kingspp/curious_agent

Pytorch implementation of large-scale-curiosity

Curious Agent

large-scale-curiosity implementation and improvements-

Paper
Github

Requirements

1. Python 3.7
2. PyTorch
3. Open AI

# Linux
sudo apt-get install ffmpeg libav-tools

# Mac
brew install ffmpeg libav-tools

Guidelines

1. Do not use print function. Instead use

# Create logger object
import logging
logger = logging.getLogger(__name__)

# Log according to the needs.
logger.info("Hello World")

2. Implement the "train" function with the "Continuing" Idiomatic Restriction:

continuing = True
state = {}


if continuing:  # the training is being resumed from a halted state
    pass  # i_episode is reloaded as part of the state
else:  # the training is being started from scratch
    self.state.i_episode = 0  # initialize the current number of episodes


# **OK!**: the current episode index can be reset and the algorithm would work seamlessly
while self.state.i_episode < 100000:
    # .
    # .
    # .
    # more training logic
    # .
    # .
    # .
    self.state.i_episode += 1

# **NOT OK!**: the current episode index starts from 0 regardless
for self.state.i_episode in range(0, 100000):
    # .
    # .
    # .
    # more training logic
    # .
    # .
    # .
    pass

3. Use the state object when implementing the "train" function to store and retrieve state information

# **OK!**: use the **state** munch object to store and retrieve all state information that the algorithm needs
self.state.i_episode = self.state.i_episode + 1
self.state.eps -= self.state.eps_reduction_rate

# **NOT OK!**: define state variables outside of the **state** munch object.
self.i_episode = self.i_episode + 1
self.eps -= self.eps_reduction_rate