fix bug: SAC/TD3/DDPG

Author: Vinson-sheep

Char05 DDPG/DDPG.py

@@ -63,7 +63,7 @@ def __init__(self, state_dim, action_dim, init_w=3e-3):
         self.l2 = nn.Linear(256, 256)
         self.l3 = nn.Linear(256, action_dim)
 
-        self.l3.weight.data.uniform_(init_w, init_w)
+        self.l3.weight.data.uniform_(-init_w, init_w)
         self.l3.bias.data.uniform_(-init_w, init_w)
 
     def forward(self, state):

Char09 SAC/SAC_v1.py

@@ -66,7 +66,7 @@ def __init__(self, state_dim, action_dim, init_w=3e-3):
         self.l2 = nn.Linear(256, 256)
 
         self.mu_head = nn.Linear(256, action_dim)
-        self.mu_head.weight.data.uniform_(init_w, init_w)
+        self.mu_head.weight.data.uniform_(-init_w, init_w)
         self.mu_head.bias.data.uniform_(-init_w, init_w)
 
         self.log_std_head = nn.Linear(256, action_dim)
@@ -210,7 +210,7 @@ def update(self):
 
         state = torch.tensor(state, dtype=torch.float).to(device)
         action = torch.tensor(action, dtype=torch.float).to(device)
-        reward = torch.tensor(reward, dtype=torch.float).view(batch_size, -1).to(device)
+        reward = torch.tensor(reward, dtype=torch.float).view(batch_size, 1).to(device)
         next_state = torch.tensor(next_state, dtype=torch.float).to(device)
         done = torch.tensor(done, dtype=torch.float).to(device).view(batch_size, -1).to(device)
 

Char09 SAC/SAC_v2.py

@@ -66,7 +66,7 @@ def __init__(self, state_dim, action_dim, init_w=3e-3):
         self.l2 = nn.Linear(256, 256)
 
         self.mu_head = nn.Linear(256, action_dim)
-        self.mu_head.weight.data.uniform_(init_w, init_w)
+        self.mu_head.weight.data.uniform_(-init_w, init_w)
         self.mu_head.bias.data.uniform_(-init_w, init_w)
 
         self.log_std_head = nn.Linear(256, action_dim)

Char10 TD3/TD3.py

@@ -66,7 +66,7 @@ def __init__(self, state_dim, action_dim, init_w=3e-3):
         self.l2 = nn.Linear(256, 256)
         self.l3 = nn.Linear(256, action_dim)
 
-        self.l3.weight.data.uniform_(init_w, init_w)
+        self.l3.weight.data.uniform_(-init_w, init_w)
         self.l3.bias.data.uniform_(-init_w, init_w)
 
     def forward(self, state):