Prerequisites

Your must install PyTorch and PyTorch Lightning in your Python environment in order to follow this tutorial.

Tip

Using hydra-zen for your research project? Cite us! 😊

Run Boilerplate-Free ML Experiments with PyTorch Lightning & hydra-zen#

PyTorch Lightning is a library designed to eliminate the boilerplate code that is associated with training and testing neural networks in PyTorch. This is a natural bedfellow of Hydra and hydra-zen, which eliminate the boilerplate associated with designing software that is configurable, repeatable, and scalable.

Let’s use Hydra, hydra-zen, and PyTorch Lightning to configure and train multiple single-layer neural networks without any boilerplate code. For the sake of simplicity, we will train it to simply fit \(\cos{x}\) on \(x \in [-2\pi, 2\pi]\).

In this tutorial we will do the following:

  1. Define a simple neural network and lightning module.

  2. Create configs for our lighting module, data loader, optimizer, and trainer.

  3. Define a task-function for training and testing a model.

  4. Train four different models using combinations of two batch-sizes and two model-sizes (i.e. the number of neurons).

  5. Analyze our models’ results.

  6. Load our best model using the checkpoints saved by PyTorch Lightning and the job-config saved by Hydra.

Defining Our Model#

Create a script called zen_model.py (or, open a Jupyter notebook and include the following code. Here, we define our single-layer neural network and the lightning module that describes how to train and evaluate our model.

Contents of zen_model.py#
from typing import Callable, Type

import pytorch_lightning as pl
import torch as tr
import torch.nn as nn
import torch.nn.functional as F
from torch.optim import Optimizer  # type: ignore
from torch.utils.data import DataLoader, TensorDataset

from hydra_zen.typing import Partial

__all__ = ["UniversalFuncModule", "single_layer_nn", "train_and_eval"]


def single_layer_nn(num_neurons: int) -> nn.Module:
    """y = sum(V sigmoid(X W + b))"""
    return nn.Sequential(
        nn.Linear(1, num_neurons),
        nn.Sigmoid(),
        nn.Linear(num_neurons, 1, bias=False),
    )


class UniversalFuncModule(pl.LightningModule):
    def __init__(
        self,
        model: nn.Module,
        optim: Partial[Optimizer],
        dataloader: Type[DataLoader],
        target_fn: Callable[[tr.Tensor], tr.Tensor],
        training_domain: tr.Tensor,
    ):
        super().__init__()
        self.optim = optim
        self.dataloader = dataloader
        self.training_domain = training_domain
        self.target_fn = target_fn

        self.model = model

    def forward(self, x):
        return self.model(x)

    def configure_optimizers(self):
        # provide optimizer with model parameters
        return self.optim(self.parameters())

    def training_step(self, batch, batch_idx):
        x, y = batch
        # compute |cos(x) - model(x)|^2
        return F.mse_loss(self.model(x), y)

    def train_dataloader(self):
        # generate dataset: x, cos(x)
        x = self.training_domain.reshape(-1, 1)
        y = self.target_fn(x)
        return self.dataloader(TensorDataset(x, y))


def train_and_eval(
    model: tr.nn.Module,
    optim: Partial[Optimizer],
    dataloader: Type[DataLoader],
    target_fn: Callable[[tr.Tensor], tr.Tensor],
    training_domain: tr.Tensor,
    lit_module: Type[UniversalFuncModule],
    trainer: pl.Trainer,
):


    lit = lit_module(
        model=model,
        optim=optim,
        dataloader=dataloader,
        target_fn=target_fn,
        training_domain=training_domain,
    )

    # train the model
    trainer.fit(lit)

    # evaluate the model over the domain to assess the fit
    final_eval = lit(training_domain.reshape(-1, 1))
    final_eval = final_eval.detach().cpu().numpy().ravel()

    # return the final evaluation of our model:
    # a shape-(N,) numpy-array
    return final_eval

Attention

Type-annotations are not required by hydra-zen. However, they do enable runtime type-checking of configured values for our app.

Creating Our Configs and Task Function#

Create another script - named experiment.py - in the same directory as zen_model.py. Here, we will create the configs for our optimizer, model, data-loader, lightning module, and trainer. We’ll also define the task function that trains and tests our model.

