lr_finder.py

# Based on https://github.com/davidtvs/pytorch-lr-finder
# which is licensed as follows:
#
# MIT License
#
# Copyright (c) 2018 davidtvs
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.

import copy
import os
import sys
import tempfile

import torch
import numpy as np
from tqdm.autonotebook import tqdm
from torch.optim.lr_scheduler import _LRScheduler
from torch.utils.data import DataLoader
import matplotlib.pyplot as plt

from packaging import version

PYTORCH_VERSION = version.parse(torch.__version__)
TQDM_OPTIONS = {
    "file": sys.stdout,
    "dynamic_ncols": True,
}

try:
    from apex import amp

    IS_AMP_AVAILABLE = True
except ImportError:
    IS_AMP_AVAILABLE = False


class DataLoaderIter:
    def __init__(self, data_loader):
        self.data_loader = data_loader
        self._iterator = iter(data_loader)

    @property
    def dataset(self):
        return self.data_loader.dataset

    @staticmethod
    def inputs_labels_from_batch(batch_data):
        if not isinstance(batch_data, list) and not isinstance(batch_data, tuple):
            raise ValueError(
                f"Your batch type is not supported: {type(batch_data)}. Please inherit from "
                "`TrainDataLoaderIter` or `ValDataLoaderIter` and override the "
                "`inputs_labels_from_batch` method."
            )

        inputs, labels, *_ = batch_data

        return inputs, labels

    def __iter__(self):
        return self

    def __next__(self):
        batch = next(self._iterator)
        return self.inputs_labels_from_batch(batch)


class TrainDataLoaderIter(DataLoaderIter):
    def __init__(self, data_loader, auto_reset=True):
        super().__init__(data_loader)
        self.auto_reset = auto_reset

    def __next__(self):
        try:
            batch = next(self._iterator)
            inputs, labels = self.inputs_labels_from_batch(batch)
        except StopIteration:
            if not self.auto_reset:
                raise
            self._iterator = iter(self.data_loader)
            batch = next(self._iterator)
            inputs, labels = self.inputs_labels_from_batch(batch)

        return inputs, labels


class ValDataLoaderIter(DataLoaderIter):
    """This iterator will reset itself **only** when it is acquired by
    the syntax of normal `iterator`. That is, this iterator just works
    like a `torch.data.DataLoader`. If you want to restart it, you
    should use it like:

        ```
        loader_iter = ValDataLoaderIter(data_loader)
        for batch in loader_iter:
            ...

        # `loader_iter` should run out of values now, you can restart it by:
        # 1. the way we use a `torch.data.DataLoader`
        for batch in loader_iter:        # __iter__ is called implicitly
            ...

        # 2. passing it into `iter()` manually
        loader_iter = iter(loader_iter)  # __iter__ is called by `iter()`
        ```
    """

    def __init__(self, data_loader):
        super().__init__(data_loader)
        self.run_limit = len(self.data_loader)
        self.run_counter = 0

    def __iter__(self):
        if self.run_counter >= self.run_limit:
            self._iterator = iter(self.data_loader)
            self.run_counter = 0
        return self

    def __next__(self):
        self.run_counter += 1
        return super().__next__()


class LRFinder:
    """Learning rate range test.

    The learning rate range test increases the learning rate in a pre-training run
    between two boundaries in a linear or exponential manner. It provides valuable
    information on how well the network can be trained over a range of learning rates
    and what is the optimal learning rate.

    Arguments:
        model (torch.nn.Module): wrapped model.
        optimizer (torch.optim.Optimizer): wrapped optimizer where the defined learning
            is assumed to be the lower boundary of the range test.
        criterion (torch.nn.Module): wrapped loss function.
        device (str or torch.device, optional): a string ("cpu" or "cuda") with an
            optional ordinal for the device type (e.g. "cuda:X", where is the ordinal).
            Alternatively, can be an object representing the device on which the
            computation will take place. Default: None, uses the same device as `model`.
        memory_cache (boolean, optional): if this flag is set to True, `state_dict` of
            model and optimizer will be cached in memory. Otherwise, they will be saved
            to files under the `cache_dir`.
        cache_dir (string, optional): path for storing temporary files. If no path is
            specified, system-wide temporary directory is used. Notice that this
            parameter will be ignored if `memory_cache` is True.

