def store_experience(self, state, action, nextState, reward, info):
if self.experienceProcessor is not None:
state, action, nextState, reward = self.experienceProcessor(
state, action, nextState, reward, info)
# caution: using multiple step forward return can increase variance
if self.nStepForward > 1:
# if this is final state, we want to do additional backup to increase useful learning experience
if nextState is None:
transitions = []
transitions.append(Transition(state, action, nextState,
reward))
R = reward
while len(self.nStepBuffer) > 0:
state, action, next_state, reward_old = self.nStepBuffer.pop(
0)
R = reward_old + self.gamma * R
transNew = Transition(state, action, None, R)
transitions.append(transNew)
for tran in transitions:
if self.priorityMemoryOption:
self.memory.store(tran)
else:
self.memory.push(tran)
else:
# otherwise we calculate normal n step return
self.nStepBuffer.append((state, action, nextState, reward))
if len(self.nStepBuffer) < self.nStepForward:
return
R = sum([
self.nStepBuffer[i][3] * (self.gamma**i)
for i in range(self.nStepForward)
])
state, action, _, _ = self.nStepBuffer.pop(0)
transition = Transition(state, action, nextState, R)
if self.priorityMemoryOption:
self.memory.store(transition)
else:
self.memory.push(transition)
else:
# if it is one step
transition = Transition(state, action, nextState, reward)
if self.priorityMemoryOption:
self.memory.store(transition)
else:
self.memory.push(transition)
def prepare_minibatch(self, state, action, nextState, reward, info):
# first store memory
self.store_experience(state, action, nextState, reward, info)
if len(self.memory) < self.trainBatchSize:
return
transitions_raw = self.memory.sample(self.trainBatchSize)
transitions = Transition(*zip(*transitions_raw))
action = torch.tensor(transitions.action,
device=self.device,
dtype=torch.float32) # shape(batch, numActions)
reward = torch.tensor(transitions.reward,
device=self.device,
dtype=torch.float32) # shape(batch)
# for some env, the output state requires further processing before feeding to neural network
if self.stateProcessor is not None:
state, _ = self.stateProcessor(transitions.state, self.device)
nonFinalNextState, nonFinalMask = self.stateProcessor(
transitions.next_state, self.device)
else:
state = torch.tensor(transitions.state,
device=self.device,
dtype=torch.float32)
nonFinalMask = torch.tensor(tuple(
map(lambda s: s is not None, transitions.next_state)),
device=self.device,
dtype=torch.uint8)
nonFinalNextState = torch.tensor(
[s for s in transitions.next_state if s is not None],
device=self.device,
dtype=torch.float32)
return state, nonFinalMask, nonFinalNextState, action, reward
def prepare_minibatch(self, transitions_raw):
'''
do some proprocessing work for transitions_raw
order the data
convert transition list to torch tensors
use trick from https://pytorch.org/tutorials/intermediate/reinforcement_q_learning.html
https://stackoverflow.com/questions/19339/transpose-unzip-function-inverse-of-zip/19343#19343
'''
transitions = Transition(*zip(*transitions_raw))
action = torch.tensor(transitions.action, device=self.device, dtype=torch.long).unsqueeze(-1) # shape(batch, 1)
reward = torch.tensor(transitions.reward, device=self.device, dtype=torch.float32).unsqueeze(-1) # shape(batch, 1)
# for some env, the output state requires further processing before feeding to neural network
if self.stateProcessor is not None:
state, _ = self.stateProcessor(transitions.state, self.device)
nonFinalNextState, nonFinalMask = self.stateProcessor(transitions.next_state, self.device)
else:
state = torch.tensor(transitions.state, device=self.device, dtype=torch.float32)
nonFinalMask = torch.tensor(tuple(map(lambda s: s is not None, transitions.next_state)), device=self.device,
dtype=torch.bool)
nonFinalNextState = torch.tensor([s for s in transitions.next_state if s is not None], device=self.device,
dtype=torch.float32)
return state, nonFinalMask, nonFinalNextState, action, reward
def sample(self, batch_size):
finish = random.sample(range(0, len(self.memory)), batch_size)
begin = [x - self.seq_length for x in finish]
