class A2C:
def __init__(self, config,create_env,create_train_env,create_agent):
self.config = config
# Creation of the Logger (that saves in tensorboard and CSV)
self.logger = TFLogger(log_dir=self.config["logdir"], hps=self.config)
self._create_env=create_env
self._create_train_env=create_train_env
self._create_agent=create_agent
#Creation of one env instance to get the dimensionnality of observations and number of actions
env = self._create_env(self.config["n_envs"], seed=0,env_name=self.config["env_name"])
self.n_actions = env.action_space.n
self.obs_dim = env.reset()[0]["frame"].size()[1]
del env
def run(self):
# Instantiate the learning model abd the baseline model
self.learning_model=AgentModel(self.obs_dim,self.n_actions,32)
self.critic_model=BaselineModel(self.obs_dim,32)
#We create a batcher dedicated to evaluation
model=copy.deepcopy(self.learning_model)
self.evaluation_batcher=EpisodeBatcher(
n_timesteps=self.config["max_episode_steps"],
n_slots=self.config["n_evaluation_episodes"],
create_agent=self._create_agent,
create_env=self._create_env,
env_args={
"n_envs": self.config["n_envs"],
"max_episode_steps": self.config["max_episode_steps"],
"env_name":self.config["env_name"]
},
agent_args={"n_actions": self.n_actions, "model": model},
n_threads=self.config["n_evaluation_threads"],
seeds=[self.config["env_seed"]+k*10 for k in range(self.config["n_evaluation_threads"])],
)
#Creation of the batcher for sampling complete pieces of trajectories (i.e Batcher)
#The batcher will sample n_threads*n_envs trajectories at each call
# To have a fast batcher, we have to configure it with n_timesteps=self.config["max_episode_steps"]
model=copy.deepcopy(self.learning_model)
self.train_batcher=Batcher(
n_timesteps=self.config["a2c_timesteps"],
n_slots=self.config["n_envs"]*self.config["n_threads"],
create_agent=self._create_agent,
create_env=self._create_train_env,
env_args={
"n_envs": self.config["n_envs"],
"max_episode_steps": self.config["max_episode_steps"],
"env_name":self.config["env_name"]
},
agent_args={"n_actions": self.n_actions, "model": model},
n_threads=self.config["n_threads"],
seeds=[self.config["env_seed"]+k*10 for k in range(self.config["n_threads"])],
)
#Creation of the optimizer
optimizer = torch.optim.Adam(nn.Sequential(self.learning_model,self.critic_model).parameters(), lr=self.config["lr"])
#Training Loop:
_start_time=time.time()
self.iteration=0
# #We launch the evaluation batcher (in deterministic mode)
n_episodes=self.config["n_evaluation_episodes"]
agent_info=DictTensor({"stochastic":torch.tensor([False]).repeat(n_episodes)})
self.evaluation_batcher.execute(n_episodes=n_episodes,agent_info=agent_info)
self.evaluation_iteration=self.iteration
#Initialize the training batcher such that agents will start to acqire pieces of episodes
self.train_batcher.update(self.learning_model.state_dict())
n_episodes=self.config["n_envs"]*self.config["n_threads"]
agent_info=DictTensor({"stochastic":torch.tensor([True]).repeat(n_episodes)})
self.train_batcher.reset(agent_info=agent_info)
while(time.time()-_start_time<self.config["time_limit"]):
#Call the batcher to get a sample of trajectories
#2) We get the pieces of episodes. Since the env is an infinite env, we will always receive a new piece of episode
self.train_batcher.execute()
trajectories=self.train_batcher.get(blocking=True)
#3) Now, we compute the loss
dt=self.get_loss(trajectories)
[self.logger.add_scalar(k,dt[k].item(),self.iteration) for k in dt.keys()]
# Computation of final loss
ld = self.config["critic_coef"] * dt["critic_loss"]
lr = self.config["a2c_coef"] * dt["a2c_loss"]
le = self.config["entropy_coef"] * dt["entropy_loss"]
floss = ld - le - lr
floss= floss/n_episodes*trajectories.n_elems()
optimizer.zero_grad()
floss.backward()
optimizer.step()
#Update the train batcher with the updated model
self.train_batcher.update(self.learning_model.state_dict())
self.iteration+=1
#We check the evaluation batcher
evaluation_trajectories=self.evaluation_batcher.get(blocking=False)
if not evaluation_trajectories is None: #trajectories are available
#Compute the cumulated reward
cumulated_reward=(evaluation_trajectories["_reward"]*evaluation_trajectories.mask()).sum(1).mean()
self.logger.add_scalar("evaluation_reward",cumulated_reward.item(),self.evaluation_iteration)
print("At iteration %d, reward is %f"%(self.evaluation_iteration,cumulated_reward.item()))
#We reexecute the evaluation batcher (with same value of agent_info and same number of episodes)
self.evaluation_batcher.update(self.learning_model.state_dict())
