Ctrl+K

Soft Sorting

Soft Sorting#

import sys

if "google.colab" in sys.modules:
    %pip install -q git+https://github.com/ott-jax/ott@main
%pip install -q torchvision

import collections
import functools
import io
import urllib
from typing import Any

from tqdm.notebook import tqdm

import jax
import jax.numpy as jnp
import numpy as np
import torchvision
from scipy import ndimage
from torch.utils import data

import flax.linen as nn
import optax
from flax import struct

import matplotlib.pyplot as plt

from ott.tools import soft_sort

Sorting operators#

Given an array of \(n\) numbers, several operators arise around the idea of sorting:

  • The sort() operator reshuffles the values in order, from smallest to largest.

  • The argsort() operator associates to each value its rank, when sorting in ascending order.

  • The quantile() operator considers a level value between \(0\) and \(1\), to return the element of the sorted array indexed at int(n * level), the median for instance if that level is set to \(0.5\).

  • The top_k() operator is equivalent to the sort()operator, but only returns the largest \(k\) values, namely the last \(k\) values of the sorted vector.

Here are some examples:

x = jnp.array([1.0, 5.0, 4.0, 8.0, 12.0])
jnp.sort(x)

DeviceArray([ 1.,  4.,  5.,  8., 12.], dtype=float32)

def rank(x):
    return jnp.argsort(jnp.argsort(x))


rank(x)

DeviceArray([0, 2, 1, 3, 4], dtype=int32)

jnp.quantile(x, q=0.5)

DeviceArray(5., dtype=float32)

Soft operators#

These sorting operators pop up everywhere in machine learning, but have several limitations when used in deep learning architectures.

For instance, ranks is integer valued: if used within a DL pipeline, one cannot differentiate through that step because the gradient of these integer values does not exist: the vector of ranks of a slightly perturbed vector \(x+\Delta x\) is either the same as that for \(x\), or switches values at some indices when inversions occur. This means that gradient is (almost always) 0 or (very rarely) infinite. Practically speaking, any loss or intermediary operation based on ranks will have 0 gradients.

This notebook shows ott provides soft counterparts to these operators. By soft, we mean differentiable, approximate proxies to these original “hard” operators. For instance, soft ranks() returned by ott operators won’t be integer valued, but instead floating point approximations; soft sort() will not contain exactly the \(n\) values contained in the input array, reordered, but instead \(n\) combinations of those values (sorted in increasing order) are very close to them.

These soft operators trade-off approximation for a more informative Jacobian. This trade-off is controlled by a non-negative parameter epsilon: The smaller epsilon, the closer to the original ranking and sorting operations; As epsilon gets bigger, these approximations will deviate from the true values, but will, instead, have more informative gradients; For very large epsilon, these approximations collapse to a mean value or a mean rank, and so does their gradients again . To facilitate that trade-off, we squash by default input values into the segment \([0,1]\) (using \(z\)-scores + a sigmoid) and set epsilon to \(0.01\). That epsilon corresponds to that used in regularized OT, see the documentation for Geometry and Epsilon.

The behavior of these operators is illustrated below.

Soft sort#

softsort_jitted = jax.jit(soft_sort.sort)
print(softsort_jitted(x))

[ 1.0504104  4.1228743  4.8620267  8.006994  11.957704 ]

As we can see, the values are close to what the original sorted x might be, but not exactly equal. Here, epsilon is set by default to \(10^{-2}\). A smaller epsilon reduces that gap, whereas a bigger one would tend to squash all returned values to the average of the input values.

print(softsort_jitted(x, epsilon=1e-4))
print(softsort_jitted(x, epsilon=1e-1))

[ 0.9983629  3.994969   5.006686   8.010008  11.989864 ]
[ 2.5705485  3.733634   5.462019   8.000692  10.233107 ]

Soft top-\(k\)#

The soft operators we propose build on a common idea: formulate sorting operations as optimal transports from an array of \(n\) values to a predefined target measure of \(m\) sorted points. The user is free to choose that measure, from the number of points \(m\), its locations stored in targets and weights, providing great flexibility depending on the use case.

