Reprieve: a library for evaluating representations

Everybody wants to learn good representations of data. However, defining precisely what we mean by a good representation can be tricky. In a recent paper, Evaluating representations by the complexity of learning low-loss predictors, we show that many notions of the quality of a representation for a task can be expressed as a function of the loss-data curve.

This repo contains a library called reprieve (for representation evaluation) for computing loss-data curves and the metrics of representation quality that can be derived from them. These metrics are:

• Validation loss
• Mutual information (approximate; a bound only)
• Minimum description length, from Information-Theoretic Probing with Minimum Description Length
• Surplus description length (our paper)
• ε-sample complexity (our paper)

We encourage anyone working on representation learning to bring their representations and datasets and use this library for evaluation and benchmarking. Don't settle for evaluating with linear probes or few-shot fine-tuning!

If you run into any problems, please file an issue! I'm happy to help get things working and learn how to make Reprieve better. Furthermore, if you're working in the field and find that we don't have a standard algorithm or representation that would be useful for others, send a pull request!

Features

This library is designed to be framework-agnostic and extremely efficient. Loss-data curves, and the associated measures like MDL and SDL, can be expensive to compute as they require training a probe algorithm dozens of times. This library reduces the time it takes to do this from 30 minutes to 2 when using probe algorithms in JAX.

• Bring your own dataset, representation function, and probe algorithm. We provide implementations of representation functions such as VAEs and supervised pretraining, and an MLP with Adam probe algorithm, but you can quickly and easily use your own.
• Framework-agnostic. You can implement representation functions and algorithms in any framework you choose, be it Pytorch, JAX, NumPy, or TensorFlow. Anything that can convert to and from NumPy arrays is fair game.
• Extremely fast. When using probing algorithms implemented in JAX, such as the standard MLP example we include, this library performs parallel training of dozens of networks at a time on a single GPU. Loss-data curves derived from training 100 small networks can be computed in about two minutes on one GPU. Yes, training 100 networks at once on one GPU.
• Publication-ready output. The library includes utilities for producing publication-quality plots and tables from the results. It even renders LaTeX tables including all the representation metrics.
• Simple to use. You can evaluate your representation according to five measures in only a few lines of code.

Examples

For more examples, see the examples folder. In particular examples/main.py is a complete example using fast parallel training and a JAX algorithm, and examples/main_torch.py is the same example using a Pytorch algorithm and no dependence on JAX.

import torchvision
import reprieve

# import a probing algorithm
from reprieve.algorithms import mlp as alg
from reprieve.representations import mnist_vae


# make a standard MNIST dataset
dataset_mnist = torchvision.datasets.MNIST(
    './data', train=True, download=True,
    transform=torchvision.transforms.Compose([
        torchvision.transforms.ToTensor(),
        torchvision.transforms.Normalize((0.1307,), (0.3081,))]))

# train a VAE on MNIST with an 8D latent space
vae_repr = mnist_vae.build_repr(8)

# make an MLP classifier algorithm with inputs of shape (8,) and 10 classes
# algorithms are represented by an initializer, a training step, and an eval step
init_fn, train_step_fn, eval_fn = alg.make_algorithm((8,), 10)

# construct a LossDataEstimator with this algorithm, dataset, and representation
vae_loss_data_estimator = reprieve.LossDataEstimator(
        init_fn, train_step_fn, eval_fn, dataset_mnist,
        representation_fn=vae_repr)

# compute the loss-data curve
loss_data_df = vae_loss_data_estimator.compute_curve(n_points=10)

# compute all the metrics and render the loss-data curve and a LaTeX table of results
metrics_df = reprieve.compute_metrics(
        loss_data_df, ns=[1000, 10000], epsilons=[0.5, 0.1])
reprieve.render_curve(loss_data_df, save_path='results.pdf')
reprieve.render_latex(metrics_df, save_path='results.tex')

You can also get started by running an example notebook on Colab: https://colab.research.google.com/github/willwhitney/reprieve/blob/master/examples/example.ipynb

Installation

1. git clone [email protected]:willwhitney/reprieve.git
2. Go install the Dependencies. Since installations of Pytorch and JAX are both highly context-dependent, I won't include an install script.
3. cd reprieve
4. pip install -e .

Dependencies

• Pytorch
• The standard Python data kit, including numpy and pandas.
• Optional:
  • For parallel training: JAX and Flax. Strongly recommended. Tested with jax==0.1.75, jaxlib==0.1.52, and flax==0.1.0. Everything here should work in principle with later versions, but will require some code changes to adapt to their moving APIs.
  • For generating and saving charts: Altair and altair_saver — pip install altair altair_saver. Note that altair_saver has some dependencies you need to manually install in order to produce PDFs.

