Reverb is an efficient and easy-to-use data storage and transport system designed for machine learning research.
Project description
Reverb
Reverb is an efficient and easy-to-use data storage and transport system designed for machine learning research. Reverb is primarily used as an experience replay system for distributed reinforcement learning algorithms but the system also supports multiple data structure representations such as FIFO, LIFO, and priority queues.
Table of Contents
Installation
Please keep in mind that Reverb is not hardened for production use, and while we do our best to keep things in working order, things may break or segfault.
:warning: Reverb currently only supports Linux based OSes.
The recommended way to install Reverb is with pip. We also provide instructions
to build from source using the same docker images we use for releases.
TensorFlow can be installed separately or as part of the pip install.
Installing TensorFlow as part of the install ensures compatibility.
$ pip install dm-reverb[tensorflow]
# Without Tensorflow install and version dependency check.
$ pip install dm-reverb
Nightly builds
$ pip install dm-reverb-nightly[tensorflow]
# Without Tensorflow install and version dependency check.
$ pip install dm-reverb-nightly
Debug builds
Starting with version 0.6.0, debug builds of Reverb are uploaded to Google Cloud
Storage. The builds can be downloaded or installed directly via pip following
the patterns below. gsutils can be used to navigate the directory structure
to ensure the files are there, e.g.
gsutil ls gs://rl-infra-builds/dm_reverb/builds/dbg. To build your own debug
binary, see the
build instructions.
For python 3.8 and 3.9 follow this pattern:
$ export reverb_version=0.8.0
# Python 3.9
$ export python_version=39
$ pip install https://storage.googleapis.com/rl-infra-builds/dm_reverb/builds/dbg/$reverb_version/dm_reverb-$reverb_version-cp$python_version-cp$python_version-manylinux2010_x86_64.whl
Build from source
This guide details how to build Reverb from source.
Reverb Releases
Due to some underlying libraries such as protoc and absl, Reverb has to be
paired with a specific version of TensorFlow. If installing Reverb as
pip install dm-reverb[tensorflow] the correct version of Tensorflow will be
installed. The table below lists the version of TensorFlow that each release of
Reverb is associated with and some versions of interest:
- 0.11.0 first version to support Python 3.11.
- 0.10.0 last version to support Python 3.7.
| Release | Branch / Tag | TensorFlow Version |
|---|---|---|
| Nightly | master | tf-nightly |
| 0.11.0 | v0.11.0 | 2.12.0 |
| 0.10.0 | v0.10.0 | 2.11.0 |
| 0.9.0 | v0.9.0 | 2.10.0 |
| 0.8.0 | v0.8.0 | 2.9.0 |
| 0.7.x | v0.7.0 | 2.8.0 |
Quick Start
Starting a Reverb server is as simple as:
import reverb
server = reverb.Server(tables=[
reverb.Table(
name='my_table',
sampler=reverb.selectors.Uniform(),
remover=reverb.selectors.Fifo(),
max_size=100,
rate_limiter=reverb.rate_limiters.MinSize(1)),
],
)
Create a client to communicate with the server:
client = reverb.Client(f'localhost:{server.port}')
print(client.server_info())
Write some data to the table:
# Creates a single item and data element [0, 1].
client.insert([0, 1], priorities={'my_table': 1.0})
An item can also reference multiple data elements:
# Appends three data elements and inserts a single item which references all
# of them as {'a': [2, 3, 4], 'b': [12, 13, 14]}.
with client.trajectory_writer(num_keep_alive_refs=3) as writer:
writer.append({'a': 2, 'b': 12})
writer.append({'a': 3, 'b': 13})
writer.append({'a': 4, 'b': 14})
# Create an item referencing all the data.
writer.create_item(
table='my_table',
priority=1.0,
trajectory={
'a': writer.history['a'][:],
'b': writer.history['b'][:],
})
# Block until the item has been inserted and confirmed by the server.
writer.flush()
The items we have added to Reverb can be read by sampling them:
# client.sample() returns a generator.
print(list(client.sample('my_table', num_samples=2)))
Continue with the Reverb Tutorial for an interactive tutorial.
Detailed overview
Experience replay has become an important tool for training off-policy
reinforcement learning policies. It is used by algorithms such as
Deep Q-Networks (DQN), Soft Actor-Critic (SAC),
Deep Deterministic Policy Gradients (DDPG), and
Hindsight Experience Replay, ... However building an efficient, easy to
use, and scalable replay system can be challenging. For good performance Reverb
is implemented in C++ and to enable distributed usage it provides a gRPC service
for adding, sampling, and updating the contents of the tables. Python clients
expose the full functionality of the service in an easy to use fashion.
Furthermore native TensorFlow ops are available for performant integration with
TensorFlow and tf.data.
Although originally designed for off-policy reinforcement learning, Reverb's flexibility makes it just as useful for on-policy reinforcement -- or even (un)supervised learning. Creative users have even used Reverb to store and distribute frequently updated data (such as model weights), acting as an in-memory lightweight alternative to a distributed file system where each table represents a file.
