wassname/DeepRL
1
0
Fork 0
mirror of https://github.com/wassname/DeepRL.git synced 2026-09-01 11:50:22 +08:00
Code Issues Packages Projects Releases Wiki Activity

Update DDPG

Browse Source
This commit is contained in:
Shangtong Zhang
2018-01-19 11:24:32 -07:00
parent db49a76c53
commit eb16eeb2a1
3 changed files with 20 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
agent/DDPG_agent.py
+1 -1
View File
@@ -97,7 +97,7 @@ class DDPGAgent:
actions = actor.predict(states, False)
var_actions = Variable(actions.data, requires_grad=True)
q = critic.predict(states, var_actions)
q.backward(torch.ones(q.size()))
q.backward(critic.FloatTensor(np.ones(q.size())))
actor.zero_grad()
self.actor_opt.zero_grad()
main.py
+5 -6
View File
@@ -286,15 +286,14 @@ def ddpg_continuous():
# config.task_fn = lambda: BipedalWalker()
task = config.task_fn()
config.actor_network_fn = lambda: DeterministicActorNet(
task.state_dim, task.action_dim, F.tanh, 1, non_linear=F.relu, batch_norm=False)
task.state_dim, task.action_dim, F.tanh, 1, non_linear=F.relu, batch_norm=False, gpu=False)
config.critic_network_fn = lambda: DeterministicCriticNet(
task.state_dim, task.action_dim, non_linear=F.relu, batch_norm=False)
task.state_dim, task.action_dim, non_linear=F.relu, batch_norm=False, gpu=False)
config.network_fn = lambda: DisjointActorCriticNet(config.actor_network_fn, config.critic_network_fn)
config.actor_optimizer_fn = lambda params: torch.optim.Adam(params, lr=1e-4)
config.critic_optimizer_fn =\
lambda params: torch.optim.Adam(params, lr=1e-3, weight_decay=0.01)
config.replay_fn = lambda: SharedReplay(memory_size=1000000, batch_size=64,
state_shape=(task.state_dim, ), action_shape=(task.action_dim, ))
config.replay_fn = lambda: HighDimActionReplay(memory_size=1000000, batch_size=64)
config.discount = 0.99
config.max_episode_length = task.max_episode_steps
config.random_process_fn = \
@@ -320,11 +319,11 @@ if __name__ == '__main__':
# dqn_cart_pole()
# async_cart_pole()
a3c_cart_pole()
# a3c_cart_pole()
# a3c_continuous()
# p3o_continuous()
# d3pg_continuous()
# ddpg_continuous()
ddpg_continuous()
# dqn_fruit()
# hrdqn_fruit()
network/continuous_action_network.py
+14 -22
View File
@@ -29,22 +29,18 @@ class DeterministicActorNet(nn.Module, BasicNet):
self.layer2 = nn.Linear(hidden_size, hidden_size)
self.batch_norm = batch_norm
BasicNet.__init__(self, None, gpu, False)
self.init_weights()
BasicNet.__init__(self, None, gpu, False)
def init_weights(self):
bound = 3e-3
self.layer3.weight.data.uniform_(-bound, bound)
self.layer3.bias.data.fill_(0)
nn.init.uniform(self.layer3.weight.data, -bound, bound)
nn.init.constant(self.layer3.bias.data, 0)
def fanin(size):
v = 1.0 / np.sqrt(size[1])
return torch.FloatTensor(size).uniform_(-v, v)
self.layer1.weight.data = fanin(self.layer1.weight.data.size())
self.layer1.bias.data.fill_(0)
self.layer2.weight.data = fanin(self.layer2.weight.data.size())
self.layer2.bias.data.fill_(0)
nn.init.xavier_uniform(self.layer1.weight.data)
nn.init.constant(self.layer1.bias.data, 0)
nn.init.xavier_uniform(self.layer2.weight.data)
nn.init.constant(self.layer2.bias.data, 0)
def forward(self, x):
x = self.to_torch_variable(x)
@@ -83,22 +79,18 @@ class DeterministicCriticNet(nn.Module, BasicNet):
self.bn2 = nn.BatchNorm1d(hidden_size)
self.batch_norm = batch_norm
BasicNet.__init__(self, None, gpu, False)
self.init_weights()
BasicNet.__init__(self, None, gpu, False)
def init_weights(self):
bound = 3e-3
self.layer3.weight.data.uniform_(-bound, bound)
self.layer3.bias.data.fill_(0)
nn.init.uniform(self.layer3.weight.data, -bound, bound)
nn.init.constant(self.layer3.bias.data, 0)
def fanin(size):
v = 1.0 / np.sqrt(size[1])
return torch.FloatTensor(size).uniform_(-v, v)
self.layer1.weight.data = fanin(self.layer1.weight.data.size())
self.layer1.bias.data.fill_(0)
self.layer2.weight.data = fanin(self.layer2.weight.data.size())
self.layer2.bias.data.fill_(0)
nn.init.xavier_uniform(self.layer1.weight.data)
nn.init.constant(self.layer1.bias.data, 0)
nn.init.xavier_uniform(self.layer2.weight.data)
nn.init.constant(self.layer2.bias.data, 0)
def forward(self, x, action):
x = self.to_torch_variable(x)