self.evaluation_iteration=self.iteration
self.evaluation_batcher.reexecute()
self.train_batcher.close()
self.evaluation_batcher.get() # To wait for the last trajectories
self.evaluation_batcher.close()
self.logger.update_csv() # To save as a CSV file in logdir
self.logger.close()
def get_loss(self,trajectories):
#First, we want to compute the cumulated reward per trajectory
#The reward is a t+1 in each iteration (since it is btained after the aaction), so we use the '_reward' field in the trajectory
# The 'reward' field corresopnds to the reward at time t
reward=trajectories["_reward"]
#We get the mask that tells which transition is in a trajectory (1) or not (0)
mask=trajectories.mask()
#We remove the reward values that are not in the trajectories
reward=reward*mask
max_length=trajectories.lengths.max().item()
#Now, we want to compute the action probabilities over the trajectories such that we will be able to do 'backward'
action_probabilities=[]
for t in range(max_length):
proba=self.learning_model(trajectories["frame"][:,t])
action_probabilities.append(proba.unsqueeze(1)) # We append the probability, and introduces the temporal dimension (2nde dimension)
action_probabilities=torch.cat(action_probabilities,dim=1) #Now, we have a B x T x n_actions tensor
#We compute the critic value for t=0 to T (i.e including the very last observation)
critic=[]
for t in range(max_length):
b=self.critic_model(trajectories["frame"][:,t])
critic.append(b.unsqueeze(1))
critic=torch.cat(critic+[b.unsqueeze(1)],dim=1).squeeze(-1) #Now, we have a B x (T+1) tensor
#We also need to compute the critic value at for the last observation of the trajectories (to compute the TD)
# It may be the last element of the trajectories (if episode is not finished), or on the last frame of the episode
idx=torch.arange(trajectories.n_elems())
last_critic=self.critic_model(trajectories["_frame"][idx,trajectories.lengths-1]).squeeze(-1)
critic[idx,trajectories.lengths]=last_critic
#We compute the temporal difference
target=reward+self.config["discount_factor"]*(1-trajectories["_done"].float())*critic[:,1:].detach()
td=critic[:,:-1]-target
critic_loss=td**2
#We sum the loss for each episode (considering the mask)
critic_loss= (critic_loss*mask).sum(1)/mask.sum(1)
#We average the loss over all the trajectories
avg_critic_loss = critic_loss.mean()
#We do the same on the reinforce loss
action_distribution=torch.distributions.Categorical(action_probabilities)
log_proba=action_distribution.log_prob(trajectories["action"])
a2c_loss = -log_proba * td.detach()
a2c_loss = (a2c_loss*mask).sum(1)/mask.sum(1)
avg_a2c_loss=a2c_loss.mean()
#We compute the entropy loss
entropy=action_distribution.entropy()
entropy=(entropy*mask).sum(1)/mask.sum(1)
avg_entropy=entropy.mean()
return DictTensor({"critic_loss":avg_critic_loss,"a2c_loss":avg_a2c_loss,"entropy_loss":avg_entropy})
class SAC(BaseExperiment):
def __init__(self, config, create_env, create_agent):
super().__init__(config,create_env,create_agent)
env = self._create_env(
self.config["n_envs"], seed=0,**{k:self.config[k] for k in self.config if k.startswith("environment/")}
)
self.action_dim = env.action_space.sample().shape[0]
self.obs_dim = env.reset()[0]["frame"].size()[1]
del env
def check_arguments(self,args):
assert args["n_evaluation_rollouts"]%(args["n_envs"]*args["n_evaluation_threads"])==0
assert args["evaluation_mode"]=="deterministic" or args["evaluation_mode"]=="stochastic"
return True
def save(self):
super().save()
reward=self.evaluate(relaunch=False)
while(reward is None):
reward=self.evaluate(relaunch=False)
f=open(self.config["logdir"]+"/out.out","wb")
pickle.dump({"reward":reward},f)
time.sleep(np.random.rand()*10)
f.close()
def reset(self):
self.q1 = self._create_q()
self.q2 = self._create_q()
self.target_q1=self._create_q()
self.target_q2=self._create_q()
self.target_q1.load_state_dict(self.q1.state_dict())
self.target_q2.load_state_dict(self.q2.state_dict())
model=copy.deepcopy(self.learning_model)
self.train_batcher=Batcher(
n_timesteps=self.config["batch_timesteps"],
n_slots=self.config["n_envs"]*self.config["n_threads"],
create_agent=self._create_agent,
create_env=self._create_env,
env_args={
"mode":"train",
"n_envs": self.config["n_envs"],
"max_episode_steps": self.config["max_episode_steps"],
**{k:self.config[k] for k in self.config if k.startswith("environment/")}
},
agent_args={"action_dim": self.action_dim, "policy": model},
n_threads=self.config["n_threads"],
seeds=self.config["env_seed"],
)
model=copy.deepcopy(self.learning_model)