Tables
A Reverb Server consists of one or more tables. A table holds items, and each
item references one or more data elements. Tables also define sample and
removal selection strategies, a maximum item
capacity, and a rate limiter.
Multiple items can reference the same data element, even if these items exist in different tables. This is because items only contain references to data elements (as opposed to a copy of the data itself). This also means that a data element is only removed when there exists no item that contains a reference to it.
For example, it is possible to set up one Table as a Prioritized Experience Replay (PER) for transitions (sequences of length 2), and another Table as a (FIFO) queue of sequences of length 3. In this case the PER data could be used to train DQN, and the FIFO data to train a transition model for the environment.
Items are automatically removed from the Table when one of two conditions are met:
-
Inserting a new item would cause the number of items in the Table to exceed its maximum capacity. Table's removal strategy is used to determine which item to remove.
-
An item has been sampled more than the maximum number of times permitted by the Table's rate limiter. Such item is deleted.
Data elements not referenced anymore by any item are also deleted.
Users have full control over how data is sampled and removed from Reverb
tables. The behavior is primarily controlled by the
item selection strategies provided to the Table
as the sampler and remover. In combination with the
rate_limiter and max_times_sampled, a wide range of
behaviors can be achieved. Some commonly used configurations include:
Uniform Experience Replay
A set of N=1000 most recently inserted items are maintained. By setting
sampler=reverb.selectors.Uniform(), the probability to select an item is the
same for all items. Due to reverb.rate_limiters.MinSize(100), sampling
requests will block until 100 items have been inserted. By setting
remover=reverb.selectors.Fifo() when an item needs to be removed the oldest
item is removed first.
reverb.Table(
name='my_uniform_experience_replay_buffer',
sampler=reverb.selectors.Uniform(),
remover=reverb.selectors.Fifo(),
max_size=1000,
rate_limiter=reverb.rate_limiters.MinSize(100),
)
Examples of algorithms that make use of uniform experience replay include SAC and DDPG.
Prioritized Experience Replay
A set of N=1000 most recently inserted items. By setting
sampler=reverb.selectors.Prioritized(priority_exponent=0.8), the probability
to select an item is proportional to the item's priority.
Note: See Schaul, Tom, et al. for the algorithm used in this implementation of Prioritized Experience Replay.
reverb.Table(
name='my_prioritized_experience_replay_buffer',
sampler=reverb.selectors.Prioritized(0.8),
remover=reverb.selectors.Fifo(),
max_size=1000,
rate_limiter=reverb.rate_limiters.MinSize(100),
)
Examples of algorithms that make use of Prioritized Experience Replay are DQN (and its variants), and Distributed Distributional Deterministic Policy Gradients.
Queue
Collection of up to N=1000 items where the oldest item is selected and removed
in the same operation. If the collection contains 1000 items then insert calls
are blocked until it is no longer full, if the collection is empty then sample
calls are blocked until there is at least one item.
reverb.Table(
name='my_queue',
sampler=reverb.selectors.Fifo(),
remover=reverb.selectors.Fifo(),
max_size=1000,
max_times_sampled=1,
rate_limiter=reverb.rate_limiters.Queue(size=1000),
)
# Or use the helper classmethod `.queue`.
reverb.Table.queue(name='my_queue', max_size=1000)
Examples of algorithms that make use of Queues are IMPALA and asynchronous implementations of Proximal Policy Optimization.
Item selection strategies
Reverb defines several selectors that can be used for item sampling or removal:
- Uniform: Sample uniformly among all items.
- Prioritized: Samples proportional to stored priorities.
- FIFO: Selects the oldest data.
- LIFO: Selects the newest data.
- MinHeap: Selects data with the lowest priority.
- MaxHeap: Selects data with the highest priority.
Any of these strategies can be used for sampling or removing items from a Table. This gives users the flexibility to create customized Tables that best fit their needs.
Rate Limiting
Rate limiters allow users to enforce conditions on when items can be inserted and/or sampled from a Table. Here is a list of the rate limiters that are currently available in Reverb:
- MinSize: Sets a minimum number of items that must be in the Table before anything can be sampled.
- SampleToInsertRatio: Sets that the average ratio of inserts to samples by blocking insert and/or sample requests. This is useful for controlling the number of times each item is sampled before being removed.
- Queue: Items are sampled exactly once before being removed.
- Stack: Items are sampled exactly once before being removed.
Sharding
Reverb servers are unaware of each other and when scaling up a system to a multi server setup data is not replicated across more than one node. This makes Reverb unsuitable as a traditional database but has the benefit of making it trivial to scale up systems where some level of data loss is acceptable.
Distributed systems can be horizontally scaled by simply increasing the number of Reverb servers. When used in combination with a gRPC compatible load balancer, the address of the load balanced target can simply be provided to a Reverb client and operations will automatically be distributed across the different nodes. You'll find details about the specific behaviors in the documentation of the relevant methods and classes.
If a load balancer is not available in your setup or if more control is required then systems can still be scaled in almost the same way. Simply increase the number of Reverb servers and create separate clients for each server.