Contents of experiment.py#
from math import pi

import pytorch_lightning as pl
from hydra_zen import builds, make_config, make_custom_builds_fn, zen
import torch as tr
from torch.optim import Adam  # type: ignore
from torch.utils.data import DataLoader

from zen_model import UniversalFuncModule, train_and_eval, single_layer_nn

pbuilds = make_custom_builds_fn(zen_partial=True, populate_full_signature=True)


ExperimentConfig = make_config(
    seed=1,
    lit_module=UniversalFuncModule,
    trainer=builds(pl.Trainer, max_epochs=100),
    model=builds(single_layer_nn, num_neurons=10),
    optim=pbuilds(Adam),
    dataloader=pbuilds(DataLoader, batch_size=25, shuffle=True, drop_last=True),
    target_fn=tr.cos,
    training_domain=builds(tr.linspace, start=-2 * pi, end=2 * pi, steps=1000),
)

# Wrapping `train_and_eval` with `zen` makes it compatible with Hydra as a task function
#
# We must specify `pre_call` to ensure that pytorch lightning seeds everything
# *before* any of our configs are instantiated (which will initialize the pytorch
# model whose weights depend on the seed)
pre_seed = zen(lambda seed: pl.seed_everything(seed))
task_function = zen(train_and_eval, pre_call=pre_seed)

if __name__ == "__main__":
    # enables us to call
    from hydra_zen import ZenStore

    store = ZenStore(deferred_hydra_store=False)
    store(ExperimentConfig, name="lit_app")

    task_function.hydra_main(
        config_name="lit_app",
        version_base="1.1",
        config_path=".",
    )

Be Mindful of What Your Task Function Returns

We could make this train_and_eval return our trained neural network, which would enable convenient access to it, in-memory, after our Hydra job completes. However, launching this task function in a multirun fashion will train multiple models and thus would keep all of those models in-memory (and perhaps on-GPU) simultaneously!

By not returning the model from our task function, we avoid the risk of hitting out-of-memory errors when training multiple large models.

Running Our Experiments#

We will use hydra_zen.launch() to run four jobs: training our model with all four combinations of:

  • a batch-size of 20 and 200

  • a model with 10 and 100 neurons

Open a Python console (or Jupyter notebook) in the same directory as experiment.py and run the following code.

Launching four jobs from a Python console.#
>>> from hydra_zen import launch
>>> from experiment import ExperimentConfig, task_function
>>> (jobs,) = launch(
...     ExperimentConfig,
...     task_function,
...     overrides=[
...         "dataloader.batch_size=20,200",
...         "model.num_neurons=10,100",
...     ],
...     multirun=True,
... )
[2021-10-24 21:23:32,556][HYDRA] Launching 4 jobs locally
[2021-10-24 21:23:32,558][HYDRA]     #0 : dataloader.batch_size=20 model.num_neurons=10
[2021-10-24 21:23:45,809][HYDRA]     #1 : dataloader.batch_size=20 model.num_neurons=100
[2021-10-24 21:23:58,656][HYDRA]     #2 : dataloader.batch_size=200 model.num_neurons=10
[2021-10-24 21:24:01,796][HYDRA]     #3 : dataloader.batch_size=200 model.num_neurons=100

Keep this Python console open; we will be making use of jobs in order to inspect our results.

Note that this is equivalent to running the following from the CLI:

Launching four jobs from the CLI.#
$ python experiment.py dataloader.batch_size=20,200 model.num_neurons=10,100 -m
[2021-10-24 21:23:32,556][HYDRA] Launching 4 jobs locally
[2021-10-24 21:23:32,558][HYDRA]     #0 : dataloader.batch_size=20 model.num_neurons=10
[2021-10-24 21:23:45,809][HYDRA]     #1 : dataloader.batch_size=20 model.num_neurons=100
[2021-10-24 21:23:58,656][HYDRA]     #2 : dataloader.batch_size=200 model.num_neurons=10
[2021-10-24 21:24:01,796][HYDRA]     #3 : dataloader.batch_size=200 model.num_neurons=100

Inspecting Our Results#

Visualizing Our Results#

Let’s begin inspecting our results by plotting our four models on \(x \in [-2\pi, 2\pi]\), alongside the target function: \(\cos{x}\). Continuing to work in our current Python console (or Jupyter notebook), run the following code and verify that you see the plot shown below.

Plotting our models#
>>> from hydra_zen import instantiate
>>> import matplotlib.pyplot as plt
>>> from matplotlib.axes import Axes

>>> x = instantiate(ExperimentConfig.training_domain)
>>> target_fn = instantiate(ExperimentConfig.target_fn)

>>> fig, ax = plt.subplots()
>>> assert isinstance(ax, Axes)
>>> ax.plot(x, target_fn(x), ls="--", label="Target")

>>> for j in jobs:
...     out = j.return_value
...     ax.plot(x, out, label=",".join(s.split(".")[-1] for s in j.overrides))
...
>>> ax.grid(True)
>>> ax.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
>>> plt.show()
Plot of four trained models vs the target function

Loading the Model of Best-Fit#

The 100-neuron model trained with a batch-size of 20 best fits our target function. Let’s load the model weights that were saved by PyTorch Lightning during training.

