From 61223394892db3ca63768c4e39eb2caefcc02b65 Mon Sep 17 00:00:00 2001
From: Jae Ro <jaero@google.com>
Date: Thu, 28 Oct 2021 10:44:25 -0700
Subject: [PATCH] Delay importing tensorflow instead of directly depending on
 it in setup

This will require users to manually install tensorflow if they happen to be using a part of fedjax that depends on tensorflow (a relatively small subset of the api).

PiperOrigin-RevId: 406173096
---
 fedjax/core/client_datasets_benchmark.py     |  3 +-
 fedjax/core/serialization.py                 |  5 +--
 fedjax/core/sqlite_federated_data_test.py    |  3 +-
 fedjax/core/util.py                          | 32 ++++++++++++++++++++
 fedjax/datasets/cifar100.py                  |  3 +-
 fedjax/datasets/stackoverflow.py             |  5 ++-
 fedjax/training/checkpoint.py                |  4 ++-
 fedjax/training/federated_experiment.py      |  4 ++-
 fedjax/training/federated_experiment_test.py | 10 +++---
 fedjax/training/logging.py                   |  4 +--
 fedjax/training/logging_test.py              |  9 +++---
 setup.py                                     |  1 -
 12 files changed, 63 insertions(+), 20 deletions(-)

diff --git a/fedjax/core/client_datasets_benchmark.py b/fedjax/core/client_datasets_benchmark.py
index 48ff0ea..e6d0d17 100644
--- a/fedjax/core/client_datasets_benchmark.py
+++ b/fedjax/core/client_datasets_benchmark.py
@@ -28,8 +28,8 @@
 
 from absl import app
 from fedjax.core import client_datasets
+from fedjax.core import util
 import numpy as np
-import tensorflow as tf
 
 # pylint: disable=cell-var-from-loop
 
@@ -78,6 +78,7 @@ def bench_client_dataset(preprocess, mode, batch_size=128, num_steps=100):
 
 def bench_tf_dataset(preprocess, mode, batch_size=128, num_steps=100):
   """Benchmarks TF Dataset."""
+  tf = util.import_tf()
   shuffle_buffer = 1000  # size of FAKE_MNIST
   dataset = tf.data.Dataset.from_tensor_slices(FAKE_MNIST)
   if mode == 'train':
diff --git a/fedjax/core/serialization.py b/fedjax/core/serialization.py
index ef41cb4..6848307 100644
--- a/fedjax/core/serialization.py
+++ b/fedjax/core/serialization.py
@@ -38,15 +38,15 @@
 import pickle
 
 from absl import logging
+from fedjax.core import util
 import jax
 import msgpack
 import numpy as np
-# TODO(b/188556866): Remove dependency on TensorFlow.
-import tensorflow as tf
 
 
 def save_state(state, path):
   """Saves state to file path."""
+  tf = util.import_tf()
   logging.info('Saving state to %s.', path)
   with tf.io.gfile.GFile(path, 'wb') as f:
     pickle.dump(state, f)
@@ -54,6 +54,7 @@ def save_state(state, path):
 
 def load_state(path):
   """Loads saved state from file path."""
+  tf = util.import_tf()
   logging.info('Loading params from %s.', path)
   with tf.io.gfile.GFile(path, 'rb') as f:
     return pickle.load(f)
diff --git a/fedjax/core/sqlite_federated_data_test.py b/fedjax/core/sqlite_federated_data_test.py
index 17310a8..7159743 100644
--- a/fedjax/core/sqlite_federated_data_test.py
+++ b/fedjax/core/sqlite_federated_data_test.py
@@ -21,10 +21,10 @@
 from absl.testing import absltest
 from fedjax.core import serialization
 from fedjax.core import sqlite_federated_data
+from fedjax.core import util
 import numpy as np
 import numpy.testing as npt
 import sqlite3
-import tensorflow as tf
 
 FLAGS = flags.FLAGS
 
@@ -251,6 +251,7 @@ class TFFSQLiteClientsIteratorTest(absltest.TestCase):
   @classmethod
   def setUpClass(cls):
     super().setUpClass()
+    tf = util.import_tf()
     path = os.path.join(FLAGS.test_tmpdir, 'test_tff_sqlite_reader.sqlite')
     split_name = 'train'
 