    Example:
        >>> lr_finder = LRFinder(net, optimizer, criterion, device="cuda")
        >>> lr_finder.range_test(dataloader, end_lr=100, num_iter=100)
        >>> lr_finder.plot() # to inspect the loss-learning rate graph
        >>> lr_finder.reset() # to reset the model and optimizer to their initial state

    Reference:
    Cyclical Learning Rates for Training Neural Networks: https://arxiv.org/abs/1506.01186
    fastai/lr_find: https://github.com/fastai/fastai
    """

    def __init__(
        self,
        model,
        optimizer,
        criterion,
        device=None,
        memory_cache=True,
        cache_dir=None,
    ):
        # Check if the optimizer is already attached to a scheduler
        self.optimizer = optimizer
        self._check_for_scheduler()

        self.model = model
        self.criterion = criterion
        self.history = {"lr": [], "loss": []}
        self.best_loss = None
        self.memory_cache = memory_cache
        self.cache_dir = cache_dir

        # Save the original state of the model and optimizer so they can be restored if
        # needed
        self.model_device = next(self.model.parameters()).device
        self.state_cache = StateCache(memory_cache, cache_dir=cache_dir)
        self.state_cache.store("model", self.model.state_dict())
        self.state_cache.store("optimizer", self.optimizer.state_dict())

        # If device is None, use the same as the model
        if device:
            self.device = device
        else:
            self.device = self.model_device

    def reset(self):
        """Restores the model and optimizer to their initial states."""

        self.model.load_state_dict(self.state_cache.retrieve("model"))
        self.optimizer.load_state_dict(self.state_cache.retrieve("optimizer"))
        self.model.to(self.model_device)

    def range_test(
        self,
        train_loader,
        val_loader=None,
        start_lr=None,
        end_lr=10,
        num_iter=100,
        step_mode="exp",
        smooth_f=0.05,
        diverge_th=5,
        accumulation_steps=1,
        non_blocking_transfer=True,
    ):
        """Performs the learning rate range test.

        Arguments:
            train_loader (`torch.utils.data.DataLoader`
                or child of `TrainDataLoaderIter`, optional):
                the training set data loader.
                If your dataset (data loader) returns a tuple (inputs, labels,*) then
                Pytorch data loader object can be provided. However, if a dataset
                returns different outputs e.g. dicts, then you should inherit
                from `TrainDataLoaderIter` class and redefine `inputs_labels_from_batch`
                method so that it outputs (inputs, labels).
            val_loader (`torch.utils.data.DataLoader`
                or child of `ValDataLoaderIter`, optional): if `None` the range test
                will only use the training loss. When given a data loader, the model is
                evaluated after each iteration on that dataset and the evaluation loss
                is used. Note that in this mode the test takes significantly longer but
                generally produces more precise results.
                Similarly to `train_loader`, if your dataset outputs are not standard
                you should inherit from `ValDataLoaderIter` class and
                redefine method `inputs_labels_from_batch` so that
                it outputs (inputs, labels). Default: None.
            start_lr (float, optional): the starting learning rate for the range test.
                Default: None (uses the learning rate from the optimizer).
            end_lr (float, optional): the maximum learning rate to test. Default: 10.
            num_iter (int, optional): the number of iterations over which the test
                occurs. Default: 100.
            step_mode (str, optional): one of the available learning rate policies,
                linear or exponential ("linear", "exp"). Default: "exp".
            smooth_f (float, optional): the loss smoothing factor within the [0, 1[
                interval. Disabled if set to 0, otherwise the loss is smoothed using
                exponential smoothing. Default: 0.05.
            diverge_th (int, optional): the test is stopped when the loss surpasses the
                threshold:  diverge_th * best_loss. Default: 5.
            accumulation_steps (int, optional): steps for gradient accumulation. If it
                is 1, gradients are not accumulated. Default: 1.
            non_blocking_transfer (bool, optional): when non_blocking_transfer is set,
                tries to convert/move data to the device asynchronously if possible,
                e.g., moving CPU Tensors with pinned memory to CUDA devices. Default: True.

