mlfab 0.2.6

A collection of core machine learning tools

mlfab

What is this?

This is a framework for trying out machine learning ideas.

Getting Started

Install the package using:

pip install mlfab

Or, to install the latest branch:

pip install 'mlfab @ git+https://github.com/kscalelabs/mlfab.git@master'

Simple Example

This framework provides an abstraction for quickly implementing and training PyTorch models. The workhorse for doing this is mlfab.Task, which wraps all of the training logic into a single cohesive unit. We can override functions on that method to get special functionality, but the default functionality is often good enough. Here's an example for training an MNIST model:

from dataclasses import dataclass

import torch.nn.functional as F
from dpshdl.dataset import Dataset
from dpshdl.impl.mnist import MNIST
from torch import Tensor, nn
from torch.optim.optimizer import Optimizer

import mlfab


@dataclass(kw_only=True)
class Config(mlfab.Config):
    in_dim: int = mlfab.field(1, help="Number of input dimensions")
    learning_rate: float = mlfab.field(1e-3, help="Learning rate to use for optimizer")
    betas: tuple[float, float] = mlfab.field((0.9, 0.999), help="Beta values for Adam optimizer")
    weight_decay: float = mlfab.field(1e-4, help="Weight decay to use for the optimizer")
    warmup_steps: int = mlfab.field(100, help="Number of warmup steps to use for the optimizer")


class MnistClassification(mlfab.Task[Config]):
    def __init__(self, config: Config) -> None:
        super().__init__(config)

        self.model = nn.Sequential(
            nn.Conv2d(config.in_dim, 32, 3, padding=1),
            nn.BatchNorm2d(32),
            nn.ReLU(),
            nn.Conv2d(32, 32, 3, padding=1),
            nn.BatchNorm2d(32),
            nn.ReLU(),
            nn.MaxPool2d(2),
            nn.Conv2d(32, 64, 3, padding=1),
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.Conv2d(64, 64, 3, padding=1),
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.MaxPool2d(2),
            nn.Flatten(),
            nn.Linear(64 * 7 * 7, 128),
            nn.BatchNorm1d(128),
            nn.ReLU(),
            nn.Linear(128, 10),
        )

    def set_loggers(self) -> None:
        self.add_logger(
            mlfab.StdoutLogger(),
            mlfab.TensorboardLogger(self.exp_dir),
        )

    def get_dataset(self, phase: mlfab.Phase) -> MNIST:
        root_dir = mlfab.get_data_dir() / "mnist"
        return MNIST(root_dir=root_dir, train=phase == "train")

    def forward(self, x: Tensor) -> Tensor:
        return self.model(x)

    def get_loss(self, batch: tuple[Tensor, Tensor], state: mlfab.State) -> Tensor:
        x_bhw, y_b = batch
        x_bchw = (x_bhw.float() / 255.0).unsqueeze(1)
        yhat_bc = self(x_bchw)
        self.log_step(batch, yhat_bc, state)
        return F.cross_entropy(yhat_bc, y_b)

    def log_valid_step(self, batch: tuple[Tensor, Tensor], output: Tensor, state: mlfab.State) -> None:
        (x_bhw, y_b), yhat_bc = batch, output

        def get_label_strings() -> list[str]:
            ypred_b = yhat_bc.argmax(-1)
            return [f"ytrue={y_b[i]}, ypred={ypred_b[i]}" for i in range(len(y_b))]

        self.log_labeled_images("images", lambda: (x_bhw, get_label_strings()))


if __name__ == "__main__":
    # python -m examples.mnist
    MnistClassification.launch(Config(batch_size=16))

Let's break down each part individually.

Config

Tasks are parametrized using a config dataclass. The ml.field function is a lightweight wrapper around dataclasses.field which is a bit more ergonomic, and ml.Config is a bigger dataclass which contains a bunch of other options for configuring training.

@dataclass(kw_only=True)
class Config(mlfab.Config):
    in_dim: int = mlfab.field(1, help="Number of input dimensions")

Model

All tasks should subclass ml.Task and override the generic Config with the task-specific config. This is very important, not just because it makes your life easier by working nicely with your typechecker, but because the framework looks at the generic type when resolving the config for the given task.

class MnistClassification(mlfab.Task[Config]):
    def __init__(self, config: Config) -> None:
        super().__init__(config)

        self.model = nn.Sequential(
            nn.Conv2d(config.in_dim, 32, 3, padding=1),
            nn.BatchNorm2d(32),
            nn.ReLU(),
            nn.Conv2d(32, 32, 3, padding=1),
            nn.BatchNorm2d(32),
            nn.ReLU(),
            nn.MaxPool2d(2),
            nn.Conv2d(32, 64, 3, padding=1),
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.Conv2d(64, 64, 3, padding=1),
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.MaxPool2d(2),
            nn.Flatten(),
            nn.Linear(64 * 7 * 7, 128),
            nn.BatchNorm1d(128),
            nn.ReLU(),
            nn.Linear(128, 10),
        )

Loggers

mlfab supports logging to multiple downstream loggers, and provides a bunch of helper functions for doing common logging operations, like rate limiting, converting image resolution to normal sizes, overlaying captions on images, and more.

If this function is not overridden, the task will just log to stdout.

def set_loggers(self) -> None:
    self.add_logger(
        mlfab.StdoutLogger(),
        mlfab.TensorboardLogger(self.exp_dir),
    )

Datasets

The task should return the dataset used for training, based on the phase. ml.Phase is a string literal with values in ["train", "valid"]. mlfab.get_data_dir() returns the data directory, which can be set in a configuration file which lives in ~/.mlfab.yml. The default configuration file will be written on first run if it doesn't exist yet.

def get_dataset(self, phase: mlfab.Phase) -> Dataset[tuple[Tensor, Tensor]]:
    root_dir = mlfab.get_data_dir() / "mnist"
    return MNIST(root_dir=root_dir, train=phase == "train")

Compute Loss

Each mlfab model should either implement the forward function, which should take a batch from the dataset and return the loss, or, if more control is desired, the get_loss function can be overridden.

def forward(self, x: Tensor) -> Tensor:
    return self.model(x)

def get_loss(self, batch: tuple[Tensor, Tensor], state: mlfab.State) -> Tensor:
    x_bhw, y_b = batch
    x_bchw = (x_bhw.float() / 255.0).unsqueeze(1)
    yhat_bc = self(x_bchw)
    self.log_step(batch, yhat_bc, state)
    return F.cross_entropy(yhat_bc, y_b)

Logging

When we call log_step in the get_loss function, it delegates to either log_train_step or log_valid_step, depending on what state.phase is. In this case, on each validation step we log images of the MNIST digits with the labels that our model predicts.

def log_valid_step(self, batch: tuple[Tensor, Tensor], output: Tensor, state: mlfab.State) -> None:
    (x_bhw, y_b), yhat_bc = batch, output

    def get_label_strings() -> list[str]:
        ypred_b = yhat_bc.argmax(-1)
        return [f"ytrue={y_b[i]}, ypred={ypred_b[i]}" for i in range(len(y_b))]

    self.log_labeled_images("images", lambda: (x_bhw, get_label_strings()))

Running

We can launch a training job using the launch class method. The config can be a Config object, or it can be the path to a config.yaml file located in the same directory as the task file. You can additionally provide the launcher argument, which supports training the model across multiple GPUs or nodes.

if __name__ == "__main__":
    MnistClassification.launch(Config(batch_size=16))