Continuing our work in the same Python console, let’s verify that job-1 corresponds to our desired model. Verify that you see the following outputs.

Job 1 corresponds to the 100-neuron model trained with batch-size 20.#
>>> best = jobs[1]
>>> best.cfg.dataloader.batch_size
20
>>> best.cfg.model.num_neurons
100

Next, we’ll load the config for this job. Recall that Hydra saves a .hydra/config.yaml file, which contains the complete configuration of this job – we can reproduce all aspects of it from this YAML.

Loading the complete config for this job#
>>> from hydra_zen import load_from_yaml, get_target, to_yaml
>>> from pathlib import Path

>>> outdir = Path(best.working_dir)
>>> cfg = load_from_yaml(outdir / ".hydra" / "config.yaml")

It is worth printing our this config to appreciate all of the exhaustive details that it captures about this job.

>>> print(to_yaml(cfg))  # fully details this job's config
seed: 1
lit_module:
  path: zen_model.UniversalFuncModule
  _target_: hydra_zen.funcs.get_obj
trainer:
  _target_: pytorch_lightning.trainer.trainer.Trainer
  max_epochs: 100
model:
  _target_: zen_model.single_layer_nn
  num_neurons: 100
optim:
  _target_: torch.optim.adam.Adam
  _partial_: true
  lr: 0.001
  betas:
  - 0.9
  - 0.999
  eps: 1.0e-08
  weight_decay: 0
  amsgrad: false
dataloader:
  _target_: torch.utils.data.dataloader.DataLoader
  _partial_: true
  batch_size: 20
  shuffle: true
  sampler: null
  batch_sampler: null
  num_workers: 0
  collate_fn: null
  pin_memory: false
  drop_last: true
  timeout: 0.0
  worker_init_fn: null
  multiprocessing_context: null
  generator: null
  prefetch_factor: 2
  persistent_workers: false
target_fn:
  path: torch.cos
  _target_: hydra_zen.funcs.get_obj
training_domain:
  _target_: torch.linspace
  start: -6.283185307179586
  end: 6.283185307179586
  steps: 1000

PyTorch Lightning saved the model’s trained weights as a .ckpt file in this job’s working directory. Let’s load these weights and use them to instantiate our lighting module.

Loading our lighting module with trained weights#
>>> from hydra_zen import zen
>>> from functools import partial
>>> *_, last_ckpt = sorted(outdir.glob("**/*.ckpt"))
>>> LitModule = get_target(cfg.lit_module)

>>> pload = partial(LitModule.load_from_checkpoint, last_ckpt)
>>> # extract top-level fields from `cfg`, instantiate them, and pass to `load_from_checkpoint`
>>> loaded = zen(pload, unpack_kwargs=True)(cfg)  # type: ignore

Finally, let’s double check that this loaded model behaves as-expected. Evaluating it at \(-\pi/2\), \(0\), and \(\pi/2\) should return, approximately, \(0\), \(1\), and \(0\), respectively.

Checkout our loaded model’s behavior#
>>> import torch as tr
>>> loaded(tr.tensor([-3.1415 / 2, 0.0, 3.1415 / 2]).reshape(-1, 1))
tensor([[0.0110],
        [0.9633],
        [0.0364]], grad_fn=<MmBackward>)

Math Details

For the interested reader… In this toy-problem we are optimizing arbitrary-width universal function approximators to fit \(\cos{x}\) on \(x \in [-2\pi, 2\pi]\). In mathematical notation, we want to solve the following optimization problem:

\[ \begin{align}\begin{aligned}F(\vec{v}, \vec{w}, \vec{b}; x) &= \sum_{i=1}^{N}{v_{i}\sigma(x w_i + b_i)}\\\vec{v}^*, \vec{w}^*, \vec{b}^* &= \operatorname*{arg\,min}_{\vec{v}, \vec{w}, \vec {b}\in\mathbb{R}^{N}} \; \|F(\vec{v}, \vec{w}, \vec{b}; x)\ - \cos{x}\|_{2}\\x &\in [-2\pi, 2\pi]\end{aligned}\end{align} \]

where \(N\) – the number of “neurons” in our layer – is a hyperparameter.

Attention

Cleaning Up: To clean up after this tutorial, delete the multirun directory that Hydra created upon launching our app. You can find this in the same directory as your experiment.py file.