        Example (fastai approach):
            >>> lr_finder = LRFinder(net, optimizer, criterion, device="cuda")
            >>> lr_finder.range_test(dataloader, end_lr=100, num_iter=100)

        Example (Leslie Smith's approach):
            >>> lr_finder = LRFinder(net, optimizer, criterion, device="cuda")
            >>> lr_finder.range_test(trainloader, val_loader=val_loader, end_lr=1, num_iter=100, step_mode="linear")

        Gradient accumulation is supported; example:
            >>> train_data = ...    # prepared dataset
            >>> desired_bs, real_bs = 32, 4         # batch size
            >>> accumulation_steps = desired_bs // real_bs     # required steps for accumulation
            >>> dataloader = torch.utils.data.DataLoader(train_data, batch_size=real_bs, shuffle=True)
            >>> acc_lr_finder = LRFinder(net, optimizer, criterion, device="cuda")
            >>> acc_lr_finder.range_test(dataloader, end_lr=10, num_iter=100, accumulation_steps=accumulation_steps)

        If your DataLoader returns e.g. dict, or other non standard output, intehit from TrainDataLoaderIter,
        redefine method `inputs_labels_from_batch` so that it outputs (inputs, lables) data:
            >>> class TrainIter(TrainDataLoaderIter):
            >>>     def inputs_labels_from_batch(self, batch_data):
            >>>         return (batch_data['user_features'], batch_data['user_history']), batch_data['y_labels']
            >>> train_data_iter = TrainIter(train_dl)
            >>> finder = LRFinder(model, optimizer, criterion)
            >>> finder.range_test(train_data_iter, end_lr=10, num_iter=300, diverge_th=10)

        Reference:
        [Training Neural Nets on Larger Batches: Practical Tips for 1-GPU, Multi-GPU & Distributed setups](https://medium.com/huggingface/ec88c3e51255)
        [thomwolf/gradient_accumulation](https://gist.github.com/thomwolf/ac7a7da6b1888c2eeac8ac8b9b05d3d3)
        """

        # Reset test results
        self.history = {"lr": [], "loss": [], "apt": []}
        self.best_loss = None

        # Move the model to the proper device
        self.model.to(self.device)

        # Check if the optimizer is already attached to a scheduler
        self._check_for_scheduler()

        # Set the starting learning rate
        if start_lr:
            self._set_learning_rate(start_lr)

        # Initialize the proper learning rate policy
        if step_mode.lower() == "exp":
            lr_schedule = ExponentialLR(self.optimizer, end_lr, num_iter)
        elif step_mode.lower() == "linear":
            lr_schedule = LinearLR(self.optimizer, end_lr, num_iter)
        else:
            raise ValueError(f"expected one of (exp, linear), got {step_mode}")

        if smooth_f < 0 or smooth_f >= 1:
            raise ValueError("smooth_f is outside the range [0, 1[")

        # Create an iterator to get data batch by batch
        if isinstance(train_loader, DataLoader):
            train_iter = TrainDataLoaderIter(train_loader)
        elif isinstance(train_loader, TrainDataLoaderIter):
            train_iter = train_loader
        else:
            raise ValueError(
                f"`train_loader` has unsupported type: {type(train_loader)}."
                "Expected types are `torch.utils.data.DataLoader`"
                "or child of `TrainDataLoaderIter`."
            )

        if val_loader:
            if isinstance(val_loader, DataLoader):
                val_iter = ValDataLoaderIter(val_loader)
            elif isinstance(val_loader, ValDataLoaderIter):
                val_iter = val_loader
            else:
                raise ValueError(
                    f"`val_loader` has unsupported type: {type(val_loader)}."
                    "Expected types are `torch.utils.data.DataLoader`"
                    "or child of `ValDataLoaderIter`."
                )

        for iteration in tqdm(range(num_iter), **TQDM_OPTIONS):
            # Train on batch and retrieve loss
            loss, apt = self._train_batch(
                train_iter,
                accumulation_steps,
                non_blocking_transfer=non_blocking_transfer,
            )
            if val_loader:
                loss = self._validate(
                    val_iter, non_blocking_transfer=non_blocking_transfer
                )

