Merge pull request #9 from apaszke/master

A cleaner solution to grad sharing problem

Author: ikostrikov2

.gitignore

@@ -0,0 +1,2 @@
+__pycache__
+*.pyc

model.py

@@ -59,16 +59,6 @@ def __init__(self, num_inputs, action_space):
         self.lstm.bias_hh.data.fill_(0)
 
         self.train()
-        self.__dummy_backprob()
-
-    def __dummy_backprob(self):
-        # See: https://discuss.pytorch.org/t/problem-on-variable-grad-data/957/7
-        # An ugly hack until there is a better solution.
-        inputs = Variable(torch.randn(1, 1, 42, 42))
-        hx, cx = Variable(torch.randn(1, 256)), Variable(torch.randn(1, 256))
-        outputs = self((inputs, (hx, cx)))
-        loss = (outputs[0].mean() + outputs[1].mean()) * 0.0
-        loss.backward()
 
     def forward(self, inputs):
         inputs, (hx, cx) = inputs

train.py

@@ -11,6 +11,13 @@
 from torchvision import datasets, transforms
 
 
+def ensure_shared_grads(model, shared_model):
+    for param, shared_param in zip(model.parameters(), shared_model.parameters()):
+        if shared_param.grad is not None:
+            return
+        shared_param._grad = param.grad
+
+
 def train(rank, args, shared_model):
     torch.manual_seed(args.seed + rank)
 
@@ -19,10 +26,6 @@ def train(rank, args, shared_model):
 
     model = ActorCritic(env.observation_space.shape[0], env.action_space)
 
-    for param, shared_param in zip(model.parameters(), shared_model.parameters()):
-        # Use gradients from the local model
-        shared_param.grad.data = param.grad.data
-
     optimizer = optim.Adam(shared_model.parameters(), lr=args.lr)
 
     model.train()
@@ -102,14 +105,9 @@ def train(rank, args, shared_model):
                 log_probs[i] * Variable(gae) - 0.01 * entropies[i]
 
         optimizer.zero_grad()
+
         (policy_loss + 0.5 * value_loss).backward()
+        torch.nn.utils.clip_grad_norm(model.parameters(), 40)
 
-        global_norm = 0
-        for param in model.parameters():
-            global_norm += param.grad.data.pow(2).sum()
-        global_norm = math.sqrt(global_norm)
-        ratio = 40 / global_norm
-        if ratio < 1:
-            for param in model.parameters():
-                param.grad.data.mul_(ratio)
+        ensure_shared_grads(model, shared_model)
         optimizer.step()