fluidml 0.1.3

FluidML is a lightweight framework for developing machine learning pipelines. Focus only on your tasks and not the boilerplate!

Develop ML pipelines fluently with no boilerplate code. Focus only on your tasks and not the boilerplate!

---

Key Features • Getting Started • Examples • Citation

---

FluidML is a lightweight framework for developing machine learning pipelines.

Developing machine learning models is a challenging process, with a wide range of sub-tasks: data collection, pre-processing, model development, hyper-parameter tuning and deployment. Each of these tasks is iterative in nature and requires lot of iterations to get it right with good performance.

Due to this, each task is generally developed sequentially, with artifacts from one task being fed as inputs to the subsequent tasks. For instance, raw datasets are first cleaned, pre-processed, featurized and stored as iterable datasets (on disk), which are then used for model training. However, this type of development can become messy and un-maintenable quickly for several reasons:

• pipeline code may be split across multiple scripts whose dependencies are not modeled explicitly
• each of this task contains boilerplate code to collect results from previous tasks (eg: reading from disk)
• hard to keep track of task artifacts and their different versions
• hyper-parameter tuning adds further complexity and boilerplate code

Key Features

FluidML provides following functionalities out-of-the-box:

• Task Graphs - Create ML pipelines or task graph using simple APIs
• Results Forwarding - Results from tasks are automatically forwarded to downstream tasks based on dependencies
• Parallel Processing - Execute the task graph parallely with multi-processing
• Grid Search - Extend the task graph by enabling grid search on tasks with just one line of code
• Result Caching - Task results are cached in a results store (eg: Local File Store or a MongoDB Store) and made available for subsequent runs without executing the tasks again and again
• Flexibility - Provides full control on your task implementations. You are free to choose any framework of your choice (Sklearn, TensorFlow, Pytorch, Keras, or any of your favorite library)

---

Getting Started

Installation

1. From Pip

Simply execute:

$ pip install fluidml

2. From Source

1. Clone the repository,
2. Navigate into the cloned directory (contains the setup.py file),
3. Execute $ pip install .

Note: To run demo examples, execute $ pip install fluidml[examples,rich-logging] (Pip) or $ pip install .[examples,rich-logging] (Source) to install the additional requirements.

Minimal Example

This minimal toy example showcases how to get started with FluidML. For real machine learning examples, check the "Examples" section below.

1. Basic imports

First, we import necessary classes from FluidML.

from fluidml import Flow, Swarm
from fluidml.common import Task, Resource
from fluidml.flow import GridTaskSpec, TaskSpec
from fluidml.storage import MongoDBStore, LocalFileStore, ResultsStore

2. Define Tasks

Next, we define some toy machine learning tasks. A Task can be implemented as a function or as a class inheriting from our Task class.

In case of the class approach, each task should implement the run() method, which takes some inputs and performs the desired functionality. These inputs are actually the results from predecessor tasks and are automatically forwarded by FluidML based on registered task dependencies. If the task has any hyper-parameters, they can be defined as arguments in the constructor. Additionally, within each task, users have access to methods and attributes like self.save() and self.resource to save its result and access task resources (more on that later).

class MyTask(Task):
    def __init__(self, kwarg_1, kwarg_2):
        ...
    def run(self, result_1, result2):
        ...

or

def my_task(result_1, result_2, kwarg_1, kwarg_2, task: Task):
    ...

In the case of defining the task as callable, an extra task object is provided to the task, which makes important internal attributes and functions like task.save() and task.resource available to the user.

Below, we define standard machine learning tasks such as dataset preparation, pre-processing, featurization and model training using Task classes. Notice that:

• Each task is implemented individually and it's clear what the inputs are (check arguments of run() method)
• Each task saves its results using self.save(...) by providing the object to be saved and a unique name for it. This unique name corresponds to input names in successor task definitions.

class DatasetFetchTask(Task):
    def run(self):
        ...
        # For InMemoryStore (default) and MongoDBStore type_ is NOT required
        # For LocalFileStore type_ IS required               
        self.save(obj=data_fetch_result, name='data_fetch_result', type_='json')


class PreProcessTask(Task):
    def __init__(self, pre_processing_steps: List[str]):
        super().__init__()
        self._pre_processing_steps = pre_processing_steps

    def run(self, data_fetch_result):
        ...
        self.save(obj=pre_process_result, name='pre_process_result')


class TFIDFFeaturizeTask(Task):
    def __init__(self, min_df: int, max_features: int):
        super().__init__()
        self._min_df = min_df
        self._max_features = max_features