self.evaluation_batcher=EpisodeBatcher(
n_timesteps=self.config["max_episode_steps"],
n_slots=self.config["n_evaluation_rollouts"],
create_agent=self._create_agent,
create_env=self._create_env,
env_args={
"mode":"evaluation",
"max_episode_steps": self.config["max_episode_steps"],
"n_envs": self.config["n_envs"],
**{k:self.config[k] for k in self.config if k.startswith("environment/")}
},
agent_args={"action_dim": self.action_dim, "policy": model},
n_threads=self.config["n_evaluation_threads"],
seeds=self.config["env_seed"]*10,
)
self.register_batcher(self.train_batcher)
self.register_batcher(self.evaluation_batcher)
def _state_dict(self,model,device):
sd = model.state_dict()
for k, v in sd.items():
sd[k] = v.to(device)
return sd
def soft_update_params(self,net, target_net, tau):
for param, target_param in zip(net.parameters(), target_net.parameters()):
target_param.data.copy_(tau * param.data +(1 - tau) * target_param.data)
def run(self):
self.replay_buffer=ReplayBuffer(self.config["replay_buffer_size"])
device = torch.device(self.config["learner_device"])
self.learning_model.to(device)
self.q1.to(device)
self.q2.to(device)
self.target_q1.to(device)
self.target_q2.to(device)
optimizer = torch.optim.Adam(
self.learning_model.parameters(), lr=self.config["lr"]
)
optimizer_q1 = torch.optim.Adam(
self.q1.parameters(), lr=self.config["lr"]
)
optimizer_q2 = torch.optim.Adam(
self.q2.parameters(), lr=self.config["lr"]
)
self.train_batcher.update(self._state_dict(self.learning_model,torch.device("cpu")))
self.evaluation_batcher.update(self._state_dict(self.learning_model,torch.device("cpu")))
n_episodes=self.config["n_envs"]*self.config["n_threads"]
self.train_batcher.reset(agent_info=DictTensor({"stochastic":torch.zeros(n_episodes).eq(0.0)}))
logging.info("Sampling initial transitions")
n_iterations=int(self.config["n_starting_transitions"]/(n_episodes*self.config["batch_timesteps"]))
for k in range(n_iterations):
self.train_batcher.execute()
trajectories=self.train_batcher.get()
self.replay_buffer.push(trajectories)
print("replay_buffer_size = ",self.replay_buffer.size())
n_episodes=self.config["n_evaluation_rollouts"]
stochastic=torch.tensor([self.config["evaluation_mode"]=="stochastic"]).repeat(n_episodes)
self.evaluation_batcher.execute(agent_info=DictTensor({"stochastic":stochastic}), n_episodes=n_episodes)
logging.info("Starting Learning")
_start_time=time.time()
logging.info("Learning")
while time.time()-_start_time <self.config["time_limit"]:
self.train_batcher.execute()
trajectories=self.train_batcher.get()
self.replay_buffer.push(trajectories)
self.logger.add_scalar("replay_buffer_size",self.replay_buffer.size(),self.iteration)
# avg_reward = 0
for k in range(self.config["n_batches_per_epochs"]):
transitions=self.replay_buffer.sample(n=self.config["size_batches"])
#print(dt)
dt,transitions = self.get_q_loss(transitions,device)
[self.logger.add_scalar(k,dt[k].item(),self.iteration) for k in dt.keys()]
optimizer_q1.zero_grad()
dt["q1_loss"].backward()
optimizer_q1.step()
optimizer_q2.zero_grad()
dt["q2_loss"].backward()
optimizer_q2.step()
optimizer.zero_grad()
dt = self.get_policy_loss(transitions)
[self.logger.add_scalar(k,dt[k].item(),self.iteration) for k in dt.keys()]
dt["policy_loss"].backward()
optimizer.step()
tau=self.config["tau"]
self.soft_update_params(self.q1,self.target_q1,tau)
self.soft_update_params(self.q2,self.target_q2,tau)
self.iteration+=1
self.train_batcher.update(self._state_dict(self.learning_model,torch.device("cpu")))
self.evaluate()
def evaluate(self,relaunch=True):
evaluation_trajectories = self.evaluation_batcher.get(blocking=False)
if (evaluation_trajectories is None):
return
avg_reward = (
(
evaluation_trajectories["_reward"]
* evaluation_trajectories.mask()
)
.sum(1)
.mean()
.item()
)
self.logger.add_scalar("avg_reward/"+self.config["evaluation_mode"], avg_reward, self.iteration)
if (self.config["verbose"]):
print("Iteration "+str(self.iteration)+", Reward = "+str(avg_reward))
if (relaunch):
cpu_parameters=self._state_dict(self.learning_model,torch.device("cpu"))
self.evaluation_batcher.update(cpu_parameters)
self.evaluation_batcher.reexecute()
return avg_reward
def get_q_loss(self, transitions,device):
transitions = transitions.to(device)
B=transitions.n_elems()
Bv=torch.arange(B)
action = transitions["action"]
reward = transitions["_reward"]
frame = transitions["frame"]
_frame = transitions["_frame"]
_done = transitions["_done"].float()
# action for s_prime
mean_prime,var_prime=self.learning_model(_frame)
_id = torch.eye(self.action_dim).unsqueeze(0).repeat(B, 1, 1)
