Distributed wrapper for Tensorboard
Project description
Tensorplex is a multiplexed extension of the popular Tensorboard visualization tool. When you have a cluster, you can collect the learning curves from multiple running nodes and display them side-by-side on a single tensorboard web page.
Tensorplex makes extensive use of ZeroMQ under the hood, an efficient, robust, and lightweight distributed communication protocol.
Loggerplex is a subcomponent of Tensorplex that does lightweight distributed logging. It collects the real-time logs from multiple nodes and send them to a single master node for persistent book-keeping.
Tensorplex is not tied to Tensorflow and can be used with any machine learning frameworks that support numpy.
THIS DOC IS INCOMPLETE. MORE COMING SOON.
Installation
git clone https://github.com/StanfordVL/Tensorplex.git
pip install -e Tensorplex/
Demo
Go to Tensorplex/examples/. Change the tensorboard log folder in run_server.py script.
In one command line window, run python run_server.py. Then in another window, run python run_client.py. The server script should print out a list of dones.
Use tensorboard --logdir ~/Temp/tensorplex/ --port 8009 to view the results.
Manual
Tensorplex requires one long-running server script. Client scripts can connect and disconnect (e.g. client crashes) without impacting the server.
Tensorplex server
There are 3 steps to create the server script.
First, initialize a Tensorplex object with the root logging folder. Different clients will write to different sub-folders that are created automatically. max_processes is the number of processes that the server uses internally. Set it to 4 should be a sweet spot.
tplex = Tensorplex(
'~/my/logging/folder',
max_processes=4,
)
Second, register the client groups, which helps group Tensorflow curves into the same or different graph windows. The “client IDs” (explained later) in your client scripts must be consistent with the groups you register in the server.
There are 3 types of client groups:
register_normal_group(name): each graph will have only one curve in a normal group.
register_indexed_group(name, bin_size): each graph will have at most bin_size number of curves. Suppose you launch 42 agents with bin_size=10, the curves of agent 0-9 will be displayed in the same graph window; likewise, the curves of 10-19, 20-29, 30-39, 40-41 will be grouped in their respective graphs.
register_combined_group(name, group_criterion): TODO
To register multiple groups, you can chain the commands:
(tplex
.register_normal_group('learner') # 1 curve per graph
.register_indexed_group('agent', 8) # 8 agent learning curves per graph
.register_indexed_group('eval', 4) # 4 eval curves per graph
.register_combined_group('eval', get_eval_bin_name)
)
Third, you specify a port and launch the server. The script will be blocking:
tplex.start_server(8008)
# block main thread forever
Tensorplex client
Every TensorplexClient object must have a client ID that looks like <group_name>/<client_name>, i.e. two string names separated by /.
client = TensorplexClient(
client_id='agent/0', # for indexed group
# client_id='learner/system_stats', # for normal group, here client_name is `system_stats`
host='123.45.6.7', # server address to connect
port=8008, # server port to connect
)
Then you can write statistics to TensorplexClient
# most fundamental method
client.add_scalar(tag, 3.1415, integer_step)
# add_scalars is equivalent to multiple add_scalar() in one line
client.add_scalars({tag: 3.1415, tag2: 2.71828, tag3: 42}, integer_step)
There are
Note that tag in add_scalar behaves differently for different client group types.
For normal group,
Project details
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 Distribution
Built Distribution
File details
Details for the file Tensorplex-0.9.post1.tar.gz
.
File metadata
- Download URL: Tensorplex-0.9.post1.tar.gz
- Upload date:
- Size: 22.2 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/1.11.0 pkginfo/1.4.2 requests/2.18.4 setuptools/39.1.0 requests-toolbelt/0.8.0 tqdm/4.23.0 CPython/3.6.4
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 37c44fa6618bb7f0fc5be236ef2dbfb891c81ddca379630198170709129a2170 |
|
MD5 | 437f5f151604aa585440c19d1c0b6c73 |
|
BLAKE2b-256 | f0ebd94ee47b8b307294aa1439abe33c41ab8455c4b7eb150d5d84330cea3054 |
File details
Details for the file Tensorplex-0.9.post1-py3-none-any.whl
.
File metadata
- Download URL: Tensorplex-0.9.post1-py3-none-any.whl
- Upload date:
- Size: 27.1 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/1.11.0 pkginfo/1.4.2 requests/2.18.4 setuptools/39.1.0 requests-toolbelt/0.8.0 tqdm/4.23.0 CPython/3.6.4
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 754385c2c9f85e02243d12377b2bab041f5d25dde79f55b6d4268e4ab91fa8f7 |
|
MD5 | d38bc309ca1bc70e35e91457387c8fa0 |
|
BLAKE2b-256 | 6c65c84034380e9c58fd8aabded39f7ee8bdc770b1cbb9ea21e03d107e5fd411 |