            # append APT
            self.history["apt"].append(apt)

            # Update the learning rate
            self.history["lr"].append(lr_schedule.get_lr()[0])
            lr_schedule.step()

            # Track the best loss and smooth it if smooth_f is specified
            if iteration == 0:
                self.best_loss = loss
            else:
                if smooth_f > 0:
                    loss = smooth_f * loss + (1 - smooth_f) * self.history["loss"][-1]
                if loss < self.best_loss:
                    self.best_loss = loss

            # Check if the loss has diverged; if it has, stop the test
            self.history["loss"].append(loss)
            if loss > diverge_th * self.best_loss:
                print("Stopping early, the loss has diverged")
                break

        print("Learning rate search finished. See the graph with LRFinder.plot()")

    def _set_learning_rate(self, new_lrs):
        if not isinstance(new_lrs, list):
            new_lrs = [new_lrs] * len(self.optimizer.param_groups)
        if len(new_lrs) != len(self.optimizer.param_groups):
            raise ValueError(
                "Length of `new_lrs` is not equal to the number of parameter groups "
                + "in the given optimizer"
            )

        for param_group, new_lr in zip(self.optimizer.param_groups, new_lrs):
            param_group["lr"] = new_lr

    def _check_for_scheduler(self):
        for param_group in self.optimizer.param_groups:
            if "initial_lr" in param_group:
                raise RuntimeError("Optimizer already has a scheduler attached to it")

    def _train_batch(self, train_iter, accumulation_steps, non_blocking_transfer=True):
        self.model.train()
        total_loss = None  # for late initialization
        net_APT_temp = None
        self.optimizer.zero_grad()
        for i in range(accumulation_steps):
            inputs, labels = next(train_iter)
            inputs, labels = self._move_to_device(
                inputs, labels, non_blocking=non_blocking_transfer
            )

            # enable APT
            if hasattr(self.model, "temperature") and i == accumulation_steps - 1:
                self.model.reset_APT()

            # Forward pass
            outputs = self.model(inputs)
            loss = self.criterion(outputs, labels)

            # Loss should be averaged in each step
            loss /= accumulation_steps

            # Backward pass
            if IS_AMP_AVAILABLE and hasattr(self.optimizer, "_amp_stash"):
                # For minor performance optimization, see also:
                # https://nvidia.github.io/apex/advanced.html#gradient-accumulation-across-iterations
                delay_unscale = ((i + 1) % accumulation_steps) != 0

                with amp.scale_loss(
                    loss, self.optimizer, delay_unscale=delay_unscale
                ) as scaled_loss:
                    scaled_loss.backward()
            else:
                loss.backward()

            if total_loss is None:
                total_loss = loss
            else:
                total_loss += loss

        self.optimizer.step()

        # get APT
        if hasattr(self.model, "temperature") and accumulation_steps == 1:
            self.model(inputs)
            net_APT_temp = self.model.temperature.item()
            self.model.disable_APT()

        return total_loss.item(), net_APT_temp

    def _move_to_device(self, inputs, labels, non_blocking=True):
        def move(obj, device, non_blocking=True):
            if hasattr(obj, "to"):
                return obj.to(device, non_blocking=non_blocking)
            if isinstance(obj, tuple):
                return tuple(move(o, device, non_blocking) for o in obj)
            if isinstance(obj, list):
                return [move(o, device, non_blocking) for o in obj]
            if isinstance(obj, dict):
                return {k: move(o, device, non_blocking) for k, o in obj.items()}
            return obj

        inputs = move(inputs, self.device, non_blocking=non_blocking)
        labels = move(labels, self.device, non_blocking=non_blocking)
        return inputs, labels

    def _validate(self, val_iter, non_blocking_transfer=True):
        # Set model to evaluation mode and disable gradient computation
        running_loss = 0
        self.model.eval()
        with torch.no_grad():
            for inputs, labels in val_iter:
                # Move data to the correct device
                inputs, labels = self._move_to_device(
                    inputs, labels, non_blocking=non_blocking_transfer
                )