Transporting an input discrete measure of \(n\) points towards one of \(m\) points results in a \(O(nm)\) complexity. The bigger \(m\), the more fine grained the quantities we recover. For instance, if we wish to get both a fine grained yet differentiable sorted vector, or vector of ranks, one can define a target measure of size \(m=n\), leading to a \(O(n^2)\) complexity.

On the contrary, if we are only interested in singling out a few important ranks, such as when considering \(\text{top-}k\) values, we can simply transport the inputs points onto \(k+1\) targets, \(k\) of them placed in large values weighting each \(1/n\), and a larger one on the smallest value, weighting \(1-k/n\). This also leads to a smaller complexity in \(O(nk)\).

top5 = jax.jit(functools.partial(soft_sort.sort, topk=5))

# Generates a vector of size 1000
big_x = jax.random.uniform(jax.random.PRNGKey(0), (1000,))
top5(big_x)

DeviceArray([0.9506124 , 0.97415835, 0.9832732 , 0.98791456, 0.9905736 ],            dtype=float32)

Soft ranks#

Similarly, we can compute soft ranks(), which do not output integer values, but provide instead a differentiable, float valued, approximation of the vector of ranks.

softranks = jax.jit(soft_sort.ranks)
print(softranks(x))

[0.01550213 1.8387042  1.148726   2.9977007  3.9927652 ]

Regularization effect#

As mentioned earlier, epsilon controls the trade-off between accuracy and differentiability. Larger epsilon tend to merge the soft ranks() of values that are close, up to the point where they all collapse to the average rank or average value.

epsilons = np.logspace(-3, 1, 100)
sorted_values = []
ranks = []
for e in epsilons:
    sorted_values.append(softsort_jitted(x, epsilon=e))
    ranks.append(softranks(x, epsilon=e))

fig, axes = plt.subplots(1, 2, figsize=(14, 5))

for values, ax, title in zip(
    (sorted_values, ranks), axes, ("sorted values", "ranks")
):
    ax.plot(epsilons, np.array(values), color="k", lw=11)
    ax.plot(epsilons, np.array(values), lw=7)
    ax.set_xlabel(r"$\epsilon$", fontsize=24)
    ax.tick_params(axis="both", which="both", labelsize=18)
    ax.set_title(f"Soft {title}", fontsize=24)
    ax.set_xscale("log")
../_images/0fabe32a15109abb42817bf38ad0d2540744a069f1f158153e66aca8a6e0b0a2.png

Note how none of the lines above cross. This is a fundamental property of soft sorting operators, proved in [Cuturi et al., 2019]: soft sorting and ranking operators are monotonic: the vector of soft sorted values will remain increasing for any epsilon, whereas if an input value \(x_i\) has a smaller (hard) rank than \(x_j\), its soft rank, for any value of epsilon, will also remain smaller than that for \(x_j\).

Soft quantiles#

To illustrate further the flexibility provided by setting target measures, one can notice that when a soft quantile() is targeted (for instance the soft median), the complexity becomes simply \(O(n)\). This is illustrated below to define “soft median” differentiable filter on a noisy image.

softquantile = jax.jit(soft_sort.quantile)
softquantile(x, q=0.5)

DeviceArray([4.995721], dtype=float32)

url = "https://raw.githubusercontent.com/matplotlib/matplotlib/master/doc/_static/stinkbug.png"
with urllib.request.urlopen(url) as resp:
    image = plt.imread(io.BytesIO(resp.read()))
image = image[..., 0]

def salt_and_pepper(im: np.array, amount: float = 0.05):
    result = np.copy(im)
    result = np.reshape(result, (-1,))
    num_noises = int(np.ceil(amount * im.size))
    indices = np.random.randint(0, im.size, num_noises)
    values = np.random.uniform(size=(num_noises,)) > 0.5
    result[indices] = values
    return np.reshape(result, im.shape)


noisy_image = salt_and_pepper(image, amount=0.1)