Custom representations

As functions

In Reprieve, representations are structured as functions which transform the observed data. With a dataset (data_x, data_y), a representation will transform batches of data_x.

A representation is a function with the following signature:

representation_fn: np.ndarray[bsize, *x_shape] -> np.ndarray[bsize, *any]

All inputs and outputs to the representation function should be NumPy ndarrays. For convenience reprieve.representations.common contains a function numpy_wrap_torch which takes function on Pytorch tensors (such as an nn.Module) and returns a function that has ndarray inputs and outputs instead.

Note that this representation function should operate on batches of data.

Basically if you do

my_repr_module = MyReprModule()  # some Pytorch nn.Module
my_repr_fn = reprieve.representations.common.numpy_wrap_torch(my_repr_module)
lde = reprieve.LossDataEstimator(alg_init_fn, alg_train_fn, alg_eval_fn, dataset,
representation_fn=my_repr_fn)

you'll be all set.

For an example demonstrating this method, see reprieve.representations.vae.

As datasets

If in your use case it is simpler to provide a dataset of already-transformed observations, whether as a Pytorch Dataset or as a tuple (repr_x, data_y), you can do that too. For example, you could encode a whole dataset with a VAE, then pass that encoded data repr_x to LossDataEstimator instead of the original data_x.

When doing this simply do not pass an argument for representation_fn to LossDataEstimator and it will be left as the identity.

For an example demonstrating this method, see reprieve.mnist_noisy_label.

Algorithms

In reprieve an algorithm is a set of three functions:

1. An init_fn which takes a random seed and returns some state object.
2. A train_step_fn which takes in the current state and a batch of data (batch_x, batch_y) and returns an updated state and the loss on that batch.
3. An eval_fn which takes in the current state and a batch of data (batch_x, batch_y) and returns only the loss on that batch. Note that eval_fn must not mutate the state.

Built-in algorithms

We include a simple MLP trained with Adam as the default algorithm in reprieve. To use this:

1. from reprieve.algorithms import mlp as alg for the JAX version or from reprieve.algorithms import torch_mlp as alg for the Pytorch version
2. init_fn, train_step_fn, eval_fn = alg.make_algorithm(x_shape, n_classes) where x_shape is the shape of a single input to the network and n_classes is the number of classes.
3. Pass init_fn, train_step_fn, and eval_fn to LossDataEstimator. If using the JAX version, also pass vmap=True for fast performance. If using the Pytorch version you must set vmap=False or you will see weird errors.

The code for each of these algorithms is very simple and is meant to be easy to modify to use a different network architecture or optimizer.

Custom algorithms

Implementing your own algorithm (such as a convolutional probe network or a linear model trained with AdaDelta) is as simple as implementing those three functions. For an example, see the make_algorithm function of reprieve.algorithms.mlp (a JAX algorithm) or reprieve.algorithms.torch_mlp (a Pytorch algorithm).

We recommend starting from reprieve.algorithms.mlp and modifying the architecture or optimizer if you are writing your own algorithm.

Citing

If you use Reprieve in academic work, please cite our paper:

@misc{whitney2020evaluating,
    title={Evaluating representations by the complexity of learning low-loss predictors},
    author={William F. Whitney and Min Jae Song and David Brandfonbrener and Jaan Altosaar and Kyunghyun Cho},
    year={2020},
    eprint={2009.07368},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

Full API documentation

Class LossDataEstimator

method __init__(self, init_fn, train_step_fn, eval_fn, dataset,
                representation_fn=lambda x: x,
                val_frac=0.1, n_seeds=5,
                train_steps=5e3, batch_size=256,
                cache_data=True, whiten=True,
                use_vmap=True, verbose=False)

Create a LossDataEstimator.

