class DQN:
def __init__(self,
n_states,
n_actions,
gamma=0.99,
epsilon_start=0.9,
epsilon_end=0.05,
epsilon_decay=200,
memory_capacity=10000,
policy_lr=0.01,
batch_size=128,
device="cpu"):
self.actions_count = 0
self.n_actions = n_actions
self.device = device
self.gamma = gamma
self.epsilon = 0
self.epsilon_start = epsilon_start
self.epsilon_end = epsilon_end
self.epsilon_decay = epsilon_decay
self.batch_size = batch_size
self.policy_net = FCN(n_states, n_actions).to(self.device)
self.target_net = FCN(n_states, n_actions).to(self.device)
self.target_net.load_state_dict(self.policy_net.state_dict())
self.target_net.eval() # 不启用 BatchNormalization 和 Dropout
self.optimizer = optim.Adam(self.policy_net.parameters(), lr=policy_lr)
self.loss = 0
self.memory = ReplayBuffer(memory_capacity)
def select_action(self, state):
'''选择工作
Args:
state [array]: 状态
Returns:
[array]: 动作
'''
self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \
math.exp(-1. * self.actions_count / self.epsilon_decay)
self.actions_count += 1
if random.random() > self.epsilon:
with torch.no_grad():
state = torch.tensor(
[state], device=self.device, dtype=torch.float32
) # 先转为张量便于丢给神经网络,state元素数据原本为float64；注意state=torch.tensor(state).unsqueeze(0)跟state=torch.tensor([state])等价
q_value = self.policy_net(
state
) # tensor([[-0.0798, -0.0079]], grad_fn=<AddmmBackward>)
action = q_value.max(1)[1].item()
else:
action = random.randrange(self.n_actions)
return action
def update(self):
if len(self.memory) < self.batch_size:
return
state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(
self.batch_size)
state_batch = torch.tensor(
state_batch, device=self.device, dtype=torch.float
) # 例如tensor([[-4.5543e-02, -2.3910e-01, 1.8344e-02, 2.3158e-01],...,[-1.8615e-02, -2.3921e-01, -1.1791e-02, 2.3400e-01]])
action_batch = torch.tensor(action_batch,
device=self.device).unsqueeze(
1) # 例如tensor([[1],...,[0]])
reward_batch = torch.tensor(
reward_batch, device=self.device,
dtype=torch.float) # tensor([1., 1.,...,1])
next_state_batch = torch.tensor(next_state_batch,
device=self.device,
dtype=torch.float)
done_batch = torch.tensor(np.float32(done_batch),
device=self.device).unsqueeze(
1) # 将bool转为float然后转为张量
# Compute Q(s_t, a) - the model computes Q(s_t), then we select the
# columns of actions taken. These are the actions which would've been taken
# for each batch state according to policy_net
q_values = self.policy_net(state_batch).gather(
1, action_batch) # 等价于self.forward
# Compute V(s_{t+1}) for all next states.
# Expected values of actions for non_final_next_states are computed based
# on the "older" target_net; selecting their best reward with max(1)[0].
# This is merged based on the mask, such that we'll have either the expected
# state value or 0 in case the state was final.
next_state_values = self.target_net(next_state_batch).max(
1)[0].detach() # tensor([ 0.0060, -0.0171,...,])
# Compute the expected Q values
expected_q_values = reward_batch + self.gamma * next_state_values * (
1 - done_batch[0])
# Compute Huber loss
# self.loss = nn.MSELoss(q_values, expected_q_values.unsqueeze(1))
self.loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1))
# Optimize the model
self.optimizer.zero_grad(
) # zero_grad clears old gradients from the last step (otherwise you’d just accumulate the gradients from all loss.backward() calls).
self.loss.backward(
) # loss.backward() computes the derivative of the loss w.r.t. the parameters (or anything requiring gradients) using backpropagation.
for param in self.policy_net.parameters(): # clip防止梯度爆炸
param.grad.data.clamp_(-1, 1)
self.optimizer.step(
) # causes the optimizer to take a step based on the gradients of the parameters.
