diff --git a/pfrl/agents/ddpg.py b/pfrl/agents/ddpg.py
index 08c0748da..9319f5475 100644
--- a/pfrl/agents/ddpg.py
+++ b/pfrl/agents/ddpg.py
@@ -168,7 +168,7 @@ def compute_critic_loss(self, batch):
         loss = F.mse_loss(target_q, predict_q)
 
         # Update stats
-        self.critic_loss_record.append(float(loss.detach().cpu().numpy()))
+        self.critic_loss_record.append(loss.item())
 
         return loss
 
@@ -182,7 +182,7 @@ def compute_actor_loss(self, batch):
 
         # Update stats
         self.q_record.extend(q.detach().cpu().numpy())
-        self.actor_loss_record.append(float(loss.detach().cpu().numpy()))
+        self.actor_loss_record.append(loss.item())
 
         return loss
 
diff --git a/pfrl/agents/soft_actor_critic.py b/pfrl/agents/soft_actor_critic.py
index 75e8ce98a..57f76ae94 100644
--- a/pfrl/agents/soft_actor_critic.py
+++ b/pfrl/agents/soft_actor_critic.py
@@ -246,8 +246,8 @@ def update_q_func(self, batch):
         # Update stats
         self.q1_record.extend(predict_q1.detach().cpu().numpy())
         self.q2_record.extend(predict_q2.detach().cpu().numpy())
-        self.q_func1_loss_record.append(float(loss1))
-        self.q_func2_loss_record.append(float(loss2))
+        self.q_func1_loss_record.append(loss1.item())
+        self.q_func2_loss_record.append(loss2.item())
 
         self.q_func1_optimizer.zero_grad()
         loss1.backward()
diff --git a/pfrl/agents/td3.py b/pfrl/agents/td3.py
index dc913f56d..e70e7e98b 100644
--- a/pfrl/agents/td3.py
+++ b/pfrl/agents/td3.py
@@ -213,8 +213,8 @@ def update_q_func(self, batch):
         # Update stats
         self.q1_record.extend(predict_q1.detach().cpu().numpy())
         self.q2_record.extend(predict_q2.detach().cpu().numpy())
-        self.q_func1_loss_record.append(float(loss1))
-        self.q_func2_loss_record.append(float(loss2))
+        self.q_func1_loss_record.append(loss1.item())
+        self.q_func2_loss_record.append(loss2.item())
 
         self.q_func1_optimizer.zero_grad()
         loss1.backward()
@@ -241,7 +241,7 @@ def update_policy(self, batch):
         # Since we want to maximize Q, loss is negation of Q
         loss = -torch.mean(q)
 
-        self.policy_loss_record.append(float(loss))
+        self.policy_loss_record.append(loss.item())
         self.policy_optimizer.zero_grad()
         loss.backward()
         if self.max_grad_norm is not None: