Match performance with stable-baselines (discrete case)

stable_baselines3/a2c/a2c.py

@@ -116,6 +116,7 @@ def train(self) -> None:
         # Update optimizer learning rate
         self._update_learning_rate(self.policy.optimizer)
 
+        # This will only loop once (get all data in one go)
         for rollout_data in self.rollout_buffer.get(batch_size=None):
 
             actions = rollout_data.actions

stable_baselines3/common/base_class.py

@@ -123,6 +123,7 @@ def __init__(
         self.tensorboard_log = tensorboard_log
         self.lr_schedule = None  # type: Optional[Callable]
         self._last_obs = None  # type: Optional[np.ndarray]
+        self._last_dones = None  # type: Optional[np.ndarray]
         # When using VecNormalize:
         self._last_original_obs = None  # type: Optional[np.ndarray]
         self._episode_num = 0
@@ -474,6 +475,7 @@ def _setup_learn(
         # Avoid resetting the environment when calling ``.learn()`` consecutive times
         if reset_num_timesteps or self._last_obs is None:
             self._last_obs = self.env.reset()
+            self._last_dones = np.zeros((self._last_obs.shape[0],), dtype=np.bool)
             # Retrieve unnormalized observation for saving into the buffer
             if self._vec_normalize_env is not None:
                 self._last_original_obs = self._vec_normalize_env.get_original_obs()

stable_baselines3/common/on_policy_algorithm.py

@@ -173,8 +173,9 @@ def collect_rollouts(
             if isinstance(self.action_space, gym.spaces.Discrete):
                 # Reshape in case of discrete action
                 actions = actions.reshape(-1, 1)
-            rollout_buffer.add(self._last_obs, actions, rewards, dones, values, log_probs)
+            rollout_buffer.add(self._last_obs, actions, rewards, self._last_dones, values, log_probs)
             self._last_obs = new_obs
+            self._last_dones = dones
 
         rollout_buffer.compute_returns_and_advantage(values, dones=dones)
 

stable_baselines3/common/sb2_compat/rmsprop_tf_like.py

@@ -0,0 +1,126 @@
+import torch
+from torch.optim import Optimizer
+
+
+class RMSpropTFLike(Optimizer):
+    r"""Implements RMSprop algorithm with closer match to Tensorflow version.
+
+    For reproducibility with original stable-baselines. Use this
+    version with e.g. A2C for stabler learning than with the PyTorch
+    RMSProp. Based on the PyTorch v1.5.0 implementation of RMSprop.
+
+    See a more throughout conversion in pytorch-image-models repository:
+        https://github.com/rwightman/pytorch-image-models/blob/master/timm/optim/rmsprop_tf.py
+
+    Changes to the original RMSprop:
+        - Move epsilon inside square root
+        - Initialize squared gradient to ones rather than zeros
+
+    Proposed by G. Hinton in his
+    `course <http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf>`_.
+
+    The centered version first appears in `Generating Sequences
+    With Recurrent Neural Networks <https://arxiv.org/pdf/1308.0850v5.pdf>`_.
+
+    The implementation here takes the square root of the gradient average before
+    adding epsilon (note that TensorFlow interchanges these two operations). The effective
+    learning rate is thus :math:`\alpha/(\sqrt{v} + \epsilon)` where :math:`\alpha`
+    is the scheduled learning rate and :math:`v` is the weighted moving average
+    of the squared gradient.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float, optional): learning rate (default: 1e-2)
+        momentum (float, optional): momentum factor (default: 0)
+        alpha (float, optional): smoothing constant (default: 0.99)
+        eps (float, optional): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        centered (bool, optional) : if ``True``, compute the centered RMSProp,
+            the gradient is normalized by an estimation of its variance
+        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
+
+    """
+
+    def __init__(self, params, lr=1e-2, alpha=0.99, eps=1e-8, weight_decay=0, momentum=0, centered=False):
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= momentum:
+            raise ValueError("Invalid momentum value: {}".format(momentum))
+        if not 0.0 <= weight_decay:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0.0 <= alpha:
+            raise ValueError("Invalid alpha value: {}".format(alpha))
+
+        defaults = dict(lr=lr, momentum=momentum, alpha=alpha, eps=eps, centered=centered, weight_decay=weight_decay)
+        super(RMSpropTFLike, self).__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super(RMSpropTFLike, self).__setstate__(state)
+        for