    def run(self, pre_process_result):
        ...
        self.save(obj=tfidf_featurize_result, name='tfidf_featurize_result')


class GloveFeaturizeTask(Task):
    def run(self, pre_process_result):
        ...
        self.save(obj=glove_featurize_result, name='glove_featurize_result')


class TrainTask(Task):
    def __init__(self, max_iter: int, balanced: str):
        super().__init__()
        self._max_iter = max_iter
        self._class_weight = "balanced" if balanced else None

    def run(self, tfidf_featurize_result, glove_featurize_result):
        ...
        self.save(obj=train_result, name='train_result')


class EvaluateTask(Task):
    def run(self, train_result):
        ...
        self.save(obj=evaluate_result, name='evaluate_result')

3. Task Specifications

Next, we can create the defined tasks with their specifications. We now only write their specifications, later these are used to create real instances of tasks by FluidML. For each Task specification, we also add a list of result names that the corresponding task publishes and expects. Each published result object will be considered when results are automatically collected for a successor task.

dataset_fetch_task = TaskSpec(task=DatasetFetchTask, publishes=['data_fetch_result'])
pre_process_task = TaskSpec(task=PreProcessTask,
                            task_kwargs={
                                "pre_processing_steps": ["lower_case", "remove_punct"]},
                            expects=['data_fetch_result'],
                            publishes=['pre_process_result'])
featurize_task_1 = TaskSpec(task=GloveFeaturizeTask,
                            expects=['pre_process_result'],
                            publishes=['glove_featurize_result'])
featurize_task_2 = TaskSpec(task=TFIDFFeaturizeTask, task_kwargs={"min_df": 5, "max_features": 1000},
                            expects=['pre_process_result'],
                            publishes=['tfidf_featurize_result'])
train_task = TaskSpec(task=TrainTask, task_kwargs={"max_iter": 50, "balanced": True},
                      expects=['glove_featurize_result', 'tfidf_featurize_result'],
                      publishes=['train_result'])
evaluate_task = TaskSpec(task=EvaluateTask, expects=['train_result'], publishes=['evaluate_result'])

4. Registering task dependencies

Here we create the task graph by registering dependencies between the tasks. In particular, for each task specifier, you can register a list of predecessor tasks using the requires() method.

pre_process_task.requires(dataset_fetch_task)
featurize_task_1.requires(pre_process_task)
featurize_task_2.requires(pre_process_task)
train_task.requires([dataset_fetch_task, featurize_task_1, featurize_task_2])
evaluate_task.requires([dataset_fetch_task, featurize_task_1, featurize_task_2, train_task])

5. [optional] Define and instantiate Resources to share across all Tasks

Additionally, you can pass a list of resources (eg. seed and GPU devices) that are made available to the workers, which forward them to the corresponding tasks. You just have to create your own Resource dataclass, which inherits from our Resource interface. In this dataclass you can define all resources, e.g. seed, and the cuda device, which automatically is made available to all tasks through the self.resource or task.resource attribute.

@dataclass
class TaskResource(Resource):
    device: str
    seed: int

Let's assume our resources consist of a seed and a list of cuda device ids, e.g. ['cuda:0', 'cuda:1', 'cuda:0', 'cuda:1'], and we set num_workers=4. Then we can create our list of resources object with a simple list comprehension:

# create list of resources
resources = [TaskResource(device=devices[i], seed=seed) for i in range(num_workers)]

6. [optional] Results Store/Caching

By default, results of tasks are stored in an InMemoryStore, which might be impractical for large datasets/models. Also, the results are not persistent. To have persistent storage, FluidML provides two fully implemented ResultsStore namely LocalFileStore and MongoDBStore.

Additionally, users can provide their own results store to Swarm by inheriting from ResultsStore class and implementing load() and save(). Note, these methods rely on task name and its config parameters, which act as lookup-key for results. In this way, tasks are skipped by FluidML when task results are already available for the given config. But users can override and force execute tasks by passing force parameter to the Flow.

class MyResultsStore(ResultsStore):
    def load(self, name: str, task_name: str, task_unique_config: Dict) -> Optional[Any]:
        """ Query method to load an object based on its name, task_name and task_config if it exists """
        raise NotImplementedError

    def save(self, obj: Any, name: str, type_: str, task_name: str, task_unique_config: Dict, **kwargs):
        """ Method to save/update any artifact """
        raise NotImplementedError

We can instantiate for example a LocalFileStore

results_store = LocalFileStore(base_dir='/some/dir')

and pass it in the next step to Swarm to enable persistent results storing.

