Skip to content

Latest commit

 

History

History
247 lines (137 loc) · 11.7 KB

README.md

File metadata and controls

247 lines (137 loc) · 11.7 KB
Raw

experimental

LightEx

LightEx is a lightweight experiment framework to create, monitor and record your machine learning experiments. Targeted towards individual data scientists, researchers, small teams and, in general, resource-constrained experimentation. Compatible with all machine-learning frameworks.

Project Status: Alpha

Unlike most experiment frameworks, LightExsports a modular, and highly configurable design:

  • dispatcher: run experiments using process,docker containers or kubernetes pods. Switch between modes seamlessly by minor changes to config.
  • mulogger: log metrics and artifacts to multiple logger backends, using an unified API. Supports mlflow, tensorboard and trains — add new loggers easily as plugins.
  • namedconf: python dataclass based flexible and unified configuration specification for jobs, parameters and model architectures. Config instances are named and can be locally modified.
  • qviz: query, compare and visualize your experiment results.

The run environment and parameters for your experiments are specified using a config file lxconfig.py in your project directory. Modify, inherit, and create new named config instances, on-the-fly, as you perform a series of experiments.

Learn more about the anatomy of a ML experimentation framework here. Also, why [yet another experiment framework] (#yet-another-experiment-framework) ?

Benefits

Start with a basic train or eval project. In a few minutes,

  • introduce systematic logging (multiple loggers) and visualization to your project
  • go from running a single experiment to multiple parameterized experiments, e.g.,
    • multiple training runs over a set of hyper-parameters.
    • multiple efficient-net or bert train or eval runs.
    • a neural architecture search over multiple architectures in parallel.

Installation

pip install -U lightex

Quick Start

Imagine we have an existing ML project, with the following model train command:

train.py --data-dir ./data —-lr 0.1 -—hidden_dim 512

Now, let us bring in lightex to organize and scale experiments for the project. In the main project directory, initialize lightex :

lx init

This creates files lxconfig.py and run_expts.py. The file lxconfig.py contains pre-defined dataclasses for specifying named experiment configs.

  • The main class Config, contains three fields: er , hp and run (see the example below), which store configurations for logging and storage resources, hyper-parameters, and runtime, respectively.
  • By configuring these three fields (depending on the project environment, parameters), we can seamlessly generate multiple experiments via different dispatchers and log the results to one or more logger backends.
  • The config classes Resources, HP and Run for these fields are pre-defined in lxconfig.py. See config.md for full description of the defined dataclasses.
  • Config also includes a get_experiments function, which generates a list of experiment configs to be executed by the dispatcher. .
@dataclass
class Config:
    er: Resources 					#(Logging, Storage resources)
    hp: HP 						#(Hyper-parameters of model, training)
    run: Run 						#(Run-time config)

    def get_experiments(self): #required: generate a list of experiments to run
        expts = [Experiment(er=self.er, hp=self.hp, run=self.run)]
        return expts

To create a Config instance for our existing project, we instantiate

  • class HP with parameter values (lr and hidden_dim), and
  • class Run 's cmd field with the train command template. The template refers to Experiment objects returned by get_experiments function (of this Config instance).
    • Need not create the template manually. Use argparse_to_command in your existing code to generate it.
  • class Resources with default values (stores experiments in /tmp/ltex).
cmd="python train.py --data-dir {{run.data_dir}} --lr {{hp.lr}} --hidden_dim {{hp.hidden_dim}}" #template placeholders refer to fields of Experiment instance
Ru1 = Run(cmd=cmd, experiment_name="find_hp")
H2 = HP(lr=1e-2, hidden_dim=512)
R1 = Resources()

C1 = Config(er=R1, hp=H1, run=Ru1)

Once the config instance C1 is defined, run your experiments as follows:

python run_expts.py -c C1

That's it! Now, your experiments, logs, metrics and models are executed and recorded systematically.

Modify Experiment Parameters, Experiment Groups

Modify configs from previous experiments quickly using replace and run new experiments.

Example: Create a new HP instance and replace it in C1 to create a new Config. Recursive replace also supported.

H2 = HP(lr=1e-3, hidden_dim=1024)
C2 = replace(C1, hp=H2) #inherit er=R1 and run=Ru1

python run_expts.py -c C2

Experiment Groups

We can create multiple similar experiments programmatically and run them in parallel. To specify such experiment groups, specify a set of HPs in a HPGroup and update the get_experiments function (see scripts/lxconfig.py for an example). In the usual case, these experiments have different hyper-parameters but share Resources and Run instances.