Arguments:

• init_fn: (function int -> object) a function which maps from an integer random seed to an initial state for the training algorithm. this initial state will be fed to train_step_fn, and the output of train_step_fn will replace it at each step.
• train_step_fn: (function (object, (ndarray, ndarray)) -> (object, num) a function which performs one step of training. in particular, should map (state, batch) -> (new_state, loss) where state is defined recursively, initialized by init_fn and replaced by train_step, and loss is a Python number.
• eval_fn: (function (object, (ndarray, ndarray)) -> float) a function which takes in a state as produced by init_fn or train_step_fn, plus a batch of data, and returns the mean loss over points in that batch. should not mutate anything.
• dataset: a PyTorch Dataset or tuple (data_x, data_y).
• representation_fn: (function ndarray -> ndarray) a function which takes in a batch of observations from the dataset, given as a numpy array, and gives back an ndarray of transformed observations.
• val_frac: (float) the fraction of the data in [0, 1] to use for validation
• n_seeds: (int) how many random seeds to use for estimating each point. the seed is used for randomly sampling a subset dataset and for initializing the algorithm.
• train_steps: (number) how many batches of training to use with the algorithm. that is, how many times train_step_fn will be called on a batch of data.
• batch_size: (int) the size of the batches used for training and eval
• cache_data: (bool) whether to cache the entire dataset in memory. setting this to True will greatly improve performance by only computing the representation once for each point in the dataset
• whiten: (bool) whether to normalize the dataset's Xs to have zero mean and unit variance
• use_vmap: (bool) only for JAX algorithms. parallelize the training of by using JAX's vmap function. may cause CUDA out of memory errors; if this happens, call compute_curve with fewer points at a time or use a smaller probe
• verbose: (bool) print out informative messages and results as we get them

method LossDataEstimator.compute_curve(self, n_points=10, sampling_type='log', points=None)

Computes the loss-data curve for the given algorithm and dataset.

Arguments:

• n_points: (int) the number of points at which the loss will be computed to estimate the curve
• sampling_type: (str) how to distribute the n_points between 0 and len(dataset). valid options are 'log' (np.logspace) or 'linear' (np.linspace).
• points: (list of ints) manually specify the exact points at which to estimate the loss.

Returns: the current DataFrame containing the loss-data curve.

Effects: This LossDataEstimator instance will record the results of the experiments which are run, including them in the results DataFrame and using them to compute representation quality measures.

method LossDataEstimator.refine_esc(self, epsilon, precision, parallelism=10)

Runs experiments to refine an estimate of epsilon sample complexity. Performs experiments until the gap between an upper and lower bound is at most precision. This method is implemented as an iterative grid search.

Arguments:

• epsilon: (num) the tolerance specifying the maximum acceptable loss from running algorithm on dataset.
• precision: (num) how tightly to bound eSC, in terms of number of training points required to reach loss epsilon. that is, the desired upper_bound - lower_bound
• parallelism: (int) the number of experiments to run in each round of grid search.

Returns: an upper bound on the epsilon sample complexity

Effects: runs compute_curve multiple times and adds points to the loss-data curve

method LossDataEstimator.to_dataframe(self)

Return the current data for estimating the loss-data curve.

Functions

compute_metrics(df, ns=None, epsilons=[1.0, 0.1, 0.01])

Compute val loss, MDL, SDL, and eSC at the specified ns and epsilons.

Arguments:

• df: (pd.DataFrame) the dataframe containing a loss-data curve as returned by LossDataEstimator.compute_curve or LossDataEstimator.to_dataframe.
• ns: (list<num>) the list of training set sizes to use for computing metrics. this will be rounded up to the nearest point where the loss has been computed. set this to [len(dataset)] to compute canonical results.
• epsilons: (list<num>) the settings of epsilon used for computing SDL and eSC.

render_curve(df, ns=[], epsilons=[], save_path=None)

Render, and optionally save, a plot of the loss-data curve. Optionally takes arguments ns and epsilons to draw lines on the plot illustrating where metrics were calculated.

Arguments:

• df: (pd.DataFrame) the dataframe containing a loss-data curve as returned by LossDataEstimator.compute_curve or LossDataEstimator.to_dataframe.
• ns: (list<num>) the list of training set sizes to use for computing metrics.
• epsilons: (list<num>) the settings of epsilon used for computing SDL and eSC.
• save_path: (str) optional: a path (ending in .pdf or .png) to save the chart. saving requires the altair-saver package and its dependencies.

Returns: an Altair chart. Note that this chart displays well in notebooks, so calling render_curve(df) without a save path will work well with Jupyter.

render_latex(metrics_df, display=False, save_path=None)

Given a df of metrics from compute_metrics, renders a LaTeX table.

Arguments:

• metrics_df: (pd.DataFrame) a dataframe as returned by compute_metrics
• display: (bool) Jupyter only. render an output widget containing the latex string. necessary because otherwise lots of things will be double-escaped.
• save_path: (str) if specified, saves the text for the LaTeX table in a file.