Google :混合不同的tf.distribute.Strategy

Question

我使用Talos和Google TPU来运行Keras模型的超参数调优。请注意，我使用的是Tensorflow 2.0.0和Keras2.2.4-TF。

# pip install --upgrade tensorflow
# pip install --upgrade --force-reinstall tensorflow-gpu

import os
import tensorflow as tf
import talos as ta
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense

tf.compat.v1.disable_eager_execution()

def iris_model(x_train, y_train, x_val, y_val, params):

    # Specify a distributed strategy to use TPU
    resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='grpc://' + os.environ['COLAB_TPU_ADDR'])
    tf.config.experimental_connect_to_host(resolver.master())
    tf.tpu.experimental.initialize_tpu_system(resolver)
    strategy = tf.distribute.experimental.TPUStrategy(resolver)

    # Use the strategy to create and compile a Keras model
    with strategy.scope():
      model = Sequential()
      model.add(Dense(32, input_dim=4, activation=params['activation']))
      model.add(Dense(3, activation='softmax'))
      model.compile(optimizer=params['optimizer'], loss=params['losses'])

    # Convert the train set to a Dataset to use TPU
    dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
    dataset = dataset.cache().shuffle(1000, reshuffle_each_iteration=True).repeat().batch(params['batch_size'], drop_remainder=True)

    # Fit the Keras model on the dataset
    out = model.fit(dataset, 
                    batch_size=params['batch_size'], 
                    epochs=params['epochs'],
                    validation_data=[x_val, y_val],
                    verbose=0,
                    steps_per_epoch=4)

    return out, model

x, y = ta.templates.datasets.iris()

# Create a hyperparameter distributions 
p = {'activation': ['relu', 'elu'],
       'optimizer': ['Nadam', 'Adam'],
       'losses': ['logcosh'],
       'batch_size': (20, 50, 5),
       'epochs': [10, 20]}

# Use Talos to scan the best hyperparameters of the Keras model
scan_object = ta.Scan(x, y, model=iris_model, params=p, fraction_limit=0.1, experiment_name='first_test')

在使用tf.data.Dataset，将火车集转换为数据集后，当使用out = model.fit对模型进行拟合时，会出现以下错误

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/distribute/distribute_lib.py in _wrong_strategy_scope(strategy, context)
    218     raise RuntimeError(
    219         "Mixing different tf.distribute.Strategy objects: %s is not %s" %
--> 220         (context.strategy, strategy))
    221 
    222 

RuntimeError: Mixing different tf.distribute.Strategy objects: <tensorflow.python.distribute.tpu_strategy.TPUStrategy object at 0x7f9886506c50> is not <tensorflow.python.distribute.tpu_strategy.TPUStrategy object at 0x7f988aa04080>

Answer 1

TensorFlow 2.0.0发行版不支持TPU培训。不过，这在TensorFlow 2.2中不应该是一个问题。我对您的代码做了一些小的修正，并让它在Colab上运行：

• 使用Talos 1.0 (pip install git+https://github.com/autonomio/talos@1.0)
• 将tf.config.experimental_connect_to_host(resolver.master())替换为tf.config.experimental_connect_to_cluster(resolver)
• 使用tf.data.Dataset进行验证数据。

%tensorflow_version 2.x
import os
import tensorflow as tf
import talos as ta
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense

def iris_model(x_train, y_train, x_val, y_val, params):

    # Specify a distributed strategy to use TPU
    resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='grpc://' + os.environ['COLAB_TPU_ADDR'])
    tf.config.experimental_connect_to_cluster(resolver)
    tf.tpu.experimental.initialize_tpu_system(resolver)
    strategy = tf.distribute.experimental.TPUStrategy(resolver)

    # Use the strategy to create and compile a Keras model
    with strategy.scope():
      model = Sequential()
      model.add(Dense(32, input_dim=4, activation=params['activation']))
      model.add(Dense(3, activation='softmax'))
      model.compile(optimizer=params['optimizer'], loss=params['losses'])

    # Convert the train set to a Dataset to use TPU
    dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
    dataset = dataset.cache().shuffle(1000, reshuffle_each_iteration=True).repeat().batch(params['batch_size'], drop_remainder=True)
    val_dataset = tf.data.Dataset.from_tensor_slices((x_val, y_val)).batch(params['batch_size'], drop_remainder=True)

    # Fit the Keras model on the dataset
    out = model.fit(dataset, 
                    batch_size=params['batch_size'], 
                    epochs=params['epochs'],
                    validation_data=val_dataset,
                    verbose=0,
                    steps_per_epoch=4)

    return out, model