mlpug

MLPug

MLPug is a library for training and evaluating Machine Learning (ML) models, able to use different ML libraries as backends.

A lot of the functionality you need to train and evaluate your model is independent of the ML library you're using, such as PyTorch, Jax, Apple MLX or TinyGrad. MLPug aims to provide a single framework with a unified API for all such training and evaluation functionality, independent of the ML library you are using.

Thus, when switching ML library, you don't have to learn a new training API and you can reuse your own training code with no, or minimal, change! 🤩🎉

MLPug is at version 0.1!

MLPug is still in development. If you are having trouble using MLPug for your use case, or when you have found a bug, please file an issue.

Dive right in! Installing MLPug as a Python package for your project.

pip install mlpug

# Using MLPug with PyTorch
import mlpug.pytorch as mlp

# Using MLPug with PyTorch/XLA (Training with Pytorch on TPUs)
import mlpug.pytorch.xla as mlp

# Using MLPug with Tensorflow
# WARNING: limited support, will not be developed and maintained further in the future
import mlpug.tensorflow as mlp

Basic template of usage

# ################# SETUP ##################
# A TrainModel uses your model to calculate and return the training loss and other basic data
train_model = TrainModel(classifier)

# The trainier knows how to train the model on a batch
trainer = mlp.trainers.DefaultTrainer(
    optimizers=optimizer,
    model_components=classifier
)

# At minimum, you want to log the loss to track training progress
# By default the batch loss and the moving average of the loss are calculated and logged
loss_evaluator = mlp.evaluation.MetricEvaluator(trainer=trainer)
callbacks = [
    mlp.callbacks.TrainingMetricsLogger(metric_evaluator=loss_evaluator),
    # Calculate validation loss only once per epoch over the whole dataset
    mlp.callbacks.DatasetMetricsLogger(validation_dataset,
                                       'validation',
                                       metric_evaluator=loss_evaluator,
                                       batch_level=False),
    # Log the progress during training                                       
    mlp.callbacks.LogProgress(log_period=progress_log_period, set_names=['training', 'validation']),
]

manager = mlp.trainers.TrainingManager(trainer,
                                       training_dataset,
                                       num_epochs=num_epochs,
                                       callbacks=callbacks)

trainer.set_training_model(train_model)

# ################# START! #################
manager.start_training()

Also see section Explaining the MLPug Hello World example.

Running the repository examples

You can find the example code here.

In the MLPug examples package there are three examples:

• A simple Hello World example
• A Fashion MNIST example with multiple commandline options for common MLPug features
• [PyTorch only] A complex example where we train a GPT2 chatbot based on the Persona dataset

To get started with the examples clone the MLPug repo:

git clone https://github.com/nuhame/mlpug.git

To run the examples, be sure to be in the root of the MLPug repository.

cd mlpug

Ensure that the current directory (the MLPug repo root) is added to the python path, such that python can find the example scripts:

export PYTHONPATH=.:$PYTHONPATH

It is advised to run the examples in a Python virtual environment:

python3 -m venv  ~/.virtualenvs/mlpug
# Activate the virtual environment
source ~/.virtualenvs/mlpug/bin/activate

In your >=Python3.9 virtual environment, install the basic requirements:

pip install -r requirements.txt 

Examples of using MLPug with PyTorch

Next to examples for PyTorch, there are equivalent examples for using MLPug with PyTorch/XLA and Tensorflow, see the tensorflow and pytorch/xla directories of the examples. For XLA you need to use a TPU on Google Cloud, or use Google Colab. Here we focus on PyTorch.

Hello world

Next to the default requirements, for the Hello World example, install the example specific requirements in your >=Python3.9 virtual environment:

pip install -r requirements.txt
pip install -r examples/hello_world/pytorch/requirements.txt

# MLPug Hello World example
python examples/hello_world/pytorch/train.py

Fashion MNIST

For the Fashion MNIST example install the following requirements:

pip install -r requirements.txt
pip install -r examples/fashion_mnist/pytorch/requirements.txt

# MLPug Fashion MNIST example
python examples/fashion_mnist/pytorch/train.py

# If you have multiple GPUs
python examples/fashion_mnist/pytorch/train.py --distributed

# To a batch size of 32, evaluated in chunks of 8 samples for gradient accumulation 
python examples/fashion_mnist/pytorch/train.py --batch-size 32 --batch-chunk-size 8

Some other useful flags are:

• --use-mixed-precision: When flag is set, mixed precision will be applied during training
• --eager-mode: When flag is set, forward and backward computation graphs will NOT be compiled (i.e. eager mode)