Checkpointing
Reverb supports checkpointing; the state and content of Reverb servers can be
stored to permanent storage. While checkpointing, the Server serializes all of
its data and metadata needed to reconstruct it. During this process the Server
blocks all incoming insert, sample, update, and delete requests.
Checkpointing is done with a call from the Reverb Client:
# client.checkpoint() returns the path the checkpoint was written to.
checkpoint_path = client.checkpoint()
To restore the reverb.Server from a checkpoint:
# The checkpointer accepts the path of the root directory in which checkpoints
# are written. If we pass the root directory of the checkpoints written above
# then the new server will load the most recent checkpoint written from the old
# server.
checkpointer = reverb.platform.checkpointers_lib.DefaultCheckpointer(
path=checkpoint_path.rsplit('/', 1)[0])
# The arguments passed to `tables=` must be the same as those used by the
# `Server` that wrote the checkpoint.
server = reverb.Server(tables=[...], checkpointer=checkpointer)
Refer to tfrecord_checkpointer.h for details on the implementation of checkpointing in Reverb.
Starting Reverb using reverb_server (beta)
Installing dm-reverb using pip will install a reverb_server script, which
accepts its config as a textproto. For example:
$ reverb_server --config="
port: 8000
tables: {
table_name: \"my_table\"
sampler: {
fifo: true
}
remover: {
fifo: true
}
max_size: 200 max_times_sampled: 5
rate_limiter: {
min_size_to_sample: 1
samples_per_insert: 1
min_diff: $(python3 -c "import sys; print(-sys.float_info.max)")
max_diff: $(python3 -c "import sys; print(sys.float_info.max)")
}
}"
The rate_limiter config is equivalent to the Python expression MinSize(1),
see rate_limiters.py.
Citation
If you use this code, please cite the Reverb paper as
@misc{cassirer2021reverb,
title={Reverb: A Framework For Experience Replay},
author={Albin Cassirer and Gabriel Barth-Maron and Eugene Brevdo and Sabela Ramos and Toby Boyd and Thibault Sottiaux and Manuel Kroiss},
year={2021},
eprint={2102.04736},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distributions
Built Distributions
File details
Details for the file dm_reverb_nightly-0.11.0.dev20230609-cp311-cp311-manylinux2014_x86_64.whl.
File metadata
- Download URL: dm_reverb_nightly-0.11.0.dev20230609-cp311-cp311-manylinux2014_x86_64.whl
- Upload date:
- Size: 6.3 MB
- Tags: CPython 3.11
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.1 CPython/3.8.10
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 | 54227220dfe4f31281e0d412da3a78592df1a2e6f649aa62d2218550542ae34d |
|
| MD5 | 38ae8d9e19fac40b3fd1f72a1700b103 |
|
| BLAKE2b-256 | 306daa627535ead9fd6cd255e42cc09c4dc7c5e69780e111f4d2cc02b5c465fd |
File details
Details for the file dm_reverb_nightly-0.11.0.dev20230609-cp310-cp310-manylinux2014_x86_64.whl.
File metadata
- Download URL: dm_reverb_nightly-0.11.0.dev20230609-cp310-cp310-manylinux2014_x86_64.whl
- Upload date:
- Size: 6.3 MB
- Tags: CPython 3.10
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.1 CPython/3.8.10
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 | 00d8c30713774bf07c65f9c862c66809c9d867f476cfa804c5a00fa095375655 |
|
| MD5 | ef491392a4080e90fce1835e0e65dbed |
|
| BLAKE2b-256 | d2092ecded387680a5b43811f6bffa1cface844e768a73541e9342deb597b450 |
File details
Details for the file dm_reverb_nightly-0.11.0.dev20230609-cp39-cp39-manylinux2014_x86_64.whl.
File metadata
- Download URL: dm_reverb_nightly-0.11.0.dev20230609-cp39-cp39-manylinux2014_x86_64.whl
- Upload date:
- Size: 6.4 MB
- Tags: CPython 3.9
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.1 CPython/3.8.10
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 | 50a6312c1ce2bbe2248b2f3e7b8da3d6a49f0c9314b7b8f8e75dec00781ddf4b |
|
| MD5 | 0bfea37cdd71d5b642a128db2518b07e |
|
| BLAKE2b-256 | 7be27218d28aafbe3e51fa1fcbc87643af4911ab0feabee6cd639c6df41f245b |
File details
Details for the file dm_reverb_nightly-0.11.0.dev20230609-cp38-cp38-manylinux2014_x86_64.whl.
File metadata
- Download URL: dm_reverb_nightly-0.11.0.dev20230609-cp38-cp38-manylinux2014_x86_64.whl
- Upload date:
- Size: 6.3 MB
- Tags: CPython 3.8
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.1 CPython/3.8.10
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 | 1959db8d56e7161909266d8256cf6d85f1265d5c7251a3b8316a724b93b8fc8c |
|
| MD5 | ab010c4b34b7f91a1ec9a0c6fa9b17a5 |
|
| BLAKE2b-256 | 44cd69edd436dd9beb0c26f6ab0f58f6c8098405cdb4d56e348d5d5a0e1581d4 |