samples = []
for start, end in zip(begin, finish):
# correct for sampling near beginning
# final is a list
final = self.memory[max(start + 1, 0):end + 1]
# correct for sampling across episodes
# remove experiences belongs to previous episode
for i in range(len(final) - 2, -1, -1):
if final[i][3] is None:
final = final[i + 1:]
break
# pad beginning to for sequence that end earlier
while (len(final) < self.seq_length):
dummyTransition = Transition(np.zeros_like(self.memory[0][0]), 0, np.zeros_like(self.memory[0][2]), 0)
final = [dummyTransition] + final
samples += final
# returns flattened version
return samples
def store_experience(self, states, actions, nextStates, reward, info):
transitions = [
Transition(states[n], actions[n], nextStates[n], reward)
for n in range(self.numAgents)
]
self.memory.push(transitions)
def prepare_minibatch(self, transitions_raw, n):
# first store memory
transitions = Transition(*zip(*transitions_raw))
action = torch.tensor(transitions.action,
device=self.device,
dtype=torch.long).unsqueeze(
-1) # shape(batch, 1)
reward = torch.tensor(transitions.reward,
device=self.device,
dtype=torch.float32).unsqueeze(
-1) # shape(batch, 1)
# for some env, the output state requires further processing before feeding to neural network
if self.stateProcessors is not None:
state, _ = self.stateProcessors[n](transitions.state, self.device)
nonFinalNextState, nonFinalMask = self.stateProcessors[n](
transitions.next_state, self.device)
else:
state = torch.tensor(transitions.state,
device=self.device,
dtype=torch.float32)
nonFinalMask = torch.tensor(tuple(
map(lambda s: s is not None, transitions.next_state)),
device=self.device,
dtype=torch.uint8)
nonFinalNextState = torch.tensor(
[s for s in transitions.next_state if s is not None],
device=self.device,
dtype=torch.float32)
return state, nonFinalMask, nonFinalNextState, action, reward
def process_experienceAugmentation(self, state, action, nextState, reward,
info):
if self.globalStepCount % self.experienceAugmentationFreq == 0:
state_Augs, action_Augs, nextState_Augs, reward_Augs = self.env.getExperienceAugmentation(
state, action, nextState, reward, info)
for i in range(len(state_Augs)):
transition = Transition(state_Augs[i], action_Augs[i],
nextState_Augs[i], reward_Augs[i])
self.memory.push(transition)
def store_experience(self, states, actions, nextStates, rewards, infos):
for i in range(len(states)):
# if it is ended due to stepLimit, we should not store the experience due to the vec env setup
if not infos[i]['endBeforeDone']:
transition = Transition(states[i], actions[i], nextStates[i], rewards[i])
self.memory.push(transition)
if self.successRepeat and nextStates[i] is None:
for _ in range(self.successRepeatTime):
self.memory.push(transition)
def process_hindSightExperience(self, state, action, nextState, reward,
info):
if nextState is not None and self.globalStepCount % self.hindSightERFreq == 0:
stateNew, actionNew, nextStateNew, rewardNew = self.env.getHindSightExperience(
state, action, nextState, info)
if stateNew is not None:
transition = Transition(stateNew, actionNew, nextStateNew,
rewardNew)
self.memory.push(transition)
if self.experienceAugmentation:
self.process_experienceAugmentation(
state, action, nextState, reward, info)
def store_experience(self, state, action, nextState, reward, info):
if self.experienceProcessor is not None:
state, action, nextState, reward = self.experienceProcessor(state, action, nextState, reward, info)
timeStep = state['timeStep']
transition = Transition(state, action, nextState, reward)
self.memories[timeStep].push(transition)
if self.experienceAugmentation:
self.process_experienceAugmentation(state, action, nextState, reward, info)
if self.hindSightER:
self.process_hindSightExperience(state, action, nextState, reward, info)
def push(self, *args):
"""Saves a transition"""
if len(args) == 1 and isinstance(*args, Transition):
transition = args[0]
else:
transition = Transition(*args)
# if it is terminal state
if transition.next_state is None:
if len(self.terminalMemory) < self.capacity:
self.terminalMemory.append(None)