The generic_filter() from scipy does not run well on GPUs thus we force CPU execution for the following computations.

with jax.default_device(jax.devices("cpu")[0]):
    softmedian = functools.partial(soft_sort.quantile, level=0.5)

    fns = {"original": None, "median": np.median}
    for e in [0.01, 1.0]:
        fns[f"soft {e}"] = jax.jit(functools.partial(softmedian, epsilon=e))
    fns.update(mean=np.mean)

    fig, axes = plt.subplots(1, len(fns), figsize=(len(fns) * 6, 4))
    for key, ax in zip(fns, axes):
        fn = fns[key]
        soft_denoised = (
            ndimage.generic_filter(
                jnp.array(noisy_image), fn, footprint=jnp.ones((3, 3))
            )
            if fn is not None
            else noisy_image
        )
        ax.imshow(soft_denoised)
        ax.set_title(key, fontsize=22)
../_images/c8a4a0fd857cd5db73b93e5dbe83462a82e383595ff7b816387d5671c1eea3ed.png

Learning through a soft ranks operator#

A crucial feature of OTT lies in the ability it provides to differentiate seamlessly through any quantities that follow an optimal transport computation, making it very easy for end-users to plug them directly into end-to-end differentiable architectures.

In this tutorial we show how OTT can be used to implement a loss based on soft ranks. That soft \(0/1\) loss is used here to train a neural network for image classification, as done by [Cuturi et al., 2019].

This implementation relies on Flax and Optax libraries for creating and training neural networks with JAX. We also use PyTorch dataset and dataloaders.

Model#

We will train a vanilla CNN, in order to classify images from the MNIST dataset.

class ConvBlock(nn.Module):
    """A simple CNN block."""

    features: int = 32
    dtype: Any = jnp.float32

    @nn.compact
    def __call__(self, x, train: bool = True):
        x = nn.Conv(features=self.features, kernel_size=(3, 3))(x)
        x = nn.relu(x)
        x = nn.Conv(features=self.features, kernel_size=(3, 3))(x)
        x = nn.relu(x)
        x = nn.max_pool(x, window_shape=(2, 2), strides=(2, 2))
        return x


class CNN(nn.Module):
    """A simple CNN model."""

    num_classes: int = 10
    dtype: Any = jnp.float32

    @nn.compact
    def __call__(self, x, train: bool = True):
        x = ConvBlock(features=32)(x)
        x = ConvBlock(features=64)(x)
        x = x.reshape((x.shape[0], -1))  # flatten
        x = nn.Dense(features=512)(x)
        x = nn.relu(x)
        x = nn.Dense(features=self.num_classes)(x)
        return x

Losses & Metrics#

The \(0/1\) loss of a classifier on a labeled example is \(0\) if the logit of the true class ranks on top (here, would have rank 9, since MNIST considers 10 classes). Of course the \(0/1\) loss is non-differentiable, which is one reason we usually rely on the cross-entropy loss instead.

Here, as in [Cuturi et al., 2019], we consider a differentiable “soft” \(0/1\) loss by measuring the gap between the soft ranks() of the logit of the right answer and the target rank 9. If that gap \(>0\), then we incur a loss equal to that gap.

def cross_entropy_loss(logits: jnp.array, labels: jnp.array):
    logits = nn.log_softmax(logits)
    return -jnp.sum(labels * logits) / labels.shape[0]


def soft_error_loss(logits: jnp.array, labels: jnp.array):
    """The average distance between the best rank and the rank of the true class."""
    ranks_fn = jax.jit(functools.partial(soft_sort.ranks, axis=-1))
    soft_ranks = ranks_fn(logits)
    return jnp.mean(
        nn.relu(labels.shape[-1] - 1 - jnp.sum(labels * soft_ranks, axis=1))
    )


def compute_metrics(logits: jnp.array, labels: jnp.array, loss_fn: Any):
    loss = loss_fn(logits, labels)
    ce = cross_entropy_loss(logits, labels)
    accuracy = jnp.argmax(logits, -1) == jnp.argmax(labels, -1)
    return {
        "loss": jnp.mean(loss),
        "cross_entropy": jnp.mean(ce),
        "accuracy": jnp.mean(accuracy),
    }

