Use stable functions (#9369)

src/transformers/benchmark/benchmark_args_tf.py

@@ -96,15 +96,15 @@ def _setup_strategy(self) -> Tuple["tf.distribute.Strategy", "tf.distribute.clus
             tf.config.experimental_connect_to_cluster(self._setup_tpu)
             tf.tpu.experimental.initialize_tpu_system(self._setup_tpu)
 
-            strategy = tf.distribute.experimental.TPUStrategy(self._setup_tpu)
+            strategy = tf.distribute.TPUStrategy(self._setup_tpu)
         else:
             # currently no multi gpu is allowed
             if self.is_gpu:
                 # TODO: Currently only single GPU is supported
-                tf.config.experimental.set_visible_devices(self.gpu_list[self.device_idx], "GPU")
+                tf.config.set_visible_devices(self.gpu_list[self.device_idx], "GPU")
                 strategy = tf.distribute.OneDeviceStrategy(device=f"/gpu:{self.device_idx}")
             else:
-                tf.config.experimental.set_visible_devices([], "GPU")  # disable GPU
+                tf.config.set_visible_devices([], "GPU")  # disable GPU
                 strategy = tf.distribute.OneDeviceStrategy(device=f"/cpu:{self.device_idx}")
 
         return strategy

src/transformers/trainer_tf.py

@@ -27,7 +27,6 @@
 
 import numpy as np
 import tensorflow as tf
-from packaging.version import parse
 from tensorflow.python.distribute.values import PerReplica
 
 from .modeling_tf_utils import TFPreTrainedModel
@@ -93,11 +92,6 @@ def __init__(
             None,
         ),
     ):
-        assert parse(tf.__version__).release >= (2, 2, 0), (
-            "You need to run the TensorFlow trainer with at least the version 2.2.0, your version is %r "
-            % tf.__version__
-        )
-
         self.model = model
         self.args = args
         self.train_dataset = train_dataset
@@ -141,7 +135,7 @@ def get_train_tfdataset(self) -> tf.data.Dataset:
             raise ValueError("Trainer: training requires a train_dataset.")
 
         self.total_train_batch_size = self.args.train_batch_size * self.args.gradient_accumulation_steps
-        self.num_train_examples = tf.data.experimental.cardinality(self.train_dataset).numpy()
+        self.num_train_examples = self.train_dataset.cardinality(self.train_dataset).numpy()
 
         if self.num_train_examples < 0:
             raise ValueError("The training dataset must have an asserted cardinality")
@@ -173,7 +167,7 @@ def get_eval_tfdataset(self, eval_dataset: Optional[tf.data.Dataset] = None) ->
             raise ValueError("Trainer: evaluation requires an eval_dataset.")
 
         eval_dataset = eval_dataset if eval_dataset is not None else self.eval_dataset
-        num_examples = tf.data.experimental.cardinality(eval_dataset).numpy()
+        num_examples = eval_dataset.cardinality(eval_dataset).numpy()
 
         if num_examples < 0:
             raise ValueError("The training dataset must have an asserted cardinality")
@@ -203,7 +197,7 @@ def get_test_tfdataset(self, test_dataset: tf.data.Dataset) -> tf.data.Dataset:
         Subclass and override this method if you want to inject some custom behavior.
         """
 
-        num_examples = tf.data.experimental.cardinality(test_dataset).numpy()
+        num_examples = test_dataset.cardinality(test_dataset).numpy()
 
         if num_examples < 0:
             raise ValueError("The training dataset must have an asserted cardinality")

src/transformers/training_args_tf.py

@@ -188,7 +188,7 @@ def _setup_strategy(self) -> Tuple["tf.distribute.Strategy", int]:
                 tf.config.experimental_connect_to_cluster(tpu)
                 tf.tpu.experimental.initialize_tpu_system(tpu)
 
-                strategy = tf.distribute.experimental.TPUStrategy(tpu)
+                strategy = tf.distribute.TPUStrategy(tpu)
 
             elif len(gpus) == 0:
                 strategy = tf.distribute.OneDeviceStrategy(device="/cpu:0")

tests/test_modeling_tf_common.py

@@ -51,10 +51,10 @@
         for gpu in gpus:
             # Restrict TensorFlow to only allocate x GB of memory on the GPUs
             try:
-                tf.config.experimental.set_virtual_device_configuration(
-                    gpu, [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=_tf_gpu_memory_limit)]
+                tf.config.set_logical_device_configuration(
+                    gpu, [tf.config.LogicalDeviceConfiguration(memory_limit=_tf_gpu_memory_limit)]
                 )
-                logical_gpus = tf.config.experimental.list_logical_devices("GPU")
+                logical_gpus = tf.config.list_logical_devices("GPU")
                 print("Logical GPUs", logical_gpus)
             except RuntimeError as e:
                 # Virtual devices must be set before GPUs have been initialized

tests/test_optimization_tf.py

@@ -14,8 +14,6 @@
 
 import unittest
 
-from packaging import version
-
 from transformers import is_tf_available
 from transformers.testing_utils import require_tf
 
@@ -78,18 +76,12 @@ def accumulate(grad1, grad2):
                 local_variables = strategy.experimental_local_results(gradient_placeholder)
                 local_variables[0].assign(grad1)
                 local_variables[1].assign(grad2)
-                if version.parse(tf.version.VERSION) >= version.parse("2.2"):
-                    strategy.run(accumulate_on_replica, args=(gradient_placeholder,))
-                else:
-                    strategy.experimental_run_v2(accumulate_on_replica, args=(gradient_placeholder,))
+                strategy.run(accumulate_on_replica, args=(gradient_placeholder,))
 
         @tf.function
         def apply_grad():
             with strategy.scope():
-                if version.parse(tf.version.VERSION) >= version.parse("2.2"):
-                    strategy.run(apply_on_replica)
-                else:
-                    strategy.experimental_run_v2(apply_on_replica)
+                strategy.run(apply_on_replica)
 
         def _check_local_values(grad1, grad2):
             values = strategy.experimental_local_results(accumulator._gradients[0])