refiners/tests/training_utils/test_trainer.py

import random
import warnings
from dataclasses import dataclass
from pathlib import Path
from typing import cast

import pytest
import torch
from torch import Tensor, nn
from torch.optim import SGD

from refiners.fluxion import layers as fl
from refiners.fluxion.utils import norm
from refiners.training_utils.callback import Callback, CallbackConfig
from refiners.training_utils.clock import ClockConfig
from refiners.training_utils.common import (
    Epoch,
    Step,
    count_learnable_parameters,
    human_readable_number,
    scoped_seed,
)
from refiners.training_utils.config import BaseConfig, IterationOrEpochField, ModelConfig, TimeValueField
from refiners.training_utils.data_loader import DataLoaderConfig, create_data_loader
from refiners.training_utils.trainer import (
    Trainer,
    TrainingClock,
    WarmupScheduler,
    count_learnable_parameters,
    human_readable_number,
    register_callback,
    register_model,
)


@dataclass
class MockBatch:
    inputs: torch.Tensor
    targets: torch.Tensor


class MockModelConfig(ModelConfig):
    use_activation: bool


class MockCallbackConfig(CallbackConfig):
    on_batch_end_interval: TimeValueField
    on_batch_end_seed: int
    on_optimizer_step_interval: IterationOrEpochField


class MockConfig(BaseConfig):
    # we register the `early_callback` before the `clock` callback to test the callback ordering
    early_callback: CallbackConfig = CallbackConfig()
    clock: ClockConfig = ClockConfig()

    mock_model: MockModelConfig
    mock_callback: MockCallbackConfig
    data_loader: DataLoaderConfig


class MockModel(fl.Chain):
    def __init__(self):
        super().__init__(
            fl.Linear(10, 10),
            fl.Linear(10, 10),
            fl.Linear(10, 10),
        )

    def add_activation(self) -> None:
        self.insert(1, fl.SiLU())
        self.insert(3, fl.SiLU())


class MockCallback(Callback["MockTrainer"]):
    def __init__(self, config: MockCallbackConfig) -> None:
        self.config = config
        self.optimizer_step_count = 0
        self.step_end_count = 0
        self.optimizer_step_random_int: int | None = None
        self.step_end_random_int: int | None = None

    def on_init_begin(self, trainer: "MockTrainer") -> None:
        pass

    def on_optimizer_step_begin(self, trainer: "MockTrainer") -> None:
        if not trainer.clock.is_due(self.config.on_optimizer_step_interval):
            return
        self.optimizer_step_count += 1
        self.optimizer_step_random_int = random.randint(0, 100)

    def on_step_end(self, trainer: "MockTrainer") -> None:
        if not trainer.clock.is_due(self.config.on_batch_end_interval):
            return

        # We verify that this callback is always called before the clock is updated (see `_call_callbacks` in trainer.py)
        assert trainer.clock.step // 3 <= self.step_end_count

        self.step_end_count += 1
        with scoped_seed(self.config.on_batch_end_seed):
            self.step_end_random_int = random.randint(0, 100)


class EarlyMockCallback(Callback["MockTrainer"]):
    """
    A callback that will be registered before the Clock callback to test the callback ordering.
    """

    def on_train_begin(self, trainer: "MockTrainer") -> None:
        assert trainer.clock.start_time is not None, "Clock callback should have been called before this callback."


class MockTrainer(Trainer[MockConfig, MockBatch]):
    step_counter: int = 0
    model_registration_counter: int = 0

    @property
    def dataset_length(self) -> int:
        return 20

    def get_item(self, index: int) -> MockBatch:
        return MockBatch(inputs=torch.randn(1, 10), targets=torch.randn(1, 10))

    def collate_fn(self, batch: list[MockBatch]) -> MockBatch:
        return MockBatch(
            inputs=torch.cat([b.inputs for b in batch]),
            targets=torch.cat([b.targets for b in batch]),
        )

    def create_data_iterable(self):
        return create_data_loader(
            get_item=self.get_item,
            length=self.dataset_length,
            config=self.config.data_loader,
            collate_fn=self.collate_fn,
        )

    @register_callback()
    def early_callback(self, config: CallbackConfig) -> EarlyMockCallback:
        return EarlyMockCallback()

    @register_callback()
    def mock_callback(self, config: MockCallbackConfig) -> MockCallback:
        return MockCallback(config)

    @register_model()
    def mock_model(self, config: MockModelConfig) -> MockModel:
        model = MockModel()
        if config.use_activation:
            model.add_activation()

        self.model_registration_counter += 1
        return model

    def compute_loss(self, batch: MockBatch) -> Tensor:
        self.step_counter += 1
        inputs, targets = batch.inputs.to(self.device), batch.targets.to(self.device)
        outputs = self.mock_model(inputs)
        return norm(outputs - targets)


@pytest.fixture
def mock_config() -> MockConfig:
    config = MockConfig.load_from_toml(Path(__file__).parent / "mock_config.toml")
    return config


@pytest.fixture
def mock_trainer(mock_config: MockConfig) -> MockTrainer:
    return MockTrainer(config=mock_config)


@pytest.fixture
def mock_trainer_short(mock_config: MockConfig) -> MockTrainer:
    mock_config_short = mock_config.model_copy(deep=True)
    mock_config_short.training.duration = Step(3)
    return MockTrainer(config=mock_config_short)