Note: Although LightEx pre-defines the dataclass hierarchy, it allows the developer plenty of flexibility in defining the individual fields of classes, in particular, the fields of the HP class.

Adding Logging to your Code

Use the unified MultiLogger API to log metrics and artifacts to multiple logger backends.

Supported Loggers: mlflow, tensorboard,trains, wandb.

from lightex.mulogger import MLFlowLogger, MultiLogger, PytorchTBLogger
logger = MultiLogger(['mlflow', 'trains'])
logger.start_run()
# log to trains only
logger.log('trains', ltype='hpdict', value={'alpha': alpha, 'l1_ratio': l1_ratio})
# log to mlflow only
logger.log('mlflow', ltype='scalardict', value={'mae': mae, 'rmse': rmse, 'r2': r2}, step=1)
# log to all
logger.log('*', ltype='scalardict', value={'mae': mae, 'rmse': rmse, 'r2': r2}, step=3)
# log scalars and tensors, if supported by the logger backend
logger.log('trains', ltype='1d', name='W1', value=Tensor(..), histogram=True, step=4)
logger.end_run()

Or, use one of the existing loggers' API directly.

logger = MLFlowLogger(); mlflow = logger.mlflow  # call mlflow API
logger = PytorchTBLogger() ; writer = logger.writer #call tensorboard's API
# Similarly, use TrainsLogger for trains and WBLogger for wandb

Note: Except for changes in logging, no changes are required to your existing code! Setup the logger backend using scripts here.

Switch to Docker

Setting up the lxconfig instances pays off here!

Now, add a Dockerfile to your project which builds the runtime environment with all the project dependencies. Update the Build instance inside Resources config. See examples/sklearn, for example.

python run_expts.py -c C2 -e docker

Both your code and data are mounted on the container (no copying involved) — minimal disruption in your dev cycle.

Advanced Features

More advanced features are in development stage.

Running Experiments on multiple nodes / servers

If you've setup a docker swarm or kubernetes cluster, few changes to the existing config instance allow changing the underlying experiment dispatcher.

We need to virtualize code (by adding to Dockerfile) and storage.

Create a shared NFS on your nodes. Switch storage config to the NFS partition. Setup scripts will be added.

Better QViz module, Logger Plugins

Improvements to qviz module and a better plugin system for loggers being developed.

Setup Summary

In summary, LightEx involves the following one-time setup:

  • config values in lxconfig.py
  • Setup backend logger servers (only the ones required). Instructions here. (Optional)
  • Update logging calls in your code to call mulogger API. (Optional)
  • Dockerfile for your project, if you want to use containers for dispatch. (Optional)

While LightEx is quick to start with, it is advisable to spend some time understanding the config schema.

Dependencies

Python > 3.6 (require dataclasses, included during install). Other dependencies defined in setup.py.

Design Challenges

  • The configuration system allows you to abstract away the locations of the code (project directory), data directory, and the output directory, via the cmd field of Run. This enables running your code as-is across all dispatcher environments (process, docker, kubernetes).
  • A significant part of experiment manager design is about setting up and propagating a giant web of configuration variables.
    • No optimal choice here: json,yaml,jsonnet— all formats have issues.
    • Using dataclasses, we can write complex config specs, with built-in inheritance and ability to do local updates. Tiny bit of a learning curve here, bound to python, but we gain a lot of flexibility.
  • A unified mulogger API to abstract away the API of multiple logging backends.
  • Designing multiple dispatchers with similar API, enabling containers and varying storage options.
  • Read more on challenges here.

References

Most current (2019 Q3) tools focus on the logger component and provide selective qviz components. kubeflow and polyaxon are tied to the (k8s) dispatcher. Every tool has its own version of config management — mostly yaml based, where config types are absent or have a non-nested config class. Config-specific languages have been also proposed (ksonnet, sonnet, gin).

Yet Another Experiment Framework

Systematic experimentation tools are essential for a data scientist. Unfortunately, many existing tools (kubeflow, mlflow, polyaxon) are too monolithic, kubernetes-first, cloud-first, target very diverse audiences and hence spread too thin, and yet lack important dev-friendly features. sacred 's design is' tightly coupled and requires several sacred-specific changes to your main code. Other tools cater only to a specific task , e.g., tensorboard only handles log recording and visualization. Also, contrasting different experiment frameworks is hard: there is no standardized expt-management architecture for machine learning and most open-source frameworks are undergoing a process of adhoc requirements discovery.

Author

Nishant Sinha, OffNote Labs (nishant@offnote.co, @medium, @twitter)

Contributors

Akarsh E S, github