.0
-  hparams.optimizer = "Adam"  # Can be unstable, maybe try Adam.
+  hparams.optimizer = "adam"  # Can be unstable, maybe try Adam.
   hparams.optimizer_adam_epsilon = 1e-9
   hparams.optimizer_adam_beta1 = 0.9
   hparams.optimizer_adam_beta2 = 0.997  # Needs tuning, try 0.98 to 0.999.
@@ -941,7 +941,7 @@ def transformer_ae_a3():
 def transformer_ae_a6():
   """Best hparams for transformer with semhash."""
   hparams = transformer_ae_a3()
-  hparams.optimizer = "Adam"
+  hparams.optimizer = "adam"
   hparams.noise_dev = 0.5
   return hparams
 

tensor2tensor/models/research/vqa_attention.py

@@ -335,7 +335,7 @@ def vqa_attention_base():
   hparams = common_hparams.basic_params1()
   hparams.batch_size = 128
   hparams.use_fixed_batch_size = True,
-  hparams.optimizer = "Adam"
+  hparams.optimizer = "adam"
   hparams.optimizer_adam_beta1 = 0.9
   hparams.optimizer_adam_beta2 = 0.999
   hparams.optimizer_adam_epsilon = 1e-8