To know more about training a neural network with Flax, please refer to the Flax ImageNet examples. After \(1\) epoch through the MNIST training examples, we are able to classify digits successfully, similar to what is done in [Cuturi et al., 2019] on CIFAR-10. We see that a soft \(0/1\) error loss, building on top of soft ranks(), can provide a competitive alternative to the cross entropy loss for classification tasks. As mentioned in that paper, that loss is less prone to overfitting.

class NameSpace:
    def __init__(self):
        pass


@struct.dataclass
class TrainState:
    step: int
    opt_state: Any
    model_state: Any
    params: Any


def create_train_state(rng, config, model):
    """Create initial training state."""
    params, model_state = initialized(
        rng, config.height, config.width, config.n_channels, model
    )
    opt_state = config.optimizer.init(params)
    state = TrainState(
        step=0, opt_state=opt_state, model_state=model_state, params=params
    )
    return state


def log(results, step, summary, train=True, tqdm_logger=None):
    """Log the metrics to stderr and tensorboard."""

    phase = "train" if train else "eval"
    for key in ("loss", "cross_entropy", "accuracy"):
        results[f"{phase}_{key}"].append((step + 1, summary[key]))

    prompt = "{} step: {}, loss: {:.3f}, cross entropy: {:.3f}, accuracy: {:.2%}".format(
        phase,
        step,
        summary["loss"],
        summary["cross_entropy"],
        summary["accuracy"],
    )
    if tqdm_logger is None:
        print(prompt)
    else:
        tqdm_logger.set_description_str(prompt)


def initialized(
    key, height: int, width: int, n_channels: int, model: nn.Module
):
    """Initialize the model parameters."""
    input_shape = (1, height, width, n_channels)

    @jax.jit
    def init(*args):
        return model.init(*args)

    variables = init({"params": key}, jnp.ones(input_shape, jnp.float32))
    model_state, params = variables.pop("params")
    return params, model_state


def train_step(apply_fn, loss_fn, optimizer, state, batch):
    """Perform a single training step."""

    def compute_loss(params):
        variables = {"params": params, **state.model_state}
        logits = apply_fn(variables, batch["image"])
        loss = loss_fn(logits, batch["label"])
        return loss, logits

    (loss, logits), grads = jax.value_and_grad(compute_loss, has_aux=True)(
        state.params
    )
    updates, new_opt_state = optimizer.update(
        grads, state.opt_state, state.params
    )
    new_params = optax.apply_updates(state.params, updates)
    metrics = compute_metrics(logits, batch["label"], loss_fn=loss_fn)
    new_state = state.replace(
        step=state.step + 1, opt_state=new_opt_state, params=new_params
    )
    return new_state, metrics


def eval_step(apply_fn, loss_fn, state, batch):
    params = state.params
    variables = {"params": params, **state.model_state}
    logits = apply_fn(variables, batch["image"], train=False, mutable=False)
    return compute_metrics(logits, batch["label"], loss_fn=loss_fn)


def train_and_evaluate(state, rng, config: NameSpace):
    """Execute model training and evaluation loop."""

    loss_fn = config.loss

    train_iter = data.DataLoader(
        config.train_dataset, batch_size=config.batch_size, shuffle=True
    )
    eval_iter = data.DataLoader(
        config.eval_dataset, batch_size=config.batch_size
    )

    v_train_step = jax.jit(
        functools.partial(
            train_step,
            model.apply,
            loss_fn,
            config.optimizer,
        )
    )
    v_eval_step = jax.jit(functools.partial(eval_step, model.apply, loss_fn))

    nb_batch_train = len(config.train_dataset) // config.batch_size
    nb_batch_eval = len(config.eval_dataset) // config.batch_size

    results = collections.defaultdict(list)