Run train.py -h for all options

Training a Persona Chatbot based on GPT2

Next to the default requirements, install the example specific requirements in your >=Python3.9 virtual environment:

pip install -r requirements.txt 
pip install -r examples/persona_chatbot/requirements.txt 
pip install -r examples/persona_chatbot/pytorch/requirements.txt 

What batch chunking size (micro batch size for gradient accumulation) depends on your system. The following setup assumes multiple GPUs (--distributed).

python examples/persona_chatbot/pytorch/train.py \
                 --experiment-name persona-bot-experiment \
                 --num-dataloader-workers 2 \
                 --batch-size 32 \
                 --num-choices 8 \
                 --sequence-length-outlier-threshold 0.05 \
                 --learning-rate 1e-4 \
                 --distributed \
                 --num-epochs 6 \
                 --progress-log-period 10

See train.py -h to see all options.

What is MLPug?

MLPug is a library for training and evaluating Machine Learning (ML) models, able to use different ML libraries as backends.

A lot of the functionality you need to train and evaluate your machine learning model is independent of the machine learning library you're using, such as PyTorch, Jax, Apple MLX or TinyGrad. For instance,

• checkpoint management,
• evaluation of validation set loss and other custom metrics,
• progress logging,
• progress visualization using Tensorboard,
• the use of gradient accumulation to train with large batch sizes using limited GPU memory, etc..

You need such functionality no matter what machine learning framework you are using.

MLPug aims to provide a single framework with a unified API for all such training and evaluation functionality, independent of the ML library you are using. This also implies that when you switch library you can reuse your training code with no, or minimal, changes.

Supported machine learning libraries

Currently, MLPug supports the following deep learning/machine learning libraries:

• PyTorch
• PyTorch/XLA (Training with Pytorch on TPUs)
• Tensorflow

Tensorflow is not fully supported at the moment and will not be further developed and maintained in the future, unless someone wants to contribute Tensorflow support to MLPug.

The ambition is to also add Jax, Apple MLX and TinyGrad as backends.

MLPug focus

Although MLPug should be able to deal with any training job, its functionality is mostly focussed on dealing with
training large models on large datasets, using limited hardware (GPU or TPU) resources and memory.

Detailed documentation

The following sections are documentation ToDo's, but provide insight in to MLPug's features:

Feature parity list

The logs object

Callbacks and the training life cycle

Progress Logging

Model components vs Training model

Distributed training

Checkpoint management
      Using the CheckpointManager
      Using training checkpoints
      Using model checkpoints
      Checkpointing on error or interrupt

MLPug metric evaluators
      Auxiliary batch training results
      Calculating custom metrics
      Conditional computation of metrics

Batch chunking, dealing with GPU memory limits
      Gradient Accumulation
      Chunked Metric Computation

Using Tensorboard
      Tensorboard made easy with AutoTensorboard
      More fine grained control

Learning Rate Scheduling

Multi GPU training

Mixed Precision Training

CUDA Memory tools

Using multiple optimizers

Explaining the MLPug Hello World example

Hello World with PyTorch

To use MLPug with Pytorch

import mlpug.pytorch as mlp

Before we can start training we need an iterable dataset that can provide our training batches.

training_dataset = torch.utils.data.DataLoader(training_data,
                                               batch_size=batch_size,
                                               shuffle=False,
                                               num_workers=3)

... and a model we want to train

classifier = torch.nn.Sequential(
    torch.nn.Flatten(),
    torch.nn.Linear(784, 128),
    torch.nn.ReLU(),
    torch.nn.Linear(128, 10))

MLPug needs a way to evaluate the loss of the model. One way to do that is to define a TrainModel that outputs the loss

class TrainModel(torch.nn.Module):
    def __init__(self, classifier):
        super(TrainModel, self).__init__()

        self.classifier = classifier
        self.loss_func = torch.nn.CrossEntropyLoss()

    def forward(self, batch_data, evaluate_settings, inference_mode=None):
        images, true_labels = batch_data

        logits = self.classifier(images)
        return self.loss_func(logits, true_labels)

train_model = TrainModel(classifier)

To train the model we will also need an optimizer

optimizer = torch.optim.Adam(classifier.parameters(), eps=1e-7)

To now use MLPug to start training, we need to create a Trainer which will be used by a TrainingManager.

trainer = mlp.trainers.DefaultTrainer(
    optimizers=optimizer, 
    model_components=classifier
)

MLPug uses a callback system allowing you to customize and extend the training functionality. The list of callback instances you provide the TrainingManager will be called using hooks at different stages of the training process.