                # Forward pass and loss computation
                outputs = self.model(inputs)
                loss = self.criterion(outputs, labels)
                running_loss += loss.item() * len(labels)

        return running_loss / len(val_iter.dataset)

    def plot(
        self,
        skip_start=10,
        skip_end=5,
        log_lr=True,
        show_lr=None,
        ax=None,
        suggest_lr=True,
    ):
        """Plots the learning rate range test.

        Arguments:
            skip_start (int, optional): number of batches to trim from the start.
                Default: 10.
            skip_end (int, optional): number of batches to trim from the start.
                Default: 5.
            log_lr (bool, optional): True to plot the learning rate in a logarithmic
                scale; otherwise, plotted in a linear scale. Default: True.
            show_lr (float, optional): if set, adds a vertical line to visualize the
                specified learning rate. Default: None.
            ax (matplotlib.axes.Axes, optional): the plot is created in the specified
                matplotlib axes object and the figure is not be shown. If `None`, then
                the figure and axes object are created in this method and the figure is
                shown . Default: None.
            suggest_lr (bool, optional): suggest a learning rate by
                - 'steepest': the point with steepest gradient (minimal gradient)
                you can use that point as a first guess for an LR. Default: True.

        Returns:
            The matplotlib.axes.Axes object that contains the plot,
            and the suggested learning rate (if set suggest_lr=True).
        """

        if skip_start < 0:
            raise ValueError("skip_start cannot be negative")
        if skip_end < 0:
            raise ValueError("skip_end cannot be negative")
        if show_lr is not None and not isinstance(show_lr, float):
            raise ValueError("show_lr must be float")

        # Get the data to plot from the history dictionary. Also, handle skip_end=0
        # properly so the behaviour is the expected
        lrs = self.history["lr"]
        losses = self.history["loss"]
        apt = self.history["apt"]
        if skip_end == 0:
            lrs = lrs[skip_start:]
            losses = losses[skip_start:]
            apt = apt[skip_start:]
        else:
            lrs = lrs[skip_start:-skip_end]
            losses = losses[skip_start:-skip_end]
            apt = apt[skip_start:-skip_end]

        # Create the figure and axes object if axes was not already given
        fig = None
        if ax is None:
            fig, ax = plt.subplots()
        ax2 = ax.twinx()

        # Plot loss as a function of the learning rate
        ax.plot(lrs, losses)
        ax2.plot(lrs, apt, color="green")

        # Plot the suggested LR
        if suggest_lr:
            # 'steepest': the point with steepest gradient (minimal gradient)
            print("LR suggestion: steepest gradient")
            min_grad_idx = None
            try:
                min_grad_idx = (np.gradient(np.array(losses))).argmin()
            except ValueError:
                print(
                    "Failed to compute the gradients, there might not be enough points."
                )
            if min_grad_idx is not None:
                print(f"Suggested LR: {lrs[min_grad_idx]:.2E}")
                ax.scatter(
                    lrs[min_grad_idx],
                    losses[min_grad_idx],
                    s=75,
                    marker="o",
                    color="red",
                    zorder=3,
                    label="steepest gradient",
                )
                ax.legend()

        if log_lr:
            ax.set_xscale("log")
        ax.set_xlabel("Learning rate")
        ax.set_ylabel("Loss")
        ax2.set_ylabel("APT")

        if show_lr is not None:
            ax.axvline(x=show_lr, color="red")

        # Show only if the figure was created internally
        if fig is not None:
            plt.show()

        if suggest_lr and min_grad_idx:
            return ax, lrs[min_grad_idx]

        return ax


class LinearLR(_LRScheduler):
    """Linearly increases the learning rate between two boundaries over a number of
    iterations.

