A python module desgined for RL logging, monitoring and experiments managing.
Project description
UtilsRL
A util python module designed for reinforcement learning. Bug reports are welcomed.
Installation
You can install this package directly from pypi:
pip install UtilsRL
After installation, you may still need to configure some other dependencies based on your platform, such as PyTorch.
Features & Usage
1. Experiment Management
We provide a handful of functions for experiment management. All of them are placed under UtilsRL.exp
, including:
1.1 Argument Parsing
The argument parsing utils in this package provides three features:
- Supporting for multiple types of config files.
parse_args
can parse json, yaml, or even a python config module which is imported aheadfrom UtilsRL.exp import parse_args json_config = "/path/to/json" yaml_config = "/path/to/yaml" import config_module json_args = parse_args(json_config) yaml_Args = parse_args(yaml_config) module_args = parse_args(config_module) also_module_args = parse_args("/path/to/module")
- Nested argument parsing. We do this by introducing the
NameSpace
class. To be specific, if you passconvert=True
toparse_args
, then all of the dicts in the config file (including the argument dict itself) will be converted to a subclass ofNameSpace
. The contents wrapped inNameSpace
can be accessed both in dict manner and in attribute manner, and they will be formatted for better illustration when printing. For example, if we define a config module as follows:# in config module: config_module from UtilsRL.misc.namespace import NameSpace batch_size = 256 num_epochs = 10 class TrainerArgs(NameSpace): learning_rate = 1e-3 weight_decay = 1e-5 momentum = 0.9 class ActorArgs(NameSpace): epsilon = 0.05 class NetArgs(NameSpace): layer_num = 2 layer_nodes = 256
The we import and parse it inmain.py
:import config_module args = parse_args(config_module) print(args) print(">>>>>>>>>>>>>>>>>>>>>>") print(args.trainer.learning_rate)
The outputs are<NameSpace: args> |ActorArgs: <NameSpace: ActorArgs> |epsilon: 0.05 |NetArgs: <NameSpace: NetArgs> |layer_num: 2 |layer_nodes: 256 |TrainerArgs: <NameSpace: TrainerArgs> |learning_rate: 0.001 |weight_decay: 1e-05 |momentum: 0.9 |batch_size: 256 |num_epochs: 10 >>>>>>>>>>>>>>>>>>>>>> 0.001
- Argument updating. When calling
UtilsRL.exp.parse_args
, the command-line argument is parsed and used to overwrite the arguments defined in config file. If the specific argument is nested, then you can use dot.
to separate each NameSpace, likepython main.py --TrainerArgs.momentum 0.8
.from UtilsRL.exp import parse_args # get arguments from file/config module # this will automatically parse arguments from command line and update them to args args = parse_args("/path/to/file") # manually update with command line arguments args = update_args(args, unknown)
1.2 Device and Seed Management
We provide a set of utils functions of selecting device and setting seed in UtilsRL.exp.device
UtilsRL.exp.seed
. Please take time and check these files.
A setup function is available in top-level UtilsRL.exp
, which will setup the arguments with logger, device and seed which you provide.
from UtilsRL.exp import setup
setup(args, logger=None, device="cuda:0", seed=None) # seed will be initialized randomly
setup(args, logger=None, device=None, seed="4234") # a most free gpu will be selected as device
1.3 Snapshot
You can make a snapshot of the experiment code by passing --UtilsRL.snapshot <name>
to the program. UtilsRL will commit all the changes to a new branch whose name is <name>
, and then return to the original branch.
After creating the branch, its name will be added to args
. You can find its name by args.UtilsRL.snapshot_branch
, and git diff that branch later to checkout the changes from which you made.
2. Monitor
Monitor listens at the main loop of the training process, and displays the process with tqdm meter.
from UtilsRL.monitor import Monitor
monitor = Monitor(desc="test_monitor")
for i in monitor.listen(range(5)):
time.sleep(0.1)
You can register callback functions which will be triggered at certain stage of the training. For example, we can register a callback which will email us when training is done:
monitor = Monitor(desc="test_monitor")
monitor.register_callback(
name= "email me at the end of training",
on = "exit",
callback = Monitor.email,
...
)
You can also register context variables for training, which will be automatically managed by monitor. In the example below, the registered context variables (i.e. self.actor
and local_var
) will be saved every 100 iters.
monitor = Monitor(desc="test_monitor", out_dir="./out")
def train():
local_var = ...
local_var = monitor.register_context("local_var", save_every=100)
for i_epoch in monitor.listen(range(1000)):
# do training
train()
As a more complex example, we can use the Monitor to resume training from a certain iteration, and restore the context variables from checkpoints:
class Trainer():
def __init__(self):
self.actor = ...
def train(self):
local_var = ...
# load previous saved checkpoints specified by `load_path`
self.actor, local_var = \
monitor.register_context(["self.actor", "local_var"], load_path="/path/to/checkpoint/dir").values()
# use `initial` to designate the start point
for i_epoch in monitor.listen(range(1000), initial=100):
# continue training
3. Logger
Logger provides a rather shallow capsulation for torch.utils.tensorboard.SummaryWriter
.
from UtilsRL.logger import TensorboardLogger
# create a logger, with terminal output enabled and file logging disabled
logger = TensorboardLogger(log_dir="./logs", name="debug", terminal=True, txt=False)
# log a sentence in color blue.
logger.log_str("This is a sentence", type="LOG")
# log sentence in color red.
logger.log_str("Here occurs an error", type="ERROR")
# log scalar and a dict of scalars repectively
logger.log_scala(tag="var_name", value=1.0, step=1)
logger.log_scalas(main_tag="group_name", tag_scalar_dict={
"var1": 1.0,
"var2": 2.0
}, step=1)
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 UtilsRL-0.3.5.tar.gz
.
File metadata
- Download URL: UtilsRL-0.3.5.tar.gz
- Upload date:
- Size: 19.9 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/3.8.0 pkginfo/1.8.2 readme-renderer/34.0 requests/2.27.1 requests-toolbelt/0.9.1 urllib3/1.26.9 tqdm/4.63.0 importlib-metadata/4.11.3 keyring/23.5.0 rfc3986/2.0.0 colorama/0.4.4 CPython/3.9.11
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 673d732c0a85e853f7f6eddddc9afcde024ae10684c0a3c8dfea61723daa5971 |
|
MD5 | 04ff8014c1d9b33685e83c9873398cd0 |
|
BLAKE2b-256 | 36c836e0b8680d47611db50f30cd6a8cc42d2af59703380835be87ca0371ac56 |
Provenance
File details
Details for the file UtilsRL-0.3.5-py3-none-any.whl
.
File metadata
- Download URL: UtilsRL-0.3.5-py3-none-any.whl
- Upload date:
- Size: 20.4 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/3.8.0 pkginfo/1.8.2 readme-renderer/34.0 requests/2.27.1 requests-toolbelt/0.9.1 urllib3/1.26.9 tqdm/4.63.0 importlib-metadata/4.11.3 keyring/23.5.0 rfc3986/2.0.0 colorama/0.4.4 CPython/3.9.11
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 37808214839ed1904d5f6a7b6bb4858d56701316416820584f2d4d8b16793cfa |
|
MD5 | ff5c93b0809da66e281aff8a43aff3c1 |
|
BLAKE2b-256 | 948b35d425e56ed48610872c615c00f1821fbf15620c41699325edc5800df86c |