mirror of
https://github.com/wassname/pytorch-soft-actor-critic.git
synced 2026-06-27 18:06:10 +08:00
wassname
185 lines
7.7 KiB
Python
185 lines
7.7 KiB
Python
import argparse
|
||
import datetime
|
||
import gym
|
||
import numpy as np
|
||
import itertools
|
||
from pathlib import Path
|
||
import torch
|
||
from sac import SAC
|
||
from torch.utils.tensorboard import SummaryWriter
|
||
from replay_memory import ReplayMemory
|
||
from load_demonstrations import load_demonstrations
|
||
import apple_gym.env
|
||
import pickle
|
||
from tqdm.auto import tqdm
|
||
|
||
parser = argparse.ArgumentParser(description='PyTorch Soft Actor-Critic Args')
|
||
parser.add_argument('-e', '--env-name', default="ApplePick-v0",
|
||
help='Mujoco Gym environment (default: ApplePick-v0)')
|
||
parser.add_argument('--policy', default="Gaussian",
|
||
help='Policy Type: Gaussian | Deterministic (default: Gaussian)')
|
||
parser.add_argument('--eval', type=bool, default=True,
|
||
help='Evaluates a policy a policy every 10 episode (default: True)')
|
||
parser.add_argument('--gamma', type=float, default=0.99, metavar='G',
|
||
help='discount factor for reward (default: 0.99)')
|
||
parser.add_argument('--tau', type=float, default=0.005, metavar='G',
|
||
help='target smoothing coefficient(τ) (default: 0.005)')
|
||
parser.add_argument('--lr', type=float, default=0.0003, metavar='G',
|
||
help='learning rate (default: 0.0003)')
|
||
parser.add_argument('--alpha', type=float, default=0.2, metavar='G',
|
||
help='Temperature parameter α determines the relative importance of the entropy\
|
||
term against the reward (default: 0.2)')
|
||
parser.add_argument('--automatic_entropy_tuning', type=bool, default=True, metavar='G',
|
||
help='Automaically adjust α (default: True)')
|
||
parser.add_argument('--seed', type=int, default=123456, metavar='N',
|
||
help='random seed (default: 123456)')
|
||
parser.add_argument('--batch_size', type=int, default=256, metavar='N',
|
||
help='batch size (default: 256)')
|
||
parser.add_argument('--num_steps', type=int, default=1000001, metavar='N',
|
||
help='maximum number of steps (default: 1000000)')
|
||
parser.add_argument('--hidden_size', type=int, default=256, metavar='N',
|
||
help='hidden size (default: 256)')
|
||
parser.add_argument('--updates_per_step', type=int, default=1, metavar='N',
|
||
help='model updates per simulator step (default: 1)')
|
||
parser.add_argument('--start_steps', type=int, default=10000, metavar='N',
|
||
help='Steps sampling random actions (default: 10000)')
|
||
parser.add_argument('--target_update_interval', type=int, default=1, metavar='N',
|
||
help='Value target update per no. of updates per step (default: 1)')
|
||
parser.add_argument('--replay_size', type=int, default=1000000, metavar='N',
|
||
help='size of replay buffer (default: 10000000)')
|
||
parser.add_argument('--cuda', action="store_true",
|
||
help='run on CUDA (default: False)')
|
||
parser.add_argument('--demonstrations', default=False,
|
||
help='Load demonstrations from https://github.com/erfanMhi/gym-recording-modified')
|
||
parser.add_argument('-l', '--load', default=False,
|
||
help='Load models')
|
||
parser.add_argument('-r', '--render', action="store_true",
|
||
help='show')
|
||
args = parser.parse_args()
|
||
|
||
# Environment
|
||
# env = NormalizedActions(gym.make(args.env_name))
|
||
env = gym.make(args.env_name, render=args.render)
|
||
env.seed(args.seed)
|
||
env.action_space.seed(args.seed)
|
||
|
||
torch.manual_seed(args.seed)
|
||
np.random.seed(args.seed)
|
||
|
||
# Agent
|
||
agent = SAC(env.observation_space.shape[0], env.action_space, args)
|
||
|
||
#Tensorboard
|
||
log_name = '{}_SAC_{}_{}_{}'.format(datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S"), args.env_name,
|
||