    Arguments:
        optimizer (torch.optim.Optimizer): wrapped optimizer.
        end_lr (float): the final learning rate.
        num_iter (int): the number of iterations over which the test occurs.
        last_epoch (int, optional): the index of last epoch. Default: -1.
    """

    def __init__(self, optimizer, end_lr, num_iter, last_epoch=-1):
        self.end_lr = end_lr

        if num_iter <= 1:
            raise ValueError("`num_iter` must be larger than 1")
        self.num_iter = num_iter

        super().__init__(optimizer, last_epoch)

    def get_lr(self):
        # In earlier Pytorch versions last_epoch starts at -1, while in recent versions
        # it starts at 0. We need to adjust the math a bit to handle this. See
        # discussion at: https://github.com/davidtvs/pytorch-lr-finder/pull/42
        if PYTORCH_VERSION < version.parse("1.1.0"):
            curr_iter = self.last_epoch + 1
            r = curr_iter / (self.num_iter - 1)
        else:
            r = self.last_epoch / (self.num_iter - 1)

        return [base_lr + r * (self.end_lr - base_lr) for base_lr in self.base_lrs]


class ExponentialLR(_LRScheduler):
    """Exponentially increases the learning rate between two boundaries over a number of
    iterations.

    Arguments:
        optimizer (torch.optim.Optimizer): wrapped optimizer.
        end_lr (float): the final learning rate.
        num_iter (int): the number of iterations over which the test occurs.
        last_epoch (int, optional): the index of last epoch. Default: -1.
    """

    def __init__(self, optimizer, end_lr, num_iter, last_epoch=-1):
        self.end_lr = end_lr

        if num_iter <= 1:
            raise ValueError("`num_iter` must be larger than 1")
        self.num_iter = num_iter

        super().__init__(optimizer, last_epoch)

    def get_lr(self):
        # In earlier Pytorch versions last_epoch starts at -1, while in recent versions
        # it starts at 0. We need to adjust the math a bit to handle this. See
        # discussion at: https://github.com/davidtvs/pytorch-lr-finder/pull/42
        if PYTORCH_VERSION < version.parse("1.1.0"):
            curr_iter = self.last_epoch + 1
            r = curr_iter / (self.num_iter - 1)
        else:
            r = self.last_epoch / (self.num_iter - 1)

        return [base_lr * (self.end_lr / base_lr) ** r for base_lr in self.base_lrs]


class StateCache:
    def __init__(self, in_memory, cache_dir=None):
        self.in_memory = in_memory
        self.cache_dir = cache_dir

        if self.cache_dir is None:
            self.cache_dir = tempfile.gettempdir()
        else:
            if not os.path.isdir(self.cache_dir):
                raise ValueError("Given `cache_dir` is not a valid directory.")

        self.cached = {}

    def store(self, key, state_dict):
        if self.in_memory:
            self.cached.update({key: copy.deepcopy(state_dict)})
            return

        fn = os.path.join(self.cache_dir, f"state_{key}_{id(self)}.pt")
        self.cached.update({key: fn})
        torch.save(state_dict, fn)

    def retrieve(self, key):
        if key not in self.cached:
            raise KeyError(f"Target {key} was not cached.")

        if self.in_memory:
            return self.cached.get(key)

        fn = self.cached.get(key)
        if not os.path.exists(fn):
            raise RuntimeError(
                f"Failed to load state in {fn}. File doesn't exist anymore."
            )
        state_dict = torch.load(fn, map_location=lambda storage, location: storage)
        return state_dict

    def __del__(self):
        """Check whether there are unused cached files existing in `cache_dir` before
        this instance being destroyed."""

        if self.in_memory:
            return

        for i in self.cached.values():
            if os.path.exists(i):
                os.remove(i)