class DQN:
def __init__(self,
n_states,
n_actions,
gamma=0.99,
epsilon_start=0.9,
epsilon_end=0.05,
epsilon_decay=200,
memory_capacity=10000,
policy_lr=0.01,
batch_size=128,
device="cpu"):
self.actions_count = 0
self.n_actions = n_actions # 总的动作个数
self.device = device # 设备，cpu或gpu等
self.gamma = gamma
# e-greedy 策略相关参数
self.epsilon = 0
self.epsilon_start = epsilon_start
self.epsilon_end = epsilon_end
self.epsilon_decay = epsilon_decay
self.batch_size = batch_size
self.policy_net = FCN(n_states, n_actions).to(self.device)
self.target_net = FCN(n_states, n_actions).to(self.device)
# target_net的初始模型参数完全复制policy_net
self.target_net.load_state_dict(self.policy_net.state_dict())
self.target_net.eval() # 不启用 BatchNormalization 和 Dropout
# 可查parameters()与state_dict()的区别，前者require_grad=True
self.optimizer = optim.Adam(self.policy_net.parameters(), lr=policy_lr)
self.loss = 0
self.memory = ReplayBuffer(memory_capacity)
def select_action(self, state):
'''选择动作
Args:
state [array]: [description]
Returns:
action [array]: [description]
'''
self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \
math.exp(-1. * self.actions_count / self.epsilon_decay)
self.actions_count += 1
if random.random() > self.epsilon:
with torch.no_grad():
# 先转为张量便于丢给神经网络,state元素数据原本为float64
# 注意state=torch.tensor(state).unsqueeze(0)跟state=torch.tensor([state])等价
state = torch.tensor([state],
device=self.device,
dtype=torch.float32)
# 如tensor([[-0.0798, -0.0079]], grad_fn=<AddmmBackward>)
q_value = self.policy_net(state)
# tensor.max(1)返回每行的最大值以及对应的下标，
# 如torch.return_types.max(values=tensor([10.3587]),indices=tensor([0]))
# 所以tensor.max(1)[1]返回最大值对应的下标，即action
action = q_value.max(1)[1].item()
else:
action = random.randrange(self.n_actions)
return action
def update(self):
if len(self.memory) < self.batch_size:
return
# 从memory中随机采样transition
state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(
self.batch_size)
# 转为张量
# 例如tensor([[-4.5543e-02, -2.3910e-01, 1.8344e-02, 2.3158e-01],...,[-1.8615e-02, -2.3921e-01, -1.1791e-02, 2.3400e-01]])
state_batch = torch.tensor(state_batch,
device=self.device,
dtype=torch.float)
action_batch = torch.tensor(action_batch,
device=self.device).unsqueeze(
1) # 例如tensor([[1],...,[0]])
reward_batch = torch.tensor(
reward_batch, device=self.device,
dtype=torch.float) # tensor([1., 1.,...,1])
next_state_batch = torch.tensor(next_state_batch,
device=self.device,
dtype=torch.float)
done_batch = torch.tensor(np.float32(done_batch),
device=self.device).unsqueeze(
1) # 将bool转为float然后转为张量
# 计算当前(s_t,a)对应的Q(s_t, a)
# 关于torch.gather,对于a=torch.Tensor([[1,2],[3,4]])
# 那么a.gather(1,torch.Tensor([[0],[1]]))=torch.Tensor([[1],[3]])
q_values = self.policy_net(state_batch).gather(
dim=1, index=action_batch) # 等价于self.forward
# 计算所有next states的V(s_{t+1})，即通过target_net中选取reward最大的对应states
next_state_values = self.target_net(next_state_batch).max(
1)[0].detach() # 比如tensor([ 0.0060, -0.0171,...,])
# 计算 expected_q_value
# 对于终止状态，此时done_batch[0]=1, 对应的expected_q_value等于reward
expected_q_values = reward_batch + self.gamma * \
next_state_values * (1-done_batch[0])
# self.loss = F.smooth_l1_loss(q_values,expected_q_values.unsqueeze(1)) # 计算 Huber loss
self.loss = nn.MSELoss()(q_values,