self.terminalMemory[self.positionTwo] = transition
self.positionTwo = (self.positionTwo + 1) % self.capacity
count = 1
R = transition.reward
for trans in self.nStepBuffer[::-1]:
# if nstepBackup is zero, then this is no backup
if count > self.nStepBackup:
break
R = trans.reward + self.gamma * R
transNew = Transition(trans.state, trans.action, None, R)
if len(self.terminalMemory) < self.capacity:
self.terminalMemory.append(None)
self.terminalMemory[self.positionTwo] = transNew
self.positionTwo = (self.positionTwo + 1) % self.capacity
count += 1
self.nStepBuffer.clear()
else: # if it is terminal state
if len(self.memory) < self.capacity:
self.memory.append(None)
self.nStepBuffer.append(transition)
# write on the earlier experience
self.memory[self.positionOne] = transition
self.positionOne = (self.positionOne + 1) % self.capacity
def update_net(self, state, action, nextState, reward):
# first store memory
self.store_experience(state, action, nextState, reward)
if len(self.memory) < self.trainBatchSize:
return
transitions_raw = self.memory.sample(self.trainBatchSize)
transitions = Transition(*zip(*transitions_raw))
# for some env, the output state requires further processing before feeding to neural network
if self.stateProcessor is not None:
state = self.stateProcessor(transitions.state)
nextState = self.stateProcessor(transitions.next_state)
else:
state = torch.tensor(transitions.state, dtype=torch.float32)
nextState = torch.tensor(transitions.next_state, dtype=torch.float32)
action = torch.tensor(transitions.action, dtype=torch.long).unsqueeze(-1) # shape(batch, 1)
reward = torch.tensor(transitions.reward, dtype=torch.float32).unsqueeze(-1) # shape(batch, 1)
batchSize = reward.shape[0]
QValues = self.policyNet(state).gather(1, action)
# note that here use policyNet for target value
QNext = self.policyNet(nextState).detach()
targetValues = reward + self.gamma * QNext.max(dim=1)[0].unsqueeze(-1)
loss = torch.mean(self.netLossFunc(QValues, targetValues))
self.optimizer.zero_grad()
loss.backward()
# for lp, gp in zip(self.localNet.parameters(), self.globalNet.parameters()):
# gp._grad = lp._grad
#
# if self.netGradClip is not None:
# torch.nn.utils.clip_grad_norm_(self.policyNet.parameters(), self.netGradClip)
#
# # global net update
# self.globalOptimizer.step()
#
# # update local net
# self.localNet.load_state_dict(self.globalNet.state_dict())
if self.globalStepCount % self.lossRecordStep == 0:
self.losses.append([self.globalStepCount, self.epIdx, loss])
def update_net_and_sync(self, state, action, nextState, reward):
self.store_experience(state, action, nextState, reward)
if self.priorityMemoryOption:
if len(self.memory) < self.config['memoryCapacity']:
return
else:
if len(self.memory) < self.trainBatchSize:
return
if self.totalStep % self.updateGlobalFrequency == 0:
transitions_raw = self.memory.sample(self.trainBatchSize)
transitions = Transition(*zip(*transitions_raw))
action = torch.tensor(transitions.action, device=self.device, dtype=torch.long).unsqueeze(
-1) # shape(batch, 1)
reward = torch.tensor(transitions.reward, device=self.device, dtype=torch.float32).unsqueeze(
-1) # shape(batch, 1)
batchSize = reward.shape[0]
# for some env, the output state requires further processing before feeding to neural network
if self.stateProcessor is not None:
state, _ = self.stateProcessor(transitions.state, self.device)
nonFinalNextState, nonFinalMask = self.stateProcessor(transitions.next_state, self.device)
else:
state = torch.tensor(transitions.state, device=self.device, dtype=torch.float32)
nonFinalMask = torch.tensor([s is not None for s in transitions.next_state],
device=self.device, dtype=torch.uint8)
nonFinalNextState = torch.tensor([s for s in transitions.next_state if s is not None],
device=self.device, dtype=torch.float32)
if self.synchLock:
self.lock.acquire()
QValues = self.globalPolicyNet(state).gather(1, action)
if self.netUpdateOption == 'targetNet':
# Here we detach because we do not want gradient flow from target values to net parameters
QNext = torch.zeros(batchSize, device=self.device, dtype=torch.float32)