@pytest.fixture
def mock_model() -> fl.Chain:
    return MockModel()


def test_count_learnable_parameters_with_params() -> None:
    params = [
        nn.Parameter(torch.randn(2, 2), requires_grad=True),
        nn.Parameter(torch.randn(5), requires_grad=False),
        nn.Parameter(torch.randn(3, 3), requires_grad=True),
    ]
    # cast because of PyTorch 2.2, see https://github.com/pytorch/pytorch/issues/118736
    assert count_learnable_parameters(cast(list[nn.Parameter], params)) == 13


def test_count_learnable_parameters_with_model(mock_model: fl.Chain) -> None:
    assert count_learnable_parameters(mock_model.parameters()) == 330


def test_human_readable_number() -> None:
    assert human_readable_number(123) == "123.0"
    assert human_readable_number(1234) == "1.2K"
    assert human_readable_number(1234567) == "1.2M"


@pytest.fixture
def training_clock() -> TrainingClock:
    return TrainingClock(
        training_duration=Epoch(5),
        gradient_accumulation=Step(1),
        lr_scheduler_interval=Epoch(1),
    )


def test_timer_functionality(training_clock: TrainingClock) -> None:
    training_clock.start_timer()
    assert training_clock.start_time is not None
    training_clock.stop_timer()
    assert training_clock.end_time is not None
    assert training_clock.time_elapsed >= 0


def test_mock_trainer_initialization(mock_config: MockConfig, mock_trainer: MockTrainer) -> None:
    assert mock_trainer.config == mock_config
    assert isinstance(mock_trainer, MockTrainer)
    assert mock_trainer.optimizer is not None
    assert mock_trainer.lr_scheduler is not None
    assert mock_trainer.model_registration_counter == 1


def test_training_cycle(mock_trainer: MockTrainer) -> None:
    clock = mock_trainer.clock
    config = mock_trainer.config

    assert mock_trainer.step_counter == 0
    assert clock.epoch == 0

    mock_trainer.train()

    assert clock.epoch == config.training.duration.number
    assert mock_trainer.step_counter == mock_trainer.clock.step


def test_callback_registration(mock_trainer: MockTrainer) -> None:
    mock_trainer.train()

    # Check that the callback skips every other iteration
    assert mock_trainer.mock_callback.optimizer_step_count == mock_trainer.clock.iteration // 2
    assert mock_trainer.mock_callback.step_end_count == mock_trainer.clock.step // 3 + 1

    # Check that the random seed was set
    assert mock_trainer.mock_callback.optimizer_step_random_int == 93
    assert mock_trainer.mock_callback.step_end_random_int == 81


def test_training_short_cycle(mock_trainer_short: MockTrainer) -> None:
    clock = mock_trainer_short.clock
    config = mock_trainer_short.config

    assert mock_trainer_short.step_counter == 0
    assert mock_trainer_short.clock.epoch == 0

    mock_trainer_short.train()

    assert clock.step == config.training.duration.number


@pytest.fixture
def warmup_scheduler():
    optimizer = SGD([nn.Parameter(torch.randn(2, 2), requires_grad=True)], lr=0.1)
    scheduler = torch.optim.lr_scheduler.ConstantLR(optimizer, 1)
    return WarmupScheduler(optimizer, scheduler, warmup_scheduler_steps=100)


def test_initial_lr(warmup_scheduler: WarmupScheduler) -> None:
    optimizer = warmup_scheduler.optimizer
    for group in optimizer.param_groups:
        assert group["lr"] == 1e-3


def test_warmup_lr(warmup_scheduler: WarmupScheduler) -> None:
    with warnings.catch_warnings():
        warnings.filterwarnings(
            "ignore",
            category=UserWarning,
            message=r"Detected call of `lr_scheduler.step\(\)` before `optimizer.step\(\)`",
        )
        for _ in range(102):
            warmup_scheduler.step()
    optimizer = warmup_scheduler.optimizer
    for group in optimizer.param_groups:
        assert group["lr"] == 0.1


class MockTrainerWith2Models(MockTrainer):
    @register_model()
    def mock_model1(self, config: ModelConfig) -> MockModel:
        return MockModel()

    @register_model()
    def mock_model2(self, config: ModelConfig) -> MockModel:
        return MockModel()

    def compute_loss(self, batch: MockBatch) -> Tensor:
        self.step_counter += 1
        inputs, targets = batch.inputs.to(self.device), batch.targets.to(self.device)
        outputs = self.mock_model2(self.mock_model1(inputs))
        return norm(outputs - targets)


class MockConfig_2_Models(BaseConfig):
    mock_model1: ModelConfig
    mock_model2: ModelConfig


@pytest.fixture
def mock_config_2_models() -> MockConfig_2_Models:
    return MockConfig_2_Models.load_from_toml(Path(__file__).parent / "mock_config_2_models.toml")


@pytest.fixture
def mock_trainer_2_models(mock_config_2_models: MockConfig) -> MockTrainerWith2Models:
    return MockTrainerWith2Models(config=mock_config_2_models)


def test_optimizer_parameters(mock_trainer_2_models: MockTrainerWith2Models) -> None:
    assert len(mock_trainer_2_models.optimizer.param_groups) == 2
    assert mock_trainer_2_models.optimizer.param_groups[0]["lr"] == 1e-5


class MockTrainerNoDataLoader(MockTrainer):
    def create_data_iterable(self) -> list[MockBatch]:  # type: ignore
        return [MockBatch(inputs=torch.randn(4, 10), targets=torch.randn(4, 10)) for _ in range(5)]


def test_trainer_no_data_loader(mock_config: MockConfig) -> None:
    trainer = MockTrainerNoDataLoader(config=mock_config)
    trainer.train()
    assert trainer.step_counter == 500
    assert trainer.clock.epoch == 100