expected_q_values.unsqueeze(1)) # 计算 均方误差loss
# 优化模型
self.optimizer.zero_grad(
) # zero_grad清除上一步所有旧的gradients from the last step
# loss.backward()使用backpropagation计算loss相对于所有parameters(需要gradients)的微分
self.loss.backward()
for param in self.policy_net.parameters(): # clip防止梯度爆炸
param.grad.data.clamp_(-1, 1)
self.optimizer.step() # 更新模型
def save_model():
pass
def load_model():
pass
class DQN:
def __init__(self,
n_states,
n_actions,
gamma=0.99,
epsilon_start=0.9,
epsilon_end=0.05,
epsilon_decay=200,
memory_capacity=10000,
policy_lr=0.01,
batch_size=128,
device="cpu"):
self.actions_count = 0
self.n_actions = n_actions # 总的动作个数
self.device = device # 设备，cpu或gpu等
self.gamma = gamma
# e-greedy策略相关参数
self.epsilon = 0
self.epsilon_start = epsilon_start
self.epsilon_end = epsilon_end
self.epsilon_decay = epsilon_decay
self.batch_size = batch_size
self.policy_net = FCN(n_states, n_actions).to(self.device)
self.target_net = FCN(n_states, n_actions).to(self.device)
# target_net的初始模型参数完全复制policy_net
self.target_net.load_state_dict(self.policy_net.state_dict())
self.target_net.eval() # 不启用 BatchNormalization 和 Dropout
# 可查parameters()与state_dict()的区别，前者require_grad=True
self.optimizer = optim.Adam(self.policy_net.parameters(), lr=policy_lr)
self.loss = 0
self.memory = ReplayBuffer(memory_capacity)
def choose_action(self, state, train=True):
'''选择动作
'''
if train:
self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \
math.exp(-1. * self.actions_count / self.epsilon_decay)
self.actions_count += 1
if random.random() > self.epsilon:
with torch.no_grad():
# 先转为张量便于丢给神经网络,state元素数据原本为float64
# 注意state=torch.tensor(state).unsqueeze(0)跟state=torch.tensor([state])等价
state = torch.tensor([state],
device=self.device,
dtype=torch.float32)
# 如tensor([[-0.0798, -0.0079]], grad_fn=<AddmmBackward>)
q_value = self.policy_net(state)
# tensor.max(1)返回每行的最大值以及对应的下标，
# 如torch.return_types.max(values=tensor([10.3587]),indices=tensor([0]))
# 所以tensor.max(1)[1]返回最大值对应的下标，即action
action = q_value.max(1)[1].item()
else:
action = random.randrange(self.n_actions)
return action
else:
with torch.no_grad():
# 先转为张量便于丢给神经网络,state元素数据原本为float64
# 注意state=torch.tensor(state).unsqueeze(0)跟state=torch.tensor([state])等价
state = torch.tensor([state],
device='cpu',
dtype=torch.float32)
# 如tensor([[-0.0798, -0.0079]], grad_fn=<AddmmBackward>)
q_value = self.target_net(state)
# tensor.max(1)返回每行的最大值以及对应的下标，
# 如torch.return_types.max(values=tensor([10.3587]),indices=tensor([0]))
# 所以tensor.max(1)[1]返回最大值对应的下标，即action
action = q_value.max(1)[1].item()
return action
def update(self):
if len(self.memory) < self.batch_size:
return
# 从memory中随机采样transition
state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(
self.batch_size)
# 转为张量
# 例如tensor([[-4.5543e-02, -2.3910e-01, 1.8344e-02, 2.3158e-01],...,[-1.8615e-02, -2.3921e-01, -1.1791e-02, 2.3400e-01]])
state_batch = torch.tensor(state_batch,
device=self.device,
dtype=torch.float)
action_batch = torch.tensor(action_batch,
device=self.device).unsqueeze(
1) # 例如tensor([[1],...,[0]])
reward_batch = torch.tensor(
reward_batch, device=self.device,
dtype=torch.float) # tensor([1., 1.,...,1])
