Keras 및 MultiWorkerMirroredStrategy를 사용한 사용자 정의 훈련 루프

import json
import os
import sys

os.environ["CUDA_VISIBLE_DEVICES"] = "-1"

os.environ.pop('TF_CONFIG', None)

if '.' not in sys.path:
  sys.path.insert(0, '.')

import tensorflow as tf

2022-12-15 02:05:55.039887: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory
2022-12-15 02:05:55.040002: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory
2022-12-15 02:05:55.040014: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.

%%writefile mnist.py

import os
import tensorflow as tf
import numpy as np

def mnist_dataset(batch_size):
  (x_train, y_train), _ = tf.keras.datasets.mnist.load_data()
  # The `x` arrays are in uint8 and have values in the range [0, 255].
  # You need to convert them to float32 with values in the range [0, 1]
  x_train = x_train / np.float32(255)
  y_train = y_train.astype(np.int64)
  train_dataset = tf.data.Dataset.from_tensor_slices(
      (x_train, y_train)).shuffle(60000)
  return train_dataset

def dataset_fn(global_batch_size, input_context):
  batch_size = input_context.get_per_replica_batch_size(global_batch_size)
  dataset = mnist_dataset(batch_size)
  dataset = dataset.shard(input_context.num_input_pipelines,
                          input_context.input_pipeline_id)
  dataset = dataset.batch(batch_size)
  return dataset

def build_cnn_model():
  return tf.keras.Sequential([
      tf.keras.Input(shape=(28, 28)),
      tf.keras.layers.Reshape(target_shape=(28, 28, 1)),
      tf.keras.layers.Conv2D(32, 3, activation='relu'),
      tf.keras.layers.Flatten(),
      tf.keras.layers.Dense(128, activation='relu'),
      tf.keras.layers.Dense(10)
  ])

Writing mnist.py

tf_config = {
    'cluster': {
        'worker': ['localhost:12345', 'localhost:23456']
    },
    'task': {'type': 'worker', 'index': 0}
}

json.dumps(tf_config)

'{"cluster": {"worker": ["localhost:12345", "localhost:23456"]}, "task": {"type": "worker", "index": 0} }'

os.environ['GREETINGS'] = 'Hello TensorFlow!'

echo ${GREETINGS}

Hello TensorFlow!

strategy = tf.distribute.MultiWorkerMirroredStrategy()

INFO:tensorflow:Single-worker MultiWorkerMirroredStrategy with local_devices = ('/device:CPU:0',), communication = CommunicationImplementation.AUTO
2022-12-15 02:05:56.272689: E tensorflow/compiler/xla/stream_executor/cuda/cuda_driver.cc:267] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected

import mnist
with strategy.scope():
  # Model building needs to be within `strategy.scope()`.
  multi_worker_model = mnist.build_cnn_model()

per_worker_batch_size = 64
num_workers = len(tf_config['cluster']['worker'])
global_batch_size = per_worker_batch_size * num_workers

with strategy.scope():
  multi_worker_dataset = strategy.distribute_datasets_from_function(
      lambda input_context: mnist.dataset_fn(global_batch_size, input_context))

with strategy.scope():
  # The creation of optimizer and train_accuracy needs to be in
  # `strategy.scope()` as well, since they create variables.
  optimizer = tf.keras.optimizers.RMSprop(learning_rate=0.001)
  train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
      name='train_accuracy')

@tf.function
def train_step(iterator):
  """Training step function."""

  def step_fn(inputs):
    """Per-Replica step function."""
    x, y = inputs
    with tf.GradientTape() as tape:
      predictions = multi_worker_model(x, training=True)
      per_batch_loss = tf.keras.losses.SparseCategoricalCrossentropy(
          from_logits=True,
          reduction=tf.keras.losses.Reduction.NONE)(y, predictions)
      loss = tf.nn.compute_average_loss(
          per_batch_loss, global_batch_size=global_batch_size)

    grads = tape.gradient(loss, multi_worker_model.trainable_variables)
    optimizer.apply_gradients(
        zip(grads, multi_worker_model.trainable_variables))
    train_accuracy.update_state(y, predictions)
    return loss

  per_replica_losses = strategy.run(step_fn, args=(next(iterator),))
  return strategy.reduce(
      tf.distribute.ReduceOp.SUM, per_replica_losses, axis=None)

from multiprocessing import util
checkpoint_dir = os.path.join(util.get_temp_dir(), 'ckpt')

def _is_chief(task_type, task_id, cluster_spec):
  return (task_type is None
          or task_type == 'chief'
          or (task_type == 'worker'
              and task_id == 0
              and "chief" not in cluster_spec.as_dict()))

def _get_temp_dir(dirpath, task_id):
  base_dirpath = 'workertemp_' + str(task_id)
  temp_dir = os.path.join(dirpath, base_dirpath)
  tf.io.gfile.makedirs(temp_dir)
  return temp_dir

def write_filepath(filepath, task_type, task_id, cluster_spec):
  dirpath = os.path.dirname(filepath)
  base = os.path.basename(filepath)
  if not _is_chief(task_type, task_id, cluster_spec):
    dirpath = _get_temp_dir(dirpath, task_id)
  return os.path.join(dirpath, base)

