Internal change

PiperOrigin-RevId: 400408816

official/nlp/projects/teams/experiments/base/wiki_books_pretrain.yaml

 task:
   model:
-    cls_heads: [{activation: tanh, cls_token_idx: 0, dropout_rate: 0.1, inner_dim: 768,
-        name: next_sentence, num_classes: 2}]
-    generator_encoder:
-      bert:
+    candidate_size: 5
+    num_shared_generator_hidden_layers: 3
+    num_discriminator_task_agnostic_layers: 11
+    tie_embeddings: true
+    generator:
       attention_dropout_rate: 0.1
       dropout_rate: 0.1
       embedding_size: 768
       hidden_activation: gelu
-        hidden_size: 256
+      hidden_size: 768
       initializer_range: 0.02
-        intermediate_size: 1024
+      intermediate_size: 3072
       max_position_embeddings: 512
-        num_attention_heads: 4
-        num_layers: 12
+      num_attention_heads: 12
+      num_layers: 6
       type_vocab_size: 2
       vocab_size: 30522
-    num_masked_tokens: 76
-    sequence_length: 512
-    num_classes: 2
-    discriminator_encoder:
-      bert:
+    discriminator:
       attention_dropout_rate: 0.1
       dropout_rate: 0.1
       embedding_size: 768
@@ -30,12 +27,9 @@ task:
       intermediate_size: 3072
       max_position_embeddings: 512
       num_attention_heads: 12
-        num_layers: 12
+      num_layers: 6
       type_vocab_size: 2
       vocab_size: 30522
-    discriminator_loss_weight: 50.0
-    disallow_correct: false
-    tie_embeddings: true
   train_data:
     drop_remainder: true
     global_batch_size: 256
@@ -55,8 +49,8 @@ task:
     use_next_sentence_label: false
     use_position_id: false
 trainer:
-  checkpoint_interval: 6000
-  max_to_keep: 50
+  checkpoint_interval: 4000
+  max_to_keep: 5
   optimizer_config:
     learning_rate:
       polynomial:
@@ -73,8 +67,8 @@ trainer:
         power: 1
         warmup_steps: 10000
       type: polynomial
-  steps_per_loop: 1000
-  summary_interval: 1000
+  steps_per_loop: 4000
+  summary_interval: 4000
   train_steps: 1000000
   validation_interval: 100
   validation_steps: 64

official/nlp/projects/teams/teams_pretrainer.py

@@ -21,6 +21,8 @@ from official.modeling import tf_utils
 from official.nlp.modeling import layers
 from official.nlp.modeling import models
 
+_LOGIT_PENALTY_MULTIPLIER = 10000
+
 
 class ReplacedTokenDetectionHead(tf.keras.layers.Layer):
   """Replaced token detection discriminator head.
@@ -273,10 +275,9 @@ class TeamsPretrainer(tf.keras.Model):
     self.mlm_activation = mlm_activation
     self.mlm_initializer = mlm_initializer
     self.output_type = output_type
-    self.embedding_table = (
-        self.discriminator_mws_network.embedding_network.get_embedding_table())
     self.masked_lm = layers.MaskedLM(
-        embedding_table=self.embedding_table,
+        embedding_table=self.generator_network.embedding_network
+        .get_embedding_table(),
         activation=mlm_activation,
         initializer=mlm_initializer,
         output=output_type,
@@ -290,7 +291,8 @@ class TeamsPretrainer(tf.keras.Model):
         name='discriminator_rtd')
     hidden_cfg = discriminator_cfg['hidden_cfg']
     self.discriminator_mws_head = MultiWordSelectionHead(
-        embedding_table=self.embedding_table,
+        embedding_table=self.discriminator_mws_network.embedding_network
+        .get_embedding_table(),
         activation=hidden_cfg['intermediate_activation'],
         initializer=hidden_cfg['kernel_initializer'],
         output=output_type,
@@ -436,7 +438,7 @@ def sample_k_from_softmax(logits, k, disallow=None, use_topk=False):
   """
   if use_topk:
     if disallow is not None:
-      logits -= 10000.0 * disallow
+      logits -= _LOGIT_PENALTY_MULTIPLIER * disallow
     uniform_noise = tf.random.uniform(
         tf_utils.get_shape_list(logits), minval=0, maxval=1)
     gumbel_noise = -tf.math.log(-tf.math.log(uniform_noise + 1e-9) + 1e-9)
@@ -445,7 +447,7 @@ def sample_k_from_softmax(logits, k, disallow=None, use_topk=False):
     sampled_tokens_list = []
     vocab_size = tf_utils.get_shape_list(logits)[-1]
     if disallow is not None:
-      logits -= 10000.0 * disallow
+      logits -= _LOGIT_PENALTY_MULTIPLIER * disallow
 
     uniform_noise = tf.random.uniform(
         tf_utils.get_shape_list(logits), minval=0, maxval=1)
@@ -454,7 +456,7 @@ def sample_k_from_softmax(logits, k, disallow=None, use_topk=False):
     for _ in range(k):
       token_ids = tf.argmax(logits, -1, output_type=tf.int32)
       sampled_tokens_list.append(token_ids)
-      logits -= 10000.0 *  tf.one_hot(
+      logits -= _LOGIT_PENALTY_MULTIPLIER *  tf.one_hot(
           token_ids, depth=vocab_size, dtype=tf.float32)
     sampled_tokens = tf.stack(sampled_tokens_list, -1)
   return sampled_tokens