saymrwulf/stable-baselines3
1
0
Fork 0
mirror of https://github.com/saymrwulf/stable-baselines3.git synced 2026-07-13 18:08:39 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Start cleanup + update docstrings

Browse source
This commit is contained in:
Antonin Raffin 2019-11-18 14:09:31 +01:00
parent 95c741c707
commit 5d353d598c
2 changed files with 43 additions and 11 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

25
tests/test_sde.py
View file

@ -1,17 +1,25 @@
import pytest
import gym
import torch as th
from torch.distributions import Normal
from torchy_baselines import A2C
from torchy_baselines.common.vec_env import DummyVecEnv, VecNormalize
from torchy_baselines.common.monitor import Monitor
def test_state_dependent_exploration():
"""
Check that the gradient correspond to the expected one
"""
n_states = 2
state_dim = 3
# TODO: fix for action_dim > 1
action_dim = 1
sigma = th.ones(state_dim, action_dim, requires_grad=True)
sigma = th.ones(state_dim, 1, requires_grad=True)
# Reduce the number of parameters
# sigma_ = th.ones(state_dim, action_dim) * sigma_
# weights_dist = Normal(th.zeros_like(log_sigma), th.exp(log_sigma))
th.manual_seed(2)
@ -42,16 +50,11 @@ def test_state_dependent_exploration():
@pytest.mark.parametrize("model_class", [A2C])
def test_state_dependent_noise(model_class):
import gym
from torchy_baselines.common.vec_env import DummyVecEnv, VecNormalize
from torchy_baselines.common.monitor import Monitor
# env_id = 'Pendulum-v0'
env_id = 'MountainCarContinuous-v0'
# env_id = 'LunarLanderContinuous-v2'
env = VecNormalize(DummyVecEnv([lambda: Monitor(gym.make(env_id))]), norm_reward=True)
eval_env = VecNormalize(DummyVecEnv([lambda: Monitor(gym.make(env_id))]), training=False, norm_reward=False)
model = model_class('MlpPolicy', env, n_steps=200, max_grad_norm=1, use_rms_prop=False,
use_sde=True, ent_coef=0.00, verbose=1, create_eval_env=True, learning_rate=3e-4,
policy_kwargs=dict(log_std_init=0.0, ortho_init=False, net_arch=[256, dict(pi=[256], vf=[256])]), seed=None)
model.learn(total_timesteps=int(20000), log_interval=5, eval_freq=10000, eval_env=eval_env)
model = model_class('MlpPolicy', env, n_steps=200, use_sde=True, ent_coef=0.00, verbose=1, learning_rate=3e-4,
policy_kwargs=dict(log_std_init=0.0, ortho_init=False), seed=None)
model.learn(total_timesteps=int(1000), log_interval=5, eval_freq=500, eval_env=eval_env)

29
torchy_baselines/common/distributions.py
View file

@ -45,6 +45,11 @@ class Distribution(object):
class DiagGaussianDistribution(Distribution):
"""
Gaussian distribution with diagonal covariance matrix.
:param action_dim: (int) Number of actions
"""
def __init__(self, action_dim):
super(DiagGaussianDistribution, self).__init__()
self.distribution = None
@ -53,12 +58,28 @@ class DiagGaussianDistribution(Distribution):
self.log_std = None
def proba_distribution_net(self, latent_dim, log_std_init=0.0):
"""
Create the layers and parameter that represent the distribution:
one output will be the mean of the gaussian, the other parameter will be the
standard deviation (log std in fact to allow negative values)
:param latent_dim: (int) Dimension og the last layer of the policy (before the action layer)
:param log_std_init: (float) Initial value for the log standard deviation
"""
mean_actions = nn.Linear(latent_dim, self.action_dim)
# TODO: allow action dependent std
log_std = nn.Parameter(th.ones(self.action_dim) * log_std_init)
return mean_actions, log_std
def proba_distribution(self, mean_actions, log_std, deterministic=False):
"""
Create and sample for the distribution given its parameters (mean, std)
:param mean_actions: (th.Tensor)
:param log_std: (th.Tensor)
:param deterministic: (bool)
:return: (th.Tensor)
"""
action_std = th.ones_like(mean_actions) * log_std.exp()
self.distribution = Normal(mean_actions, action_std)
if deterministic:
@ -77,6 +98,14 @@ class DiagGaussianDistribution(Distribution):
return self.distribution.entropy()
def log_prob_from_params(self, mean_actions, log_std):
"""
Compute the log probabilty of taking an action
given the distribution parameters.
:param mean_actions: (th.Tensor)
:param log_std: (th.Tensor)
:return: (th.Tensor, th.Tensor)
"""
action, _ = self.proba_distribution(mean_actions, log_std)
log_prob = self.log_prob(action)
return action, log_prob