next_state_batch = torch.tensor(next_state_batch,
device=self.device,
dtype=torch.float)
done_batch = torch.tensor(np.float32(done_batch),
device=self.device).unsqueeze(
1) # 将bool转为float然后转为张量
# 计算当前(s_t,a)对应的Q(s_t, a)
q_values = self.policy_net(state_batch)
next_q_values = self.policy_net(next_state_batch)
# 代入当前选择的action，得到Q(s_t|a=a_t)
q_value = q_values.gather(dim=1, index=action_batch)
'''以下是Nature DQN的q_target计算方式
# 计算所有next states的Q'(s_{t+1})的最大值，Q'为目标网络的q函数
next_q_state_value = self.target_net(
next_state_batch).max(1)[0].detach() # 比如tensor([ 0.0060, -0.0171,...,])
# 计算 q_target
# 对于终止状态，此时done_batch[0]=1, 对应的expected_q_value等于reward
q_target = reward_batch + self.gamma * next_q_state_value * (1-done_batch[0])
'''
'''以下是Double DQNq_target计算方式，与NatureDQN稍有不同'''
next_target_values = self.target_net(next_state_batch)
# 选出Q(s_t‘, a)对应的action，代入到next_target_values获得target net对应的next_q_value，即Q’(s_t|a=argmax Q(s_t‘, a))
next_target_q_value = next_target_values.gather(
1,
torch.max(next_q_values, 1)[1].unsqueeze(1)).squeeze(1)
q_target = reward_batch + self.gamma * next_target_q_value * (
1 - done_batch[0])
self.loss = nn.MSELoss()(q_value, q_target.unsqueeze(1)) # 计算 均方误差loss
# 优化模型
self.optimizer.zero_grad(
) # zero_grad清除上一步所有旧的gradients from the last step
# loss.backward()使用backpropagation计算loss相对于所有parameters(需要gradients)的微分
self.loss.backward()
for param in self.policy_net.parameters(): # clip防止梯度爆炸
param.grad.data.clamp_(-1, 1)
self.optimizer.step() # 更新模型
def save_model(self, path):
torch.save(self.target_net.state_dict(), path)
def load_model(self, path):
self.target_net.load_state_dict(torch.load(path))
class PolicyGradient:
def __init__(self,
state_dim,
device='cpu',
gamma=0.99,
lr=0.01,
batch_size=5):
self.gamma = gamma
self.policy_net = FCN(state_dim)
self.optimizer = torch.optim.RMSprop(self.policy_net.parameters(),
lr=lr)
self.batch_size = batch_size
def choose_action(self, state):
state = torch.from_numpy(state).float()
state = Variable(state)
probs = self.policy_net(state)
m = Bernoulli(probs)
action = m.sample()
action = action.data.numpy().astype(int)[0] # 转为标量
return action
def update(self, reward_pool, state_pool, action_pool):
# Discount reward
running_add = 0 # 就是那个有discount的公式
for i in reversed(range(len(reward_pool))): # 倒数
if reward_pool[i] == 0:
running_add = 0
else:
running_add = running_add * self.gamma + reward_pool[i]
reward_pool[i] = running_add
# 得到G
# Normalize reward
reward_mean = np.mean(reward_pool)
reward_std = np.std(reward_pool)
for i in range(len(reward_pool)):
reward_pool[i] = (reward_pool[i] - reward_mean) / reward_std
# 归一化
# Gradient Desent
self.optimizer.zero_grad()
for i in range(len(reward_pool)): # 从前往后
state = state_pool[i]
action = Variable(torch.FloatTensor([action_pool[i]]))
reward = reward_pool[i]
state = Variable(torch.from_numpy(state).float())
probs = self.policy_net(state)
m = Bernoulli(probs)
# Negtive score function x reward
loss = -m.log_prob(action) * reward # 核心
# print(loss)
loss.backward()
self.optimizer.step()
def save_model(self, path):
torch.save(self.policy_net.state_dict(), path)
def load_model(self, path):
self.policy_net.load_state_dict(torch.load(path))