diff --git a/fedjax/core/util.py b/fedjax/core/util.py
index 5aa10b2..3a559e6 100644
--- a/fedjax/core/util.py
+++ b/fedjax/core/util.py
@@ -22,3 +22,35 @@ def safe_div(a, b):
   safe = b != 0
   c = a / jnp.where(safe, b, 1)
   return jnp.where(safe, c, 0)
+
+
+# TODO(b/188556866): Remove dependency on TensorFlow.
+def import_tf():
+  """Imports and returns TensorFlow module if it is installed.
+
+  This is to avoid having FedJAX directly depend on TensorFlow since TensorFlow
+  is large and complex and the majority of FedJAX functions without it.
+
+  Returns:
+    TensorFlow module if it is installed.
+
+  Raises:
+    ModuleNotFoundError: If TensorFlow is not installed along with instructions
+      on how to install TensorFlow.
+  """
+  try:
+    import tensorflow  # pylint:disable=g-import-not-at-top
+    return tensorflow
+  except ModuleNotFoundError as e:
+    raise ModuleNotFoundError(
+        """This FedJAX feature requires TensorFlow, but TensorFlow is not installed.
+
+If you do not otherwise need TensorFlow, we recommend installing the smaller
+CPU-only version of TensorFlow via
+
+  pip install tensorflow-cpu
+
+If you may need to use TensorFlow with GPU, please install the full version via
+
+  pip install tensorflow
+""") from e
diff --git a/fedjax/datasets/cifar100.py b/fedjax/datasets/cifar100.py
index bc1fd75..62f2454 100644
--- a/fedjax/datasets/cifar100.py
+++ b/fedjax/datasets/cifar100.py
@@ -19,10 +19,10 @@
 from fedjax.core import client_datasets
 from fedjax.core import federated_data
 from fedjax.core import sqlite_federated_data
+from fedjax.core import util
 from fedjax.datasets import downloads
 
 import numpy as np
-import tensorflow as tf
 
 SPLITS = ('train', 'test')
 _TFF_SQLITE_COMPRESSED_HEXDIGEST = '23d3916c9caa33395737ee560cc7cb77bbd05fc7b73647ee7be3a7e764172939'
@@ -45,6 +45,7 @@ def cite():
 
 
 def _parse_tf_examples(vs: List[bytes]) -> client_datasets.Examples:
+  tf = util.import_tf()
   tf_examples = tf.io.parse_example(
       vs,
       features={
diff --git a/fedjax/datasets/stackoverflow.py b/fedjax/datasets/stackoverflow.py
index 57321e4..70e597b 100644
--- a/fedjax/datasets/stackoverflow.py
+++ b/fedjax/datasets/stackoverflow.py
@@ -19,9 +19,9 @@
 from fedjax.core import client_datasets
 from fedjax.core import federated_data
 from fedjax.core import sqlite_federated_data
+from fedjax.core import util
 from fedjax.datasets import downloads
 import numpy as np
-import tensorflow as tf
 
 SPLITS = ('train', 'held_out', 'test')
 
@@ -137,6 +137,7 @@ def preprocess_client(
 
 def default_vocab(default_vocab_size) -> List[str]:
   """Loads the deafult stackoverflow vocabulary."""
+  tf = util.import_tf()
   path = 'gs://gresearch/fedjax/stackoverflow/stackoverflow.word_count'
   vocab = []
   with tf.io.gfile.GFile(path) as f:
@@ -176,6 +177,7 @@ def __init__(self,
         OOV labels starting at `default_vocab_size + 3`.
       num_oov_buckets: Number of out of vocabulary buckets.
     """
+    tf = util.import_tf()
     if vocab is None:
       # Load default vocabulary.
       vocab = default_vocab(default_vocab_size)
@@ -197,6 +199,7 @@ def as_preprocess_batch(
     Returns:
       A function that can be used with FederatedData.preprocess_batch().
     """
+    tf = util.import_tf()
 
     @tf.function(input_signature=[tf.TensorSpec(shape=[None], dtype=tf.string)])
     def token_to_ids(tokens):
diff --git a/fedjax/training/checkpoint.py b/fedjax/training/checkpoint.py
index 443cc9f..d35745f 100644
--- a/fedjax/training/checkpoint.py
+++ b/fedjax/training/checkpoint.py
@@ -18,13 +18,14 @@
 from typing import Any, List, Optional, Tuple
 
 from fedjax.core import serialization
-import tensorflow as tf
+from fedjax.core import util
 
 _CHECKPOINT_PREFIX = 'checkpoint_'
 
 
 def _get_checkpoint_paths(base_path: str) -> List[str]:
   """Returns all checkpoint paths present."""
+  tf = util.import_tf()
   pattern = base_path + r'[0-9]{8}$'
   checkpoint_paths = []
   for path in tf.io.gfile.glob(base_path + '*'):
@@ -53,6 +54,7 @@ def save_checkpoint(root_dir: str,
                     round_num: int = 0,
                     keep: int = 1):
   """Saves checkpoint and cleans up old checkpoints."""
+  tf = util.import_tf()
   base_path = os.path.join(root_dir, _CHECKPOINT_PREFIX)
   checkpoint_path = f'{base_path}{round_num:08d}'
   serialization.save_state(state, checkpoint_path)
diff --git a/fedjax/training/federated_experiment.py b/fedjax/training/federated_experiment.py
index c091b50..28752b1 100644
--- a/fedjax/training/federated_experiment.py
+++ b/fedjax/training/federated_experiment.py
@@ -24,10 +24,10 @@
 from fedjax.core import federated_algorithm
 from fedjax.core import federated_data
 from fedjax.core import models
+from fedjax.core import util
 from fedjax.training import checkpoint
 from fedjax.training import logging as fedjax_logging
 import jax.numpy as jnp
-import tensorflow as tf
 
 
 def set_tf_cpu_only():
@@ -36,6 +36,7 @@ def set_tf_cpu_only():
   TensorFlow is only used for data loading, so we prevent it from allocating
   GPU/TPU memory.
   """
+  tf = util.import_tf()
   tf.config.experimental.set_visible_devices([], 'GPU')
   tf.config.experimental.set_visible_devices([], 'TPU')
 