7. [optional] Configure Logging

FluidML internally utilizes Python's logging library. However, we refrain from configuring a logger object with handlers and formatters since each user has different logging needs and preferences. Hence, if you want to use FluidML's logging capability, you just have to do the configuration yourself. For convenience, we provide a simple utility function which configures a visually appealing logger (using a specific handler from the rich library).

from fluidml.common.logging import configure_logging
configure_logging()

Note: If you want to use logging in your application (e.g. within FluidML Tasks) but want to disable all FluidML internal logging messages you can simply call

logging.getLogger('fluidml').propagate = False

8. Run tasks using Flow and Swarm

Now that we have all the tasks specified, we can just run the task graph. For that, we have to create an instance of theSwarm class, by specifying a number of workers (n_dolphins :wink:). If n_dolphin is not set, it defaults internally to the number of CPU's available to the machine.

Next, you can create an instance of the Flow class and run the tasks utilizing one of our persistent result stores (defaults to InMemoryStore if no store is provided). Flow under the hood constructs the task graph and Swarm executes the graph in parallel while considering the registered dependencies.

tasks = [dataset_fetch_task, pre_process_task, featurize_task_1,
         featurize_task_2, train_task, evaluate_task]

with Swarm(n_dolphins=2,                        # optional (defaults to number of CPU's)
           resources=resources,                 # optional
           return_results=True,                 # optional
           results_store=results_store,         # optional
           ) as swarm:
    flow = Flow(swarm=swarm)
    results = flow.run(tasks)

Note: If the InMemoryStore is used, results of all the tasks are always returned by flow.run(), so that the user can store them manually. For the other shipped storages the user has the option to return or not return results (return_results=True/False). Task results can be accessed via task names, e.g. results["EvaluationTask"]. Our shipped result stores can be utilized to fetch specific task results from the returned result dictionary at any point via results_store.load().

Grid Search

Users can easily enable grid search for their tasks with just one line of code. To enable grid search on a particular task, we just have to wrap it with GridTaskSpec instead of TaskSpec.

train_task = GridTaskSpec(task=TrainTask,
                          gs_config={"max_iter": [50, 100],
                                     "balanced": [True, False],
                                     "layers": [[50, 100, 50]]},
                          gs_expansion_method='product'  # or 'zip'
                          )

That's it! Internally, Flow expands this task into 4 tasks with provided cross product combinations of max_iter and balanced. Alternatively, one can select zip as the expansion method, which would result in 2 expanded tasks, with the respective max_iter and balanced combinations of (50, True), (100, False). Generally, all values of type List will be unpacked to form grid search combinations. If a list itself is an argument and should not be expanded, it has to be wrapped again in a list. That is why layers is not considered for different grid search realizations. Further, any successor tasks (for instance, evaluate task) in the task graph will also be automatically expanded. Therefore, in our example, we would have 4 evaluate tasks, each one corresponding to the 4 train tasks.

Running a complete machine learning pipeline usually yields trained models for many grid search parameter combinations. A common goal is then to automatically determine the best hyper-parameter setup and the best performing model. FluidML enables just that by providing a reduce=True argument to the TaskSpec class. Hence, to automatically compare the 4 evaluate tasks and select the best performing model, we implement an additional ModelSelectionTask which gets wrapped by our TaskSpec class.

class ModelSelectionTask(Task):
    def run(self, reduced_results: List[Dict[str, Dict]]):
        # from all trained models/hyper-parameter combinations, determine the best performing model
        ...

model_selection_task = TaskSpec(task=ModelSelectionTask, reduce=True)

model_selection_task.requires(evaluate_task)

The important reduce=True argument enables that a single ModelSelectionTask instance gets the reduced results from all grid search expanded predecessor tasks. Every reduce=True task expects only the special reduced_results argument as input to the run method. It is a list of dictionaries where each dictionary holds the results and config of one specific grid search parameter combination. For example:

reduced_results = [
    {'result': {'result_name_1': result_1,
                'result_name_2': result_2},
     'config': {...}  # first unique parameter combination config
     },
    {'result': {'result_name_1': result_1,
                'result_name_2': result_2},
     'config': {...}  # second unique parameter combination config
     }
]

---

Examples

For real machine learning pipelines including grid search implemented with FluidML, check our Jupyter Notebook tutorials:

• Transformer based Sequence to Sequence Translation (PyTorch)
• Multi-class Text Classification (Sklearn)

---

Citation

@article{fluid_ml,
  title = {FluidML - a lightweight framework for developing machine learning pipelines},
  author = {Ramamurthy, Rajkumar and Hillebrand, Lars},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/fluidml/fluidml}},
}