Merged commit includes the following changes: (#7289)

259649972  by hongkuny<hongkuny@google.com>:

    Update docs.

--
259470074  by hongkuny<hongkuny@google.com>:

    Adds a dedup phase for trainable variables.

--

PiperOrigin-RevId: 259649972

official/bert/README.md

-# BERT in TensorFlow
+# BERT (Bidirectional Encoder Representations from Transformers)
 
 Note> Please do not create pull request. This model is still under development
 and testing.

official/bert/model_training_utils.py

@@ -242,7 +242,8 @@ def run_customized_training_loop(
           model_outputs = model(inputs)
           loss = loss_fn(labels, model_outputs)
 
-        tvars = model.trainable_variables
+        # De-dupes variables due to keras tracking issues.
+        tvars = list(set(model.trainable_variables))
         grads = tape.gradient(loss, tvars)
         optimizer.apply_gradients(zip(grads, tvars))
         # For reporting, the metric takes the mean of losses.