# _nvar = var_prime.unsqueeze(-1).repeat(1, 1, self.action_dim)
# _nvar = _nvar * _id
distribution=torch.distributions.Normal(mean_prime, var_prime)
next_action=distribution.sample().detach()
#Compute targets
q1=self.target_q1(_frame,next_action).detach().squeeze(-1)
q2=self.target_q2(_frame,next_action).detach().squeeze(-1)
q = torch.min(q1,q2)
lp= distribution.log_prob(next_action).detach().sum(-1)
q = q - self.config["lambda_entropy"]*lp
target_value=q*(1.-_done)*self.config["discount_factor"]+reward
q1_loss=(target_value.detach()-self.q1(frame,action).squeeze(-1))**2
q2_loss=(target_value.detach()-self.q2(frame,action).squeeze(-1))**2
dt ={
"q1_loss": q1_loss.mean(),
"q2_loss": q2_loss.mean(),
}
return DictTensor(dt),transitions
def get_policy_loss(self,transitions):
frame = transitions["frame"]
B=transitions.n_elems()
#Now, compute the policy term
mean,var=self.learning_model(frame)
#print(var.mean().item())
#print(mean)
_id = torch.eye(self.action_dim).unsqueeze(0).repeat(B, 1, 1)
# _nvar = var.unsqueeze(-1).repeat(1, 1, self.action_dim)
# _nvar = _nvar * _id
distribution=torch.distributions.Normal(mean, var)
entropy=distribution.entropy().mean()
action_tilde=distribution.rsample()
#print(action_tilde)
q1 = self.q1(frame,action_tilde).squeeze(-1)
q2 = self.q2(frame,action_tilde).squeeze(-1)
q=torch.min(q1,q2)
loss=q-self.config["lambda_entropy"]*distribution.log_prob(action_tilde).sum(-1)
dt={"policy_loss":-loss.mean(),"entropy":entropy.detach(),"avg_var":var.mean().detach(),"avg_mean":mean.mean().detach()}
dt=DictTensor(dt)
return dt
class PPO(BaseExperiment):
def __init__(self, config, create_env, create_agent):
super().__init__(config,create_env,create_agent)
env = self._create_env(
self.config["n_envs"], seed=0,**{k:self.config[k] for k in self.config if k.startswith("environment/")}
)
self.n_actions = env.action_space.n
self.obs_dim = env.reset()[0]["frame"].size()[1]
del env
def check_arguments(self,args):
assert args["n_evaluation_rollouts"]%(args["n_envs"]*args["n_evaluation_threads"])==0
assert args["evaluation_mode"]=="deterministic" or args["evaluation_mode"]=="stochastic"
return True
def reset(self):
#Creation of the batchers
model=copy.deepcopy(self.learning_model)
print(self.config)
self.train_batcher=Batcher(
n_timesteps=self.config["learning_timesteps"],
n_slots=self.config["n_envs"]*self.config["n_threads"],
create_agent=self._create_agent,
create_env=self._create_env,
env_args={
"mode":"train",
"n_envs": self.config["n_envs"],
"max_episode_steps": self.config["max_episode_steps"],
**{k:self.config[k] for k in self.config if k.startswith("environment/")}
},
agent_args={"n_actions": self.n_actions, "model": model},
n_threads=self.config["n_threads"],
seeds=self.config["env_seed"],
)
model=copy.deepcopy(self.learning_model)
self.evaluation_batcher=EpisodeBatcher(
n_timesteps=self.config["max_episode_steps"],
n_slots=self.config["n_evaluation_rollouts"],
create_agent=self._create_agent,
create_env=self._create_env,
env_args={
"mode":"evaluation",
"max_episode_steps": self.config["max_episode_steps"],
"n_envs": self.config["n_envs"],
**{k:self.config[k] for k in self.config if k.startswith("environment/")}
},
agent_args={"n_actions": self.n_actions, "model": model},
n_threads=self.config["n_evaluation_threads"],
seeds=self.config["env_seed"]*10,
)
self.register_batcher(self.train_batcher)
self.register_batcher(self.evaluation_batcher)
def _state_dict(self,model,device):
sd = model.state_dict()
for k, v in sd.items():
sd[k] = v.to(device)
return sd
def run(self):
device = torch.device(self.config["learner_device"])
self.learning_model.to(device)
optimizer = torch.optim.Adam(
self.learning_model.parameters(), lr=self.config["lr"]
)
cpu_parameters=self._state_dict(self.learning_model,torch.device("cpu"))
self.train_batcher.update(cpu_parameters)
self.evaluation_batcher.update(cpu_parameters)
n_episodes=self.config["n_evaluation_rollouts"]
self.evaluation_batcher.execute(agent_info=DictTensor({"stochastic":torch.ones(n_episodes)}), n_episodes=n_episodes)
# Initialize the train batcher
n_episodes=self.config["n_envs"]*self.config["n_threads"]
self.train_batcher.reset(agent_info=DictTensor({"stochastic":torch.ones(n_episodes)}))
_start_time=time.time()
while time.time()-_start_time<self.config["time_limit"]:
self.train_batcher.execute()
trajectories=self.train_batcher.get()
avg_reward = 0
for K in range(self.config["k_epochs"]):
optimizer.zero_grad()
dt = self.get_loss(trajectories)
[
self.logger.add_scalar("loss/" + k, dt[k].item(), self.iteration)