args.policy, "autotune" if args.automatic_entropy_tuning else "")
|
||
writer = SummaryWriter('runs/' + log_name)
|
||
|
||
save_dir=Path("models") / log_name
|
||
|
||
# Memory
|
||
memory=ReplayMemory(args.replay_size, args.seed)
|
||
|
||
|
||
def save(save_dir):
|
||
save_dir.mkdir(exist_ok=True)
|
||
agent.save_model(save_dir/'actor.pkl', save_dir/'critic.pkl')
|
||
memory.save(save_dir/'memory.pkl')
|
||
|
||
def load(save_dir):
|
||
agent.load_model(save_dir/'actor.pkl', save_dir/'critic.pkl')
|
||
memory.load(save_dir/'memory.pkl')
|
||
|
||
if args.load:
|
||
load(args.load)
|
||
|
||
if args.demonstrations:
|
||
load_demonstrations(memory, args.demonstrations)
|
||
|
||
# Training Loop
|
||
total_numsteps = 0
|
||
updates = 0
|
||
|
||
with tqdm(unit='steps', mininterval=5) as prog:
|
||
for i_episode in itertools.count(1):
|
||
episode_reward = 0
|
||
episode_steps = 0
|
||
done = False
|
||
state = env.reset()
|
||
|
||
for i_step in itertools.count(1):
|
||
if args.start_steps > total_numsteps:
|
||
action = env.action_space.sample() # Sample random action
|
||
else:
|
||
action = agent.select_action(state) # Sample action from policy
|
||
|
||
if len(memory) > args.batch_size:
|
||
# Number of updates per step in environment
|
||
for i in range(args.updates_per_step):
|
||
# Update parameters of all the networks
|
||
critic_1_loss, critic_2_loss, policy_loss, ent_loss, alpha = agent.update_parameters(memory, args.batch_size, updates)
|
||
|
||
writer.add_scalar('loss/critic_1', critic_1_loss, updates)
|
||
writer.add_scalar('loss/critic_2', critic_2_loss, updates)
|
||
writer.add_scalar('loss/policy', policy_loss, updates)
|
||
writer.add_scalar('loss/entropy_loss', ent_loss, updates)
|
||
writer.add_scalar('entropy_temperature/alpha', alpha, updates)
|
||
|
||
updates += 1
|
||
|
||
next_state, reward, done, info = env.step(action) # Step
|
||
episode_steps += 1
|
||
total_numsteps += 1
|
||
prog.update(1)
|
||
episode_reward += reward
|
||
|
||
# log env stuff
|
||
for k in ['env_reward/apple_pick/tree/min_fruit_dist_reward',
|
||
'env_reward/apple_pick/tree/gripping_fruit_reward',
|
||
'env_reward/apple_pick/tree/force_tree_reward',
|
||
'env_reward/apple_pick/tree/force_fruit_reward']:
|
||
writer.add_scalar(k, info[k], episode_steps)
|
||
|
||
# Ignore the "done" signal if it comes from hitting the time horizon. (that is, when it's an artificial terminal signal that isn't based on the agent's state)
|
||
# (https://github.com/openai/spinningup/blob/master/spinup/algos/pytorch/sac/sac.py)
|
||
mask = 1 if episode_steps == env._max_episode_steps else float(not done)
|
||
|
||
memory.push(state, action, reward, next_state, mask) # Append transition to memory
|
||
|
||
state = next_state
|
||
|
||
if total_numsteps > args.num_steps:
|
||
break
|
||
|
||
writer.add_scalar('reward/train', episode_reward, i_episode)
|
||
print("Episode: {}, total numsteps: {}, episode steps: {}, reward: {}".format(i_episode, total_numsteps, episode_steps, round(episode_reward, 2)))
|
||
|
||
if i_episode % 10 == 0 and args.eval is True:
|
||
avg_reward = 0.
|
||
episodes = 10
|
||
for _ in range(episodes):
|
||
state = env.reset()
|
||
episode_reward = 0
|
||
done = False
|
||
while not done:
|
||
action = agent.select_action(state, evaluate=True)
|
||
|
||
next_state, reward, done, _ = env.step(action)
|
||
episode_reward += reward
|
||
|
||
|
||
state = next_state
|
||
avg_reward += episode_reward
|
||
avg_reward /= episodes
|
||
|
||
|
||
writer.add_scalar('avg_reward/test', avg_reward, i_episode)
|
||
|
||
save(save_dir)
|
||
|
||
print("----------------------------------------")
|
||
print("Test Episodes: {}, Avg. Reward: {}".format(episodes, round(avg_reward, 2)))
|
||
print("----------------------------------------")
|
||
|
||
env.close()
|
||
save(save_dir)
|