    tqdm_iter = tqdm(len(train_iter), total=nb_batch_train)
    tqdm_eval = tqdm(len(eval_iter), total=nb_batch_eval)
    for i_epoch in range(config.n_epochs):
        epoch_metrics = []
        for step, batch in enumerate(train_iter):
            state, metrics = v_train_step(
                state,
                {
                    "image": jnp.asarray(batch[0]),
                    "label": jnp.asarray(batch[1]),
                },
            )
            epoch_metrics.append(metrics)
            tqdm_iter.update(1)
            summary = jax.tree_map(lambda x: x.mean(), [metrics])[0]
            log(results, step + 1, summary, train=True, tqdm_logger=tqdm_iter)

            if step % config.nb_train_steps_between_eval == 0:
                epoch_metrics = []
                for step_eval, batch in enumerate(eval_iter):
                    metrics = v_eval_step(
                        state,
                        {
                            "image": jnp.asarray(batch[0]),
                            "label": jnp.asarray(batch[1]),
                        },
                    )
                    epoch_metrics.append(metrics)
                    tqdm_eval.update(1)
                summary = jax.tree_map(lambda x: x.mean(), [metrics])[0]
                log(
                    results,
                    step + 1,
                    summary,
                    train=False,
                    tqdm_logger=tqdm_eval,
                )
                tqdm_eval.reset()

        tqdm_iter.reset()

    return results, state

config = NameSpace()

config.batch_size = 1000
config.loss = soft_error_loss
config.learning_rate = 0.0005

config.train_dataset = torchvision.datasets.MNIST(
    "data",
    download=True,
    train=True,
    transform=lambda x: np.expand_dims(np.array(x), axis=2),
    target_transform=lambda x: np.asarray(jax.nn.one_hot(x, 10)),
)
config.eval_dataset = torchvision.datasets.MNIST(
    "data",
    download=True,
    train=False,
    transform=lambda x: np.expand_dims(np.array(x), axis=2),
    target_transform=lambda x: np.asarray(jax.nn.one_hot(x, 10)),
)

config.optimizer = optax.adamw(
    learning_rate=config.learning_rate, weight_decay=0.0001
)

config.height, config.width, config.n_channels = 28, 28, 1
config.num_classes = 10

config.n_epochs = 1
config.nb_train_steps_between_eval = 1

seed = 0
rng = jax.random.PRNGKey(seed)
model = CNN(num_classes=config.num_classes, dtype=jnp.float32)

init_state = create_train_state(rng, config, model)

As we are running this on CPU it may take a few minutes to run.

results, state = train_and_evaluate(init_state, rng, config)

Let us compare these results to training a neural net with the usual cross entropy.

config.loss = cross_entropy_loss
config.optimizer = optax.adam(learning_rate=config.learning_rate)
model = CNN(num_classes=config.num_classes, dtype=jnp.float32)

init_state = create_train_state(rng, config, model)

results_ce, state_ce = train_and_evaluate(init_state, rng, config)

fix, axes = plt.subplots(1, 3, figsize=(20, 6))
for j, ds in enumerate(["train", "eval"]):
    for i, metric in enumerate(["loss", "cross_entropy", "accuracy"]):
        axes[i].set_title(metric)
        vals = results[ds + "_" + metric]
        x = np.array(list(map(lambda x: x[0], vals)))
        y = np.array(list(map(lambda x: x[1], vals)))
        axes[i].plot(x, y, label=ds + " soft ranking", linewidth=2.0)
        if metric != "loss":
            # Only plot the loss for the soft ranking training.
            vals = results_ce[ds + "_" + metric]
            x = np.array(list(map(lambda x: x[0], vals)))
            y = np.array(list(map(lambda x: x[1], vals)))
            axes[i].plot(x, y, label=ds + " cross entropy", linewidth=2.0)

        if j == 1:
            axes[i].legend()
axes[0].set_title("Soft error loss")

Text(0.5, 1.0, 'Soft error loss')
../_images/8ccbf72c4a407337800cd54dce908b060b96088173c5d87428f22e3e1ee5c425.png
Contents