UtilsRL 0.4.0

A python module desgined for RL logging, monitoring and experiments managing.

UtilsRL

UtilsRL is a reinforcement learning utility python package, which is designed for fast integration into other RL projects. Despite its lightweightness, it still provides a full set of functions needed for RL algorithms development.

Currently UtilsRL is maintained by researchers from LAMDA-RL group. Any bug report / feature request / improvement is appreciated. We additionally thank @YuRuiii, @cmj2020, @paperplane03 and @momanto for their participation in code testing and performance benchmarking.

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. RL Modules

See sub-modules in UtilsRL/rl for details. We provide

• Actors with different output head, such as gaussian distribution, categorical distributions.
• Critics. For now we only implement single critic, but we have plans for implementing ensemble critics.
• Buffers, which replay transitions or trjectories.
• Normalizers, which normalizes data in multiple modes.

2. Experiment Managing

We provide a handful of functions for experiment management. All of them are placed under UtilsRL.exp, including:

2.1 Argument Parsing

The argument parsing utils in this package provides three features:

1. Supporting for multiple types of config files. parse_args can parse json, yaml, or even a python config module which is imported ahead

   from 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")
   

2. Nested argument parsing. We do this by introducing the NameSpace class. To be specific, if you pass convert=True to parse_args, then all of the dicts in the config file (including the argument dict itself) will be converted to a subclass of NameSpace. The contents wrapped in NameSpace 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 in main.py:

   import config_module
   args = parse_args(config_module)
   print(args)
   print(">>>>>>>>>>>>>>>>>>>>>>")
   print(args.TrainerArgs.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
   

3. 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, like python 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)
   

2.2 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

2.3 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.

3. 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

4. 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_scalar(tag="var_name", value=1.0, step=1)
logger.log_scalars(main_tag="group_name", tag_scalar_dict={
    "var1": 1.0, 
    "var2": 2.0
}, step=1)