QNext[nonFinalMask] = self.globalTargetNet(nonFinalNextState).max(1)[0].detach()
targetValues = reward + self.gamma * QNext.unsqueeze(-1)
if self.netUpdateOption == 'policyNet':
raise NotImplementedError
targetValues = reward + self.gamma * torch.max(self.globalPolicyNet(nextState).detach(), dim=1)[0].unsqueeze(-1)
if self.netUpdateOption == 'doubleQ':
# select optimal action from policy net
with torch.no_grad():
batchAction = self.globalPolicyNet(nonFinalNextState).max(dim=1)[1].unsqueeze(-1)
QNext = torch.zeros(batchSize, device=self.device, dtype=torch.float32).unsqueeze(-1)
QNext[nonFinalMask] = self.globalTargetNet(nonFinalNextState).gather(1, batchAction)
targetValues = reward + self.gamma * QNext
loss = self.netLossFunc(QValues, targetValues)
self.globalOptimizer.zero_grad()
loss.backward()
if self.netGradClip is not None:
torch.nn.utils.clip_grad_norm_(self.globalPolicyNet.parameters(), self.netGradClip)
# global net update
self.globalOptimizer.step()
#
# # update local net
self.localNet.load_state_dict(self.globalPolicyNet.state_dict())
self.lock.release()
else:
# update local net
self.localNet.load_state_dict(self.globalPolicyNet.state_dict())
QValues = self.localNet(state).gather(1, action)
if self.netUpdateOption == 'targetNet':
# Here we detach because we do not want gradient flow from target values to net parameters
QNext = torch.zeros(batchSize, device=self.device, dtype=torch.float32)
QNext[nonFinalMask] = self.globalTargetNet(nonFinalNextState).max(1)[0].detach()
targetValues = reward + self.gamma * QNext.unsqueeze(-1)
if self.netUpdateOption == 'policyNet':
raise NotImplementedError
targetValues = reward + self.gamma * torch.max(self.globalPolicyNet(nextState).detach(), dim=1)[
0].unsqueeze(-1)
if self.netUpdateOption == 'doubleQ':
# select optimal action from policy net
with torch.no_grad():
batchAction = self.localNet(nonFinalNextState).max(dim=1)[1].unsqueeze(-1)
QNext = torch.zeros(batchSize, device=self.device, dtype=torch.float32).unsqueeze(-1)
QNext[nonFinalMask] = self.globalTargetNet(nonFinalNextState).gather(1, batchAction)
targetValues = reward + self.gamma * QNext
loss = self.netLossFunc(QValues, targetValues)
loss.backward()
self.lock.acquire()
self.globalOptimizer.zero_grad()
for lp, gp in zip(self.localNet.parameters(), self.globalPolicyNet.parameters()):
if self.device == 'cpu':
gp._grad = lp._grad
else:
gp._grad = lp._grad.cpu()
if self.netGradClip is not None:
torch.nn.utils.clip_grad_norm_(self.globalPolicyNet.parameters(), self.netGradClip)
# global net update
self.globalOptimizer.step()
self.lock.release()
#
# # update local net
self.localNet.load_state_dict(self.globalPolicyNet.state_dict())
def update_net(self, state, action, nextState, reward, info):
# state, nonFinalMask, nonFinalNextState, action, reward = self.prepare_minibatch(state, action, nextState, reward, info)
self.store_experience(state, action, nextState, reward, info)
if len(self.memory) < self.trainBatchSize:
return
transitions_raw = self.memory.sample(self.trainBatchSize)
transitions = Transition(*zip(*transitions_raw))
action = torch.tensor(transitions.action,
device=self.device,
dtype=torch.float32) # shape(batch, numActions)
reward = torch.tensor(transitions.reward,
device=self.device,
dtype=torch.float32) # shape(batch)
# for some env, the output state requires further processing before feeding to neural network
if self.stateProcessor is not None:
state, _ = self.stateProcessor(transitions.state, self.device)
nonFinalNextState, nonFinalMask = self.stateProcessor(
transitions.next_state, self.device)
else:
state = torch.tensor(transitions.state,
device=self.device,
dtype=torch.float32)
nonFinalMask = torch.tensor(tuple(
map(lambda s: s is not None, transitions.next_state)),
device=self.device,
dtype=torch.uint8)
nonFinalNextState = torch.tensor(
[s for s in transitions.next_state if s is not None],
device=self.device,
dtype=torch.float32)
batchSize = reward.shape[0]
# Critic loss
QValuesOne = self.criticNetOne.forward(state, action).squeeze()
QValuesTwo = self.criticNetTwo.forward(state, action).squeeze()
actionNoise = torch.randn((nonFinalNextState.shape[0], self.numAction),
dtype=torch.float32,
device=self.device)
next_actions = self.actorNet_target.forward(
nonFinalNextState) + actionNoise * self.policySmoothNoise
# next_actions = self.actorNet_target.forward(nonFinalNextState)
QNext = torch.zeros(batchSize, device=self.device, dtype=torch.float32)
QNextCriticOne = self.criticNet_targetOne.forward(
nonFinalNextState, next_actions.detach()).squeeze()
QNextCriticTwo = self.criticNet_targetTwo.forward(
nonFinalNextState, next_actions.detach()).squeeze()
QNext[nonFinalMask] = torch.min(QNextCriticOne, QNextCriticTwo)
targetValues = reward + self.gamma * QNext
criticOne_loss = self.netLossFunc(QValuesOne, targetValues)
criticTwo_loss = self.netLossFunc(QValuesTwo, targetValues)
self.criticOne_optimizer.zero_grad()
self.criticTwo_optimizer.zero_grad()
# https://jdhao.github.io/2017/11/12/pytorch-computation-graph/
criticOne_loss.backward(retain_graph=True)
criticTwo_loss.backward()
if self.netGradClip is not None:
torch.nn.utils.clip_grad_norm_(self.criticNetOne.parameters(),
self.netGradClip)
torch.nn.utils.clip_grad_norm_(self.criticNetTwo.parameters(),
self.netGradClip)
self.criticOne_optimizer.step()
self.criticTwo_optimizer.step()
if self.learnStepCounter % self.policyUpdateFreq:
# Actor loss
# we try to maximize criticNet output(which is state value)
policy_loss = -self.criticNetOne.forward(
state, self.actorNet.forward(state)).mean()
# update networks
self.actor_optimizer.zero_grad()
policy_loss.backward()
if self.netGradClip is not None:
torch.nn.utils.clip_grad_norm_(self.actorNet.parameters(),
self.netGradClip)
self.actor_optimizer.step()
if self.globalStepCount % self.lossRecordStep == 0:
self.losses.append([
self.globalStepCount, self.epIdx,
criticOne_loss.item(),
criticTwo_loss.item(),
policy_loss.item()
])
# update target networks
for target_param, param in zip(
self.actorNet_target.parameters(),
self.actorNet.parameters()):
target_param.data.copy_(param.data * self.tau +
target_param.data *
(1.0 - self.tau))
for target_param, param in zip(
self.criticNet_targetOne.parameters(),
self.criticNetOne.parameters()):
target_param.data.copy_(param.data * self.tau +
target_param.data *
(1.0 - self.tau))
for target_param, param in zip(
self.criticNet_targetTwo.parameters(),
self.criticNetTwo.parameters()):
target_param.data.copy_(param.data * self.tau +
target_param.data *
(1.0 - self.tau))
self.learnStepCounter += 1
def store_experience(self, states, actions, nextStates, rewards, infos):
for i in range(len(states)):
transition = Transition(states[i], actions[i], nextStates[i],
rewards[i])
self.memory.push(transition)
from Agents.Core.ReplayMemory import ReplayMemory, Transition
#from ..Agents.Core.ReplayMemory import ReplayMemory, Transition
import torch
tran1 = Transition(1, 1, 1, 1)
tran2 = Transition(2, 2, 2, 2)
memory = ReplayMemory(10)
memory.push(tran1)
memory.push(tran2)
memory.push(3, 3, 3, 3)
print(memory)
memory.write_to_text('memoryOut.txt')
toTensor = memory.totensor()
toTensor2 = torch.tensor(memory.sample(2))
for i in range(5, 50):
tran = Transition(i, i, i, i)
memory.push(tran)
print(memory)
memory.clear()
print(memory)
print(toTensor)
print(toTensor2)
from Agents.Core.ReplayMemory import ReplayMemory, Transition
#from ..Agents.Core.ReplayMemory import ReplayMemory, Transition
import torch
import numpy as np
import pickle
state1 = np.random.rand(5, 5)
state2 = np.random.rand(5, 5)
state3 = np.random.rand(5, 5)
state4 = np.random.rand(5, 5)
tran1 = Transition(state1, 1, state2, 1)
tran2 = Transition(state3, 2, state4, 2)
memory = ReplayMemory(10)
memory.push(tran1)
memory.push(tran2)
print(memory)
file = open('memory.pickle', 'wb')
pickle.dump(memory, file)
file.close()
with open('memory.pickle', 'rb') as file:
memory2 = pickle.load(file)
print(memory2)
def store_experience(self, state, action, nextState, reward, info):
if self.experienceProcessor is not None:
state, action, nextState, reward = self.experienceProcessor(state, action, nextState, reward, info)
transition = Transition(state, action, nextState, reward)
self.memory.push(transition)