for k in dt.keys()
]
# Computation of final loss
ld = self.config["coef_critic"] * dt["value_loss"]
lr = self.config["coef_ppo"] * dt["ppo_loss"]
le = self.config["coef_entropy"] * dt["entropy_loss"]
floss = ld - le - lr
floss.backward()
if self.config["clip_grad"] > 0:
n = torch.nn.utils.clip_grad_norm_(
self.learning_model.parameters(), self.config["clip_grad"]
)
self.logger.add_scalar("grad_norm", n.item(), self.iteration)
optimizer.step()
self.evaluate()
self.iteration+=1
cpu_parameters=self._state_dict(self.learning_model,torch.device("cpu"))
self.train_batcher.update(cpu_parameters)
self.evaluate()
self.iteration+=1
def evaluate(self,relaunch=True):
evaluation_trajectories = self.evaluation_batcher.get(blocking=False)
if (evaluation_trajectories is None):
return
avg_reward = (
(
evaluation_trajectories["_reward"]
* evaluation_trajectories.mask()
)
.sum(1)
.mean()
.item()
)
self.logger.add_scalar("avg_reward/"+self.config["evaluation_mode"], avg_reward, self.iteration)
if (self.config["verbose"]):
print("Iteration "+str(self.iteration)+", Reward = "+str(avg_reward))
if (relaunch):
cpu_parameters=self._state_dict(self.learning_model,torch.device("cpu"))
self.evaluation_batcher.update(cpu_parameters)
self.evaluation_batcher.reexecute()
return avg_reward
def get_loss(self, trajectories):
device=self.config["learner_device"]
trajectories = trajectories.to(device)
max_length = trajectories.lengths.max().item()
assert trajectories.lengths.eq(max_length).all()
actions = trajectories["action"]
actions_probabilities = trajectories["action_probabilities"]
reward = trajectories["_reward"]
frame = trajectories["frame"]
last_action = trajectories["last_action"]
done = trajectories["_done"].float()
# Re compute model on trajectories
n_action_scores = []
n_values = []
hidden_state = trajectories["agent_state"][:, 0]
for T in range(max_length):
hidden_state = masked_tensor(hidden_state,trajectories["agent_state"][:, T],trajectories["initial_state"][:, T])
_as, _v, hidden_state = self.learning_model(
hidden_state, frame[:, T], last_action[:, T]
)
n_action_scores.append(_as.unsqueeze(1))
n_values.append(_v.unsqueeze(1))
n_action_scores = torch.cat(n_action_scores, dim=1)
n_values = torch.cat(
[*n_values, torch.zeros(trajectories.n_elems(), 1, 1).to(device)], dim=1
).squeeze(-1)
# Compute value function for last state
_idx = torch.arange(trajectories.n_elems()).to(device)
_hidden_state = hidden_state.detach() #trajectories["_agent_state"][_idx, trajectories.lengths - 1]
_frame = trajectories["_frame"][_idx, trajectories.lengths - 1]
_last_action = trajectories["_last_action"][_idx, trajectories.lengths - 1]
_, _v, _ = self.learning_model(_hidden_state, _frame, _last_action)
n_values[_idx, trajectories.lengths] = _v.squeeze(-1)
advantage = self.get_gae(
trajectories,
n_values,
discount_factor=self.config["discount_factor"],
_lambda=self.config["gae_lambda"],
)
value_loss = advantage ** 2
avg_value_loss = value_loss.mean()
n_action_probabilities = torch.softmax(n_action_scores, dim=2)
n_action_distribution = torch.distributions.Categorical(n_action_probabilities)
log_a=torch.distributions.Categorical(actions_probabilities).log_prob(actions)
log_na=n_action_distribution.log_prob(actions)
ratios=torch.exp(log_na-log_a)
surr1 = ratios * advantage
surr2 = torch.clamp(ratios,1-self.config["eps_clip"],1-self.config["eps_clip"])*advantage
ppo_loss = torch.min(surr1,surr2)
avg_ppo_loss = ppo_loss.mean()
entropy_loss = n_action_distribution.entropy()
avg_entropy_loss = entropy_loss.mean()
dt = DictTensor(
{
"entropy_loss": avg_entropy_loss,
"ppo_loss": avg_ppo_loss,
"value_loss": avg_value_loss,
}
)
return dt
def get_gae(self, trajectories, values, discount_factor=1, _lambda=0):
r = trajectories["_reward"]
v = values[:, 1:].detach()
d = trajectories["_done"].float()
delta = r + discount_factor * v * (1.0 - d) - values[:, :-1]
T = trajectories.lengths.max().item()
gae = delta[:, -1]
gaes = [gae]
for t in range(T - 2, -1, -1):
gae = delta[:, t] + discount_factor * _lambda * (1 - d[:, t]) * gae
gaes.append(gae)
gaes = list([g.unsqueeze(-1) for g in reversed(gaes)])
fgae = torch.cat(gaes, dim=1)
return fgae
class Reinforce:
def __init__(self, config, create_env, create_agent):
self.config = config
# Creation of the Logger (that saves in tensorboard and CSV)
self.logger = TFLogger(log_dir=self.config["logdir"], hps=self.config)
self._create_env = create_env
self._create_agent = create_agent
#Creation of one env instance to get the dimensionnality of observations and number of actions
