Quick Start

This section explains how to get started using Tarantella to distributedly train an existing TensorFlow 2/Keras model. First, we will examine what changes have to be made to your code, before we will look into the execution of your script with tarantella on the command line. Finally, we will present the features Tarantella currently supports and what important points need to be taken into account when using Tarantella.

Code example: LeNet-5 on MNIST

After having build and installed Tarantella we are ready to add distributed training support to an existing TensorFlow 2/Keras model. We will first illustrate all the necessary steps, using the well-known example of LeNet-5 on the MNIST dataset. Although this is not necessarily a good use case to take full advantage of Tarantella’s capabilities, it will allow you to simply copy-paste the code snippets and try them out, even on your laptop.

Let’s get started!

import tensorflow as tf
from tensorflow import keras
import tarantella as tnt

# Skip function implementations for brevity
[...]

# Initialize Tarantella (before doing anything else)
tnt.init()
args = parse_args()
              
# Create Tarantella model
model = tnt.Model(lenet5_model_generator())

# Compile Tarantella model (as with Keras)
model.compile(optimizer = keras.optimizers.SGD(learning_rate=args.learning_rate),
              loss = keras.losses.SparseCategoricalCrossentropy(),
              metrics = [keras.metrics.SparseCategoricalAccuracy()])

# Load MNIST dataset (as with Keras)
shuffle_seed = 42
(x_train, y_train), (x_val, y_val), (x_test, y_test) = \
      mnist_as_np_arrays(args.train_size, args.val_size, args.test_size)

train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_dataset = train_dataset.shuffle(len(x_train), shuffle_seed)
train_dataset = train_dataset.batch(args.batch_size)
train_dataset = train_dataset.prefetch(tf.data.experimental.AUTOTUNE)

test_dataset = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_dataset = test_dataset.batch(args.batch_size)

# Train Tarantella model (as with Keras)
model.fit(train_dataset,
          epochs = args.number_epochs,
          verbose = 1)

# Evaluate Tarantella model (as with Keras)
model.evaluate(test_dataset, verbose = 1)

As you can see from the marked lines in the code snippet, you only need to add 3 lines of code to train LeNet-5 distributedly using Tarantella! Let us go through the code in some more detail, in order to understand what is going on.

First we need to import the Tarantella library:

import tarantella as tnt

Having done that we need to initialize the library (which will setup the communication infrastructure):

tnt.init()

Note that this should be done before executing any other code. Next, we need to wrap the keras.Model object, generated by lenet5_model_generator(), into a tnt.Model object:

model = tnt.Model(lenet5_model_generator())

That’s it!

All the necessary steps to distribute training and datasets will now automatically be handled by Tarantella. In particular, we still run model.compile on the new model to generate a compute graph, just as we would have done with a typical Keras model.

Next, we load the MNIST data for training and testing, and create Dataset s from it. Note that we batch the dataset for training. This will guarantee that Tarantella is able to distribute the data later on in the correct way. Also note that the batch_size used here, is the same as for the original model, that is the global batch size. For details concerning local and global batch sizes have a look here.

Now we are able to train our model using model.fit, in the same familiar way used by the standard Keras interface. Note, however, that Tarantella is taking care of proper distribution of the train_dataset in the background. All the possibilities of how to feed datasets to Tarantella are explained in more detail below. Lastly, we can evaluate the final accuracy of our model on the test_dataset using model.evaluate.

To test and run tarantella in the next section, you can find a full version of the above example here.

Executing your model with tarantella

Next, let’s execute our model distributedly using tarantella on the command line. The simplest way to do that is by passing the Python script of the model to tarantella:

tarantella -- model.py

This will execute our model distributedly on a single node, using all the available GPUs. In case no GPUs can be found, tarantella will executed in serial mode on the CPU, and an WARNING message will be issued. In case you have GPUs available, but want to execute tarantella on CPUs nonetheless, you can specify the --no-gpu option.

tarantella --no-gpu -- model.py

We can also set command line parameters for the python script model.py, which have to succeed the name of the script:

tarantella --no-gpu -- model.py --batch_size=64 --learning_rate=0.01

On a single node, we can also explicitly specify the number of TensorFlow instances we want to use. This is done with the -n option:

tarantella -n 4 -- model.py --batch_size=64

Here, tarantella would try to execute distributedly on 4 GPUs. If there are not enough GPUs available, tarantella will print a WARNING and run 4 instances of TensorFlow on the CPU instead. If there are no GPUs installed or the --no-gpu option is use, tarantella will not print a WARNING.