def train(data_loader, model_index, x_eval_train, y_eval_train):
### Model Initiation
fcn = FCN()
#print (fcn.b_1_conv_1[0].weight.data)
d = tor.load("./models/vgg16_pretrained.pkl")
fcn.vgg16_load(d)
#d = tor.load("./models/fcn_model_1_1.pkl")
#fcn.load_state_dict(d)
fcn.cuda()
#loss_func = tor.nn.CrossEntropyLoss(weight=w)
loss_func = tor.nn.CrossEntropyLoss()
#loss_func = tor.nn.MSELoss()
#optim = tor.optim.SGD(fcn.parameters(), lr=LR, momentum=MOMENTUM)
optim1 = tor.optim.Adam(fcn.b_6_conv_1.parameters(), lr=LR)
optim2 = tor.optim.Adam(fcn.b_6_conv_2.parameters(), lr=LR)
optim3 = tor.optim.Adam(fcn.b_6_conv_3.parameters(), lr=LR)
optim4 = tor.optim.Adam(fcn.b_7_trans_1.parameters(), lr=LR)
optim = tor.optim.Adam(fcn.parameters(), lr=LR)
### Training
for epoch in range(EPOCH):
print("|Epoch: {:>4} |".format(epoch + 1), end="")
### Training
for step, (x_batch, y_batch) in enumerate(data_loader):
x = Variable(x_batch).type(tor.FloatTensor).cuda()
y = Variable(y_batch).type(tor.LongTensor).cuda()
pred = fcn(x)
optim1.zero_grad()
optim2.zero_grad()
optim3.zero_grad()
optim4.zero_grad()
optim.zero_grad()
loss = loss_func(pred, y)
loss.backward()
optim1.step()
optim2.step()
optim3.step()
optim4.step()
print(pred[:2])
print(tor.max(pred[:5], 1)[1])
### Evaluation
loss = float(loss.data)
acc = evaluate(fcn, x_eval_train, y_eval_train)
print("|Loss: {:<8} |Acc: {:<8}".format(loss, acc))
### Save model
if epoch % RECORD_MODEL_PERIOD == 0:
tor.save(
fcn.state_dict(),
os.path.join(MODEL_ROOT,
"fcn_model_{}_{}.pkl".format(model_index, epoch)))
### Record
record_data = dict()
if epoch == 0:
record_data["model_name"] = "fcn_model_{}.pkl".format(model_index)
record_data["data_size"] = AVAILABLA_SIZE
record_data["batch_size"] = BATCHSIZE
record_data["decay"] = str((LR_STEPSIZE, LR_GAMMA))
record_data["lr_init"] = float(optim1.param_groups[0]["lr"])
record_data["lr"] = float(optim1.param_groups[0]["lr"])
record_data["record_epoch"] = RECORD_MODEL_PERIOD
record_data["loss"] = loss
record_data["acc"] = acc
else:
record_data["model_name"] = "fcn_model_{}.pkl".format(model_index)
record_data["lr"] = float(optim1.param_groups[0]["lr"])
record_data["loss"] = loss
record_data["acc"] = acc
record(RECORD_FP, record_data)
for predIdx in range(numClass):
conf[(predIdx, labIdx)] /= (labCnts[labIdx] / 100.0)
meanClassAcc = 0.0
meanIoU = torch.sum(IoU / imgCnt).item() / numClass * 100
currLoss = running_loss / (i + 1)
for j in range(numClass):
meanClassAcc += conf[(j, j)] / numClass
print(
"Epoch [%d] Validation Loss: %.4f Validation Pixel Acc: %.2f Mean Class Acc: %.2f IoU: %.2f"
% (epoch + 1, running_loss / (i + 1), running_acc /
(imgCnt), meanClassAcc, meanIoU))
ploter.plot("loss", "val", epoch + 1, running_loss / (i + 1))
if bestLoss > currLoss:
conf[conf < 0.001] = 0
print(conf)
bestConf = conf
bestLoss = currLoss
bestIoU = meanIoU
bestAcc = meanClassAcc
bestTAcc = running_acc / (imgCnt)
torch.save(model.state_dict(), root + "bestModelSeg.pth")
scheduler.step(currLoss)
print(
"Optimization finished Validation Loss: %.4f Pixel Acc: %.2f Mean Class Acc: %.2f IoU: %.2f"
% (bestLoss, bestTAcc, bestAcc, bestIoU))
print(bestConf)