env = self._create_env(self.config["n_envs"],
seed=0,
env_name=self.config["env_name"])
self.n_actions = env.action_space.n
self.obs_dim = env.reset()[0]["frame"].size()[1]
del env
def run(self):
# Instantiate the learning model abd the baseline model
self.learning_model = AgentModel(self.obs_dim, self.n_actions, 16)
self.baseline_model = BaselineModel(self.obs_dim, 16)
#We create a batcher dedicated to evaluation
model = copy.deepcopy(self.learning_model)
self.evaluation_batcher = EpisodeBatcher(
n_timesteps=self.config["max_episode_steps"],
n_slots=self.config["n_evaluation_episodes"],
create_agent=self._create_agent,
create_env=self._create_env,
env_args={
"n_envs": self.config["n_envs"],
"max_episode_steps": self.config["max_episode_steps"],
"env_name": self.config["env_name"]
},
agent_args={
"n_actions": self.n_actions,
"model": model
},
n_threads=self.config["n_evaluation_threads"],
seeds=[
self.config["env_seed"] + k * 10
for k in range(self.config["n_evaluation_threads"])
],
)
#Creation of the batcher for sampling complete episodes (i.e Episode Batcher)
#The batcher will sample n_threads*n_envs trajectories at each call
# To have a fast batcher, we have to configure it with n_timesteps=self.config["max_episode_steps"]
model = copy.deepcopy(self.learning_model)
self.train_batcher = EpisodeBatcher(
n_timesteps=self.config["max_episode_steps"],
n_slots=self.config["n_envs"] * self.config["n_threads"],
create_agent=self._create_agent,
create_env=self._create_env,
env_args={
"n_envs": self.config["n_envs"],
"max_episode_steps": self.config["max_episode_steps"],
"env_name": self.config["env_name"]
},
agent_args={
"n_actions": self.n_actions,
"model": model
},
n_threads=self.config["n_threads"],
seeds=[
self.config["env_seed"] + k * 10
for k in range(self.config["n_threads"])
],
)
#Creation of the optimizer
optimizer = torch.optim.Adam(nn.Sequential(
self.learning_model, self.baseline_model).parameters(),
lr=self.config["lr"])
#Training Loop:
_start_time = time.time()
self.iteration = 0
#We launch the evaluation batcher (in deterministic mode)
n_episodes = self.config["n_evaluation_episodes"]
agent_info = DictTensor(
{"stochastic": torch.tensor([False]).repeat(n_episodes)})
self.evaluation_batcher.execute(n_episodes=n_episodes,
agent_info=agent_info)
self.evaluation_iteration = self.iteration
while (time.time() - _start_time < self.config["time_limit"]):
#Update the batcher with the last version of the learning model
self.train_batcher.update(self.learning_model.state_dict())
#Call the batcher to get a sample of trajectories
#1) The policy will be executed in "stochastic' mode
n_episodes = self.config["n_envs"] * self.config["n_threads"]
agent_info = DictTensor(
{"stochastic": torch.tensor([True]).repeat(n_episodes)})
self.train_batcher.execute(n_episodes=n_episodes,
agent_info=agent_info)
#2) We get the trajectories (and wait until the trajectories have been sampled)
trajectories = self.train_batcher.get(blocking=True)
#3) Now, we compute the loss
dt = self.get_loss(trajectories)
[
self.logger.add_scalar(k, dt[k].item(), self.iteration)
for k in dt.keys()
]
# Computation of final loss
ld = self.config["baseline_coef"] * dt["baseline_loss"]
lr = self.config["reinforce_coef"] * dt["reinforce_loss"]
le = self.config["entropy_coef"] * dt["entropy_loss"]
floss = ld - le - lr
optimizer.zero_grad()
floss.backward()
optimizer.step()
#Update the train batcher with the updated model
self.train_batcher.update(self.learning_model.state_dict())
print("At iteration %d, avg (discounted) reward is %f" %
(self.iteration, dt["avg_reward"].item()))
print("\t Avg trajectory length is %f" %
(trajectories.lengths.float().mean().item()))
print(
"\t Curves can be visualized using 'tensorboard --logdir=%s'" %
self.config["logdir"])
self.iteration += 1
#We check the evaluation batcher
evaluation_trajectories = self.evaluation_batcher.get(
blocking=False)
if not evaluation_trajectories is None: #trajectories are available
#Compute the cumulated reward
cumulated_reward = (
evaluation_trajectories["_reward"] *
evaluation_trajectories.mask()).sum(1).mean()
self.logger.add_scalar("evaluation_reward",
cumulated_reward.item(),
self.evaluation_iteration)
#We reexecute the evaluation batcher (with same value of agent_info and same number of episodes)
self.evaluation_batcher.update(
self.learning_model.state_dict())
self.evaluation_iteration = self.iteration
self.evaluation_batcher.reexecute()
self.train_batcher.close()
self.evaluation_batcher.get() # To wait for the last trajectories
self.evaluation_batcher.close()