# At minimum, you want to log the loss to track training progress
# By default the batch loss and the moving average of the loss are calculated and logged
loss_evaluator = mlp.evaluation.MetricEvaluator(trainer=trainer)
callbacks = [
    mlp.callbacks.TrainingMetricsLogger(metric_evaluator=loss_evaluator),
    # Calculate validation loss only once per epoch over the whole dataset
    mlp.callbacks.DatasetMetricsLogger(validation_dataset,
                                       'validation',
                                       metric_evaluator=loss_evaluator,
                                       batch_level=False),
    # Log the progress during training
    mlp.callbacks.LogProgress(log_period=progress_log_period, set_names=['training', 'validation']),
]

The TrainingMetricsLogger and the DatasetMetricsLogger callback instances log training and validation set loss values in a logs object that is passed through all callbacks during training. The LogProgress callback instance logs the metric values stored in the received logs object to the console (stdout).

We can now instantiate the TrainingManager and pass it the trainer.

manager = mlp.trainers.TrainingManager(trainer,
                                       training_dataset,
                                       num_epochs=num_epochs,
                                       callbacks=callbacks)

Before we can start training we still have to provide the train_model to the trainer.

trainer.set_training_model(train_model)

The final step is to actually start training:

manager.start_training()

Running examples/hello_world/pytorch/train.py finishes like this:

###############################################################################
Epoch 9/9	READY - Duration 0:00:08
Computed over sliding window:
training       : loss          0.239.

Computed over dataset:
validation     : loss          0.355.



INFO    : TrainingManager::_train : Training completed. All good! ❤️

Using the classifier ...
real label = 9, predicted label = 9

Hello World with PyTorch/XLA

The Hello World example with PyTorch/XLA, is largely the same as with PyTorch. There are only two small differences.

To use MLPug with Pytorch/XLA, load the correct backend

import mlpug.pytorch.xla as mlp

Load your model on a TPU core:

import torch_xla.core.xla_model as xm

...

device = xm.xla_device()

train_model = TrainModel(classifier, device)
classifier.to(device)

Hello World with Tensorflow

Below we will focus only on the minor differences between using MLPug with PyTorch and Tensorflow.

To use MLPug with Tensorflow

import mlpug.tensorflow as mlp

For the DefaultTrainer, you can specify the input batch_data_signature.

trainer = mlp.trainers.DefaultTrainer(optimizers=optimizer,
                                      model_components=classifier,
                                      batch_data_signature=(tf.TensorSpec(shape=(None, 28, 28), dtype=tf.float64),
                                                            tf.TensorSpec(shape=(None,), dtype=tf.uint8),))

When you run hello_world/tensorflow/train.py and hello_world/tensorflow/train_eager.py you will see that when running in eager mode, training is much slower.

Running hello_world/tensorflow/train_eager.py finishes like this:

###############################################################################
Epoch 9/9	READY - Duration 0:00:21
Computed over sliding window:
training       : loss          0.230.

Computed over dataset:
validation     : loss          0.349.



INFO    : TrainingManager::_train : Training completed. All good! ❤️

Using the classifier ...
real label = 9, predicted label = 9

Running hello_world/tensorflow/train.py finishes like this:

###############################################################################
Epoch 9/9	READY - Duration 0:00:09
Computed over sliding window:
training       : loss          0.235.

Computed over dataset:
validation     : loss          0.354.



INFO    : TrainingManager::_train : Training completed. All good! ❤️

Using the classifier ...
real label = 9, predicted label = 9

Note the difference in epoch duration!

Feature parity list

Feature	PyTorch	PyTorch/XLA	Tensorflow	JAX	Comments
Callbacks and training life cycle	✓	✓	✓
Progress Logging	✓	✓	✓
Distributed training	✓	✓	✓		Both multi-GPU and multi-TPU support for PyTorch and TF. TPU training with TF is untested
Distributed evaluation	✓	✓	✓		Both multi-GPU and multi-TPU support for PyTorch and TF. evaluation on TPU with TF is untested
Model and training checkpoint management	✓	✓	✓
Custom metric evaluation	✓	✓	✓
Conditional evaluation of metrics	✓	✓	✓
Batch Chunking: gradient accumulation	✓	✓	✓
Batch Chunking: chunked evaluation of metrics	✓	✓	✓
Tensorboard support	✓	✓	✓		Might be refactored
Learning Rate scheduling	✓	✓	✓		Might be refactored
Mixed Precision Training	✓	❌	❌		Should work with TF, but no specific support
Using multiple optimizers	✓	✓	✓