Next, let’s run tarantella on multiple nodes. In order to do this, we need to provide tarantella with a hostfile that contains the hostname s of the nodes that we want to use:

$ cat hostfile
name_of_node_1
name_of_node_2

With this hostfile we can run tarantella on multiple nodes:

tarantella --hostfile hostfile -- model.py

In this case, tarantella uses all GPUs it can find. If no GPUs are available, tarantella will start one TensorFlow instance per node on the CPUs, and will issue an WARNING message. Again, this can be disabled by explicitly using the --no-gpu option.

As before, you can specify the number of GPUs/CPUs used per node explicitly with the option --n-per-node=<number>:

tarantella --hostfile hostfile --n-per-node=4 --no-gpu -- model.py --batch_size=64

In this example, tarantella would execute 4 instances of TensorFlow on the CPUs of each node specified in hostfile.

Caution

tarantella requires all the names in the hostfile be unique, and all nodes be homogeneous (number and type of CPUs and GPUs).

In addition, tarantella can be run with different levels of logging output. The log-levels that are available are INFO, WARNING, DEBUG and ERROR, and can be set with --log-level:

tarantella --hostfile hostfile --log-level=INFO -- model.py

By default, tarantella will log on the master rank only. This can be changed by using the --log-on-all-devices option which will print log messages for each rank individually.

Similarly, by default tarantella will print outputs from functions like fit, evaluate and predict, as well as callbacks only on the master rank. Sometimes, it might be useful to print outputs from all devices (e.g., for debugging), which can be switched on with the --output-on-all-devices option.

tarantella uses GPI-2’s gaspi_run internally, taking care of export ing environment variables, and generating an execution script from the user inputs. Details of this process can be monitored using the --dry-run option.

Lastly, you can overwrite the Tensor Fusion threshold tarantella uses with --fusion-threshold FUSION_THRESHOLD_KB (cf. here and here), and set and number of environment variables, most notably TNT_TENSORBOARD_ON_ALL_DEVICES, as explained here.

Save and load Tarantella models

Storing and loading your trained Tarantella.Model is very simple.

Tarantella supports all the different ways, in which you can load and store a keras.Model (for a guide look for instance here). In particular, you can:

• save the whole model (including the architecture, the weights and the state of the optimizer)

• save the model’s architecture/configuration only

• save the model’s weights only

Whole-model saving and loading

Saving the entire model including the architecture, weights and optimizer can be done via

model = ...  # get `tnt.Model`
model.save('path/to/location')

Alternatively, you could use tnt.models.save_model('path/to/location'), which works on both keras.Model s and tnt.Model s.

You can than load your model back using

import tarantella as tnt
model = tnt.models.load_model('path/to/location')

which will return an instance of tnt.Model.

Caution

At the moment, you will need to re-compile your model after loading.

This is again done with

model.compile(optimizer = keras.optimizers.SGD(learning_rate=args.learning_rate),
              loss = keras.losses.SparseCategoricalCrossentropy(),
              metrics = [keras.metrics.SparseCategoricalAccuracy()])

or similar.

Architecture saving and loading

If you only want to save the configuration (that is the architecture) of your model (in memory), you can use one of the following functions:

• tnt.Model.get_config

• tnt.Model.to_json

• tnt.Model.to_yaml

The architecture without its original weights and optimizer can then be restored using:

• tnt.models.model_from_config / tnt.Model.from_config

• tnt.models.model_from_json

• tnt.models.model_from_yaml

respectively. Here is an example:

import tarantella as tnt
model = ...  # get `tnt.Model`
config = model.get_config()
new_model = tnt.models.model_from_config(config)

The same can be achieved through cloning:

import tarantella as tnt
model = ...  # get `tnt.Model`
new_model = tnt.models.clone_model(model)

Weights saving and loading

Storing and loading the weights of a model to/from memory can be done using the functions tnt.Model.get_weights and tnt.Model.set_weights, respectively. Saving and loading weights to/from disk is done using the functions tnt.Model.save_weights and tnt.Model.load_weights, respectively.