epoch = tf.Variable(
    initial_value=tf.constant(0, dtype=tf.dtypes.int64), name='epoch')
step_in_epoch = tf.Variable(
    initial_value=tf.constant(0, dtype=tf.dtypes.int64),
    name='step_in_epoch')
task_type, task_id = (strategy.cluster_resolver.task_type,
                      strategy.cluster_resolver.task_id)
# Normally, you don't need to manually instantiate a `ClusterSpec`, but in this 
# illustrative example you did not set `'TF_CONFIG'` before initializing the
# strategy. Check out the next section for "real-world" usage.
cluster_spec = tf.train.ClusterSpec(tf_config['cluster'])

checkpoint = tf.train.Checkpoint(
    model=multi_worker_model, epoch=epoch, step_in_epoch=step_in_epoch)

write_checkpoint_dir = write_filepath(checkpoint_dir, task_type, task_id,
                                      cluster_spec)
checkpoint_manager = tf.train.CheckpointManager(
    checkpoint, directory=write_checkpoint_dir, max_to_keep=1)

latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
if latest_checkpoint:
  checkpoint.restore(latest_checkpoint)

num_epochs = 3
num_steps_per_epoch = 70

while epoch.numpy() < num_epochs:
  iterator = iter(multi_worker_dataset)
  total_loss = 0.0
  num_batches = 0

  while step_in_epoch.numpy() < num_steps_per_epoch:
    total_loss += train_step(iterator)
    num_batches += 1
    step_in_epoch.assign_add(1)

  train_loss = total_loss / num_batches
  print('Epoch: %d, accuracy: %f, train_loss: %f.'
                %(epoch.numpy(), train_accuracy.result(), train_loss))

  train_accuracy.reset_states()

  # Once the `CheckpointManager` is set up, you're now ready to save, and remove
  # the checkpoints non-chief workers saved.
  checkpoint_manager.save()
  if not _is_chief(task_type, task_id, cluster_spec):
    tf.io.gfile.rmtree(write_checkpoint_dir)

  epoch.assign_add(1)
  step_in_epoch.assign(0)

2022-12-15 02:05:57.289580: W tensorflow/core/framework/dataset.cc:769] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.9/site-packages/tensorflow/python/autograph/pyct/static_analysis/liveness.py:83: Analyzer.lamba_check (from tensorflow.python.autograph.pyct.static_analysis.liveness) is deprecated and will be removed after 2023-09-23.
Instructions for updating:
Lambda fuctions will be no more assumed to be used in the statement where they are used, or at least in the same block. https://github.com/tensorflow/tensorflow/issues/56089
Epoch: 0, accuracy: 0.819531, train_loss: 0.572494.
Epoch: 1, accuracy: 0.929129, train_loss: 0.238400.
Epoch: 2, accuracy: 0.956250, train_loss: 0.157569.

Writing main.py

ls *.py

main.py
mnist.py

os.environ['TF_CONFIG'] = json.dumps(tf_config)

# first kill any previous runs
%killbgscripts

All background processes were killed.

python main.py &> job_0.log

import time
time.sleep(20)

cat job_0.log

2022-12-15 02:06:04.951967: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory
2022-12-15 02:06:04.952071: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory
2022-12-15 02:06:04.952083: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.
2022-12-15 02:06:06.050491: E tensorflow/compiler/xla/stream_executor/cuda/cuda_driver.cc:267] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected

tf_config['task']['index'] = 1
os.environ['TF_CONFIG'] = json.dumps(tf_config)

python main.py > /dev/null 2>&1

cat job_0.log

2022-12-15 02:06:04.951967: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory
2022-12-15 02:06:04.952071: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory
2022-12-15 02:06:04.952083: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.
2022-12-15 02:06:06.050491: E tensorflow/compiler/xla/stream_executor/cuda/cuda_driver.cc:267] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected
2022-12-15 02:06:27.412697: W tensorflow/core/framework/dataset.cc:769] Input of GeneratorDatasetOp::Dataset will not be optimized because the dataset does not implement the AsGraphDefInternal() method needed to apply optimizations.
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.9/site-packages/tensorflow/python/autograph/pyct/static_analysis/liveness.py:83: Analyzer.lamba_check (from tensorflow.python.autograph.pyct.static_analysis.liveness) is deprecated and will be removed after 2023-09-23.
Instructions for updating:
Lambda fuctions will be no more assumed to be used in the statement where they are used, or at least in the same block. https://github.com/tensorflow/tensorflow/issues/56089
Epoch: 0, accuracy: 0.823996, train_loss: 0.581225.
Epoch: 1, accuracy: 0.925558, train_loss: 0.249244.
Epoch: 2, accuracy: 0.946317, train_loss: 0.186680.

# Delete the `'TF_CONFIG'`, and kill any background tasks so they don't affect the next section.
os.environ.pop('TF_CONFIG', None)
%killbgscripts

All background processes were killed.