self.logger.update_csv() # To save as a CSV file in logdir
self.logger.close()
def get_loss(self, trajectories):
#First, we want to compute the cumulated reward per trajectory
#The reward is a t+1 in each iteration (since it is btained after the aaction), so we use the '_reward' field in the trajectory
# The 'reward' field corresopnds to the reward at time t
reward = trajectories["_reward"]
#We get the mask that tells which transition is in a trajectory (1) or not (0)
mask = trajectories.mask()
#We remove the reward values that are not in the trajectories
reward = reward * mask
#We compute the future cumulated reward at each timestep (by reverse computation)
max_length = trajectories.lengths.max().item()
cumulated_reward = torch.zeros_like(reward)
cumulated_reward[:, max_length - 1] = reward[:, max_length - 1]
for t in range(max_length - 2, -1, -1):
cumulated_reward[:, t] = reward[:, t] + self.config[
"discount_factor"] * cumulated_reward[:, t + 1]
#Now, we want to compute the action probabilities over the trajectories such that we will be able to do 'backward'
action_probabilities = []
for t in range(max_length):
proba = self.learning_model(trajectories["frame"][:, t])
action_probabilities.append(
proba.unsqueeze(1)
) # We append the probability, and introduces the temporal dimension (2nde dimension)
action_probabilities = torch.cat(
action_probabilities,
dim=1) #Now, we have a B x T x n_actions tensor
#We compute the baseline
baseline = []
for t in range(max_length):
b = self.baseline_model(trajectories["frame"][:, t])
baseline.append(b.unsqueeze(1))
baseline = torch.cat(baseline,
dim=1).squeeze(-1) #Now, we have a B x T tensor
#We compute the baseline loss
baseline_loss = (baseline - cumulated_reward)**2
#We sum the loss for each episode (considering the mask)
baseline_loss = (baseline_loss * mask).sum(1) / mask.sum(1)
#We average the loss over all the trajectories
avg_baseline_loss = baseline_loss.mean()
#We do the same on the reinforce loss
action_distribution = torch.distributions.Categorical(
action_probabilities)
log_proba = action_distribution.log_prob(trajectories["action"])
reinforce_loss = log_proba * (cumulated_reward - baseline).detach()
reinforce_loss = (reinforce_loss * mask).sum(1) / mask.sum(1)
avg_reinforce_loss = reinforce_loss.mean()
#We compute the entropy loss
entropy = action_distribution.entropy()
entropy = (entropy * mask).sum(1) / mask.sum(1)
avg_entropy = entropy.mean()
return DictTensor({
"avg_reward": cumulated_reward[:, 0].mean(),
"baseline_loss": avg_baseline_loss,
"reinforce_loss": avg_reinforce_loss,
"entropy_loss": avg_entropy
})
class DQN(BaseExperiment):
def __init__(self, config, create_env, create_agent):
super().__init__(config,create_env,create_agent)
env = self._create_env(
self.config["n_envs"], seed=0,**{k:self.config[k] for k in self.config if k.startswith("environment/")}
)
self.n_actions = env.action_space.n
self.obs_dim = env.reset()[0]["frame"].size()[1]
del env
def check_arguments(self,args):
assert args["n_evaluation_rollouts"]%(args["n_envs"]*args["n_evaluation_threads"])==0
return True
def reset(self):
self.target_model=copy.deepcopy(self.learning_model)
model=copy.deepcopy(self.learning_model)
self.train_batcher=Batcher(
n_timesteps=self.config["batch_timesteps"],
n_slots=self.config["n_envs"]*self.config["n_threads"],
create_agent=self._create_agent,
create_env=self._create_env,
env_args={
"mode":"train",
"n_envs": self.config["n_envs"],
"max_episode_steps": self.config["max_episode_steps"],
**{k:self.config[k] for k in self.config if k.startswith("environment/")}
},
agent_args={"n_actions": self.n_actions, "model": model},
n_threads=self.config["n_threads"],
seeds=self.config["env_seed"],
)
model=copy.deepcopy(self.learning_model)
self.evaluation_batcher=EpisodeBatcher(
n_timesteps=self.config["max_episode_steps"],
n_slots=self.config["n_evaluation_rollouts"],
create_agent=self._create_agent,
create_env=self._create_env,
env_args={
"mode":"evaluation",
"max_episode_steps": self.config["max_episode_steps"],
"n_envs": self.config["n_envs"],
**{k:self.config[k] for k in self.config if k.startswith("environment/")}
},
agent_args={"n_actions": self.n_actions, "model": model},
n_threads=self.config["n_evaluation_threads"],
seeds=self.config["env_seed"]*10,
)
self.register_batcher(self.train_batcher)
self.register_batcher(self.evaluation_batcher)
def _state_dict(self,model,device):
sd = model.state_dict()
for k, v in sd.items():
sd[k] = v.to(device)
return sd
def run(self):
self.replay_buffer=ReplayBuffer(self.config["replay_buffer_size"])