Here is an example how this can be used to restore a model:

import tarantella as tnt
model = ...  # get `tnt.Model`
config = model.get_config()
weights = model.get_weights()

# initialize a new model with original model's weights
new_model = tnt.models.model_from_config(config)
new_model.set_weights(weights)

Checkpointing via callbacks

Apart from saving and loading models manually, Tarantella also supports checkpointing via Keras’ ModelCheckpoint callback, as it is described for instance here.

import tensorflow as tf
import tarantella as tnt

model = ...  # get `tnt.Model`

checkpoint_path = 'path/to/checkpoint/location'
model_checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(
  filepath=checkpoint_path, monitor='val_acc', verbose=1, save_best_only=False,
  save_weights_only=False, mode='auto', save_freq='epoch', options=None)

model.fit(train_dataset,
          validation_data = val_dataset,
          epochs = 2,
          callbacks = [model_checkpoint_callback])

Note

All saving to the filesystem (including tnt.Model.save and tnt.Model.save_weights) by Tarantella will only be done on the master rank.

This is the default and will yield correct behavior when you are using a distributed filesystem. If you wish to explicitly save on all devices you can pass tnt_save_all_devices = True to tnt.Model.save, tnt.Model.save_weights and tnt.models.save_model.

Using distributed datasets

This section explains what needs to be done in order to use Tarantella’s distributed datasets correctly.

The recommended way in which to provide your dataset to Tarantella is by passing a batched tf.data.Dataset to tnt.Model.fit. In order to do this, create a Dataset and apply the batch transformation using the (global) batch size to it. However, do not provide a value to batch_size in tnt.Model.fit, which would lead to double batching, and thus modified shapes for the input data.

Tarantella also supports batched and unbatched Dataset s in tnt.Model.fit when setting the tnt_micro_batch_size argument. This can be useful to obtain maximal performance in multi-node execution, as explained here. Keep in mind however, that Tarantella still expects the Dataset to be batched with the global batch size, and that the micro-batch size has to be consistent with the global batch size. 1 This is why, it is recommended to use an unbatched Dataset when setting a tnt_micro_batch_size explicitly.

Tarantella does not support any other way to feed data to fit at the moment. In particular, Numpy arrays, TensorFlow tensors and generators are not supported.

Tarantella’s automatic data distribution can be switched off by passing tnt_distribute_dataset=False in tnt.Model.fit, in which case Tarantella will issue an INFO message. If a validation dataset is passed to tnt.Model.fit, it should also be batched with the global batch size. You can similarly switch off its automatic micro-batching mechanism by setting tnt_distribute_validation_dataset=False.

There are a few important points when using distributed datasets in Tarantella:

Note

Batch size must be a multiple of the number of devices used.

This issue will be fixed in the next release.

Note

The last incomplete batch is always dropped.

We recommend to use drop_remainder=True when generating a Dataset. If drop_remainder is set to False, Tarantella will ignore it and issue a WARNING message. This behavior will be fixed in the next release.

Note

When using shuffle without a seed, Tarantella will use a fixed default seed.

This guarantees that the input data is shuffled the same way on all devices, when no seed is given, which is necessary for consistency. However, when a random seed is provided by the user, Tarantella will use that one instead.

Callbacks

At the moment, Tarantella fully supports 3 of the Keras callbacks:

• tf.keras.callbacks.LearningRateScheduler

• tf.keras.callbacks.ModelCheckpoint

• tf.keras.callbacks.TensorBoard

The LearningRateScheduler takes a schedule which will change the learning rate on each of the devices used (for detailed explanation, cf. here and here ). If verbose=1 is set, Tarantella will only print on one device by default. This behavior can be changed by passing --output-on-all-devices to tarantella.

ModelCheckpoint can be used to automatically checkpoint the state of the model during training. For an example look here, and into the Keras documentation.

The TensorBoard callback can be used to collect training information for visualization in TensorBoard. By default, Tarantella will only collect (device local) information on one device. If you want to collect the local information on all devices use the environment variable TNT_TENSORBOARD_ON_ALL_DEVICES:

TNT_TENSORBOARD_ON_ALL_DEVICES=true tarantella -- model.py

Note

At the moment, all of the other Keras callbacks will be executed on all devices with local information only.

For instance, the BaseLogger callback will be executed on each and every rank, and will log the acculumated metric averages for the local (micro-batch) information.

Important points

There is a number of points you should be aware of when using Tarantella.

Note

tnt.init() needs to be called after import tarantella as tnt, but before any other statement.

This will make sure the GPI-2 communication infrastructure is correctly initialized.

Note

Tarantella does not support custom training loops.

Instead of using custom training loops, please use Model.fit(...).

Note

Tarantella supports all TensorFlow optimizers with the exception of tf.keras.optimizers.Ftrl.

Since the Ftrl optimizer does not use batches, it is not supported in Tarantella.

Footnotes

1

That is, the global batch size must equal the micro batch size times the number of devices used.