@@ -177,6 +178,7 @@ def run_federated_experiment(
   Returns:
     Final state of the input federated algortihm after training.
   """
+  tf = util.import_tf()
   if config.root_dir:
     tf.io.gfile.makedirs(config.root_dir)
 
diff --git a/fedjax/training/federated_experiment_test.py b/fedjax/training/federated_experiment_test.py
index 79e1392..3c67fa4 100644
--- a/fedjax/training/federated_experiment_test.py
+++ b/fedjax/training/federated_experiment_test.py
@@ -13,6 +13,7 @@
 # limitations under the License.
 """Tests for fedjax.legacy.training.federated_experiment."""
 
+import glob
 import os.path
 
 from absl.testing import absltest
@@ -26,7 +27,6 @@
 import jax
 import numpy as np
 import numpy.testing as npt
-import tensorflow as tf
 
 
 class FakeClientSampler(client_samplers.ClientSampler):
@@ -123,7 +123,7 @@ def test_checkpoint(self):
           config=config)
       self.assertEqual(state, (6, 14))
       self.assertCountEqual(
-          tf.io.gfile.glob(os.path.join(config.root_dir, 'checkpoint_*')),
+          glob.glob(os.path.join(config.root_dir, 'checkpoint_*')),
           [os.path.join(config.root_dir, 'checkpoint_00000003')])
 
     with self.subTest('checkpoint restore'):
@@ -135,7 +135,7 @@ def test_checkpoint(self):
           config=config)
       self.assertEqual(state, (6, 16))
       self.assertCountEqual(
-          tf.io.gfile.glob(os.path.join(config.root_dir, 'checkpoint_*')),
+          glob.glob(os.path.join(config.root_dir, 'checkpoint_*')),
           [os.path.join(config.root_dir, 'checkpoint_00000004')])
 
   def test_periodic_eval_fn_map(self):
@@ -159,7 +159,7 @@ def test_periodic_eval_fn_map(self):
         })
     self.assertEqual(state, (5, 15))
     self.assertCountEqual(
-        tf.io.gfile.glob(os.path.join(config.root_dir, '*eval*')), [
+        glob.glob(os.path.join(config.root_dir, '*eval*')), [
             os.path.join(config.root_dir, 'test_eval'),
             os.path.join(config.root_dir, 'train_eval')
         ])
@@ -180,7 +180,7 @@ def test_final_eval_fn_map(self):
         })
     self.assertEqual(state, (5, 15))
     self.assertCountEqual(
-        tf.io.gfile.glob(os.path.join(config.root_dir, 'final_eval.tsv')),
+        glob.glob(os.path.join(config.root_dir, 'final_eval.tsv')),
         [os.path.join(config.root_dir, 'final_eval.tsv')])
 
 
diff --git a/fedjax/training/logging.py b/fedjax/training/logging.py
index 73b8c6e..65f7978 100644
--- a/fedjax/training/logging.py
+++ b/fedjax/training/logging.py
@@ -18,8 +18,7 @@
 from typing import Any, Optional
 
 from absl import logging
-
-import tensorflow as tf
+from fedjax.core import util
 
 
 class Logger:
@@ -44,6 +43,7 @@ def log(self, writer_name: str, metric_name: str, metric_value: Any,
       metric_value: Value of metric to log.
       round_num: Round number to log.
     """
+    tf = util.import_tf()
     logging.info('round %d %s: %s = %s', round_num, writer_name, metric_name,
                  metric_value)
 
diff --git a/fedjax/training/logging_test.py b/fedjax/training/logging_test.py
index 83f923d..19b12ae 100644
--- a/fedjax/training/logging_test.py
+++ b/fedjax/training/logging_test.py
@@ -13,12 +13,13 @@
 # limitations under the License.
 """Tests for fedjax.training.logging."""
 
-import tensorflow as tf
+import os
 
+from absl.testing import absltest
 from fedjax.training import logging
 
 
-class LoggingTest(tf.test.TestCase):
+class LoggingTest(absltest.TestCase):
 
   def test_log_no_root_dir(self):
     logger = logging.Logger()
@@ -37,8 +38,8 @@ def test_log_root_dir(self):
     logger.log(
         writer_name='eval', metric_name='loss', metric_value=5.3, round_num=0)
 
-    self.assertCountEqual(['train', 'eval'], tf.io.gfile.listdir(root_dir))
+    self.assertCountEqual(['train', 'eval'], os.listdir(root_dir))
 
 
 if __name__ == '__main__':
-  tf.test.main()
+  absltest.main()
diff --git a/setup.py b/setup.py
index 74c59c8..ead7315 100644
--- a/setup.py
+++ b/setup.py
@@ -41,7 +41,6 @@
         'msgpack',
         'optax',
         'requests',
-        'tensorflow',
     ],
     python_requires='>=3.6',
     classifiers=[