device = torch.device(self.config["learner_device"])
self.learning_model.to(device)
self.target_model.to(device)
optimizer = torch.optim.Adam(
self.learning_model.parameters(), lr=self.config["lr"]
)
self.train_batcher.update(self._state_dict(self.learning_model,torch.device("cpu")))
self.evaluation_batcher.update(self._state_dict(self.learning_model,torch.device("cpu")))
n_episodes=self.config["n_envs"]*self.config["n_threads"]
self.train_batcher.reset(agent_info=DictTensor({"epsilon":torch.ones(n_episodes)*self.config["epsilon_greedy"]}))
logging.info("Sampling initial transitions")
for k in range(self.config["initial_buffer_epochs"]):
self.train_batcher.execute()
trajectories=self.train_batcher.get()
self.replay_buffer.push(trajectories)
n_episodes=self.config["n_evaluation_rollouts"]
self.evaluation_batcher.execute(agent_info=DictTensor({"epsilon":torch.zeros(n_episodes)}), n_episodes=n_episodes)
logging.info("Starting Learning")
_start_time=time.time()
logging.info("Learning")
while time.time()-_start_time <self.config["time_limit"]:
self.train_batcher.execute()
trajectories=self.train_batcher.get()
self.replay_buffer.push(trajectories)
self.logger.add_scalar("replay_buffer_size",self.replay_buffer.size(),self.iteration)
# avg_reward = 0
for k in range(self.config["qvalue_epochs"]):
optimizer.zero_grad()
dt = self.get_loss(device)
[self.logger.add_scalar(k,dt[k].item(),self.iteration) for k in dt.keys()]
floss=dt["q_loss"]
floss.backward()
if self.config["clip_grad"] > 0:
n = torch.nn.utils.clip_grad_norm_(
self.learning_model.parameters(), self.config["clip_grad"]
)
self.logger.add_scalar("grad_norm", n.item(), self.iteration)
self.iteration+=1
optimizer.step()
tau=self.config["tau"]
self.soft_update_params(self.learning_model,self.target_model,tau)
self.train_batcher.update(self._state_dict(self.learning_model,torch.device("cpu")))
self.evaluate()
self.iteration+=1
def soft_update_params(self,net, target_net, tau):
for param, target_param in zip(net.parameters(), target_net.parameters()):
target_param.data.copy_(tau * param.data +(1 - tau) * target_param.data)
def evaluate(self,relaunch=True):
evaluation_trajectories = self.evaluation_batcher.get(blocking=False)
if (evaluation_trajectories is None):
return
avg_reward = (
(
evaluation_trajectories["_reward"]
* evaluation_trajectories.mask()
)
.sum(1)
.mean()
.item()
)
self.logger.add_scalar("avg_reward", avg_reward, self.iteration)
if (self.config["verbose"]):
print("Iteration "+str(self.iteration)+", Reward = "+str(avg_reward)+", Buffer size = "+str(self.replay_buffer.size()))
if (relaunch):
self.evaluation_batcher.update(self._state_dict(self.learning_model,torch.device("cpu")))
self.evaluation_batcher.reexecute()
return avg_reward
def get_loss(self, device):
transitions=self.replay_buffer.sample(n=self.config["n_batches"])
transitions = transitions.to(device)
B=transitions.n_elems()
Bv=torch.arange(B)
action = transitions["action"]
reward = transitions["_reward"]
frame = transitions["frame"]
_frame = transitions["_frame"]
_done = transitions["_done"].float()
q=self.learning_model(frame)
qa=q[Bv,action]
qp = self.learning_model(_frame)
actionp=qp.max(1)[1]
_q_target = self.target_model(_frame).detach()
_q_target_a= _q_target[Bv,actionp]
_target_value=_q_target_a*(1-_done)*self.config["discount_factor"]+reward
td = (_target_value-qa)**2
dt = DictTensor(
{
"q_loss": td.mean(),
}
)
return dt
batcher=EpisodeBatcher(
n_timesteps=100,
n_slots=128,
n_threads=4,
seeds=[1,2,3,4],
create_agent=create_agent,
agent_args={"n_actions":2},
create_env=create_env,
env_args={"max_episode_steps":100}
)
# Execution will start the acqusition process (such that some other computations can be done in parallel to the episodes acquisition)
n_episodes=32
# Since we will sample 32 episodes, we need to configure the 32 agents and 32 environments that will interact
agent_info=DictTensor({"agent_id":torch.arange(32)})
env_info=DictTensor({"env_id":torch.arange(32)})
#Running the batcher. It is a non-blocking function that launch the acqusition
batcher.execute(n_episodes=32,agent_info=agent_info,env_info=env_info)
#Getting episodes -- not that get is a blocking function such that the process will wait until the end of the acquisition.
trajectories=batcher.get()
#*get* can also be called in non-blocking mode *trajectories.get(blocking=Flase)* that wil return *None* if the acqusition process is not terminated
# the **reexecute** method is a shortcut to call *execute* again with the same arguments
batcher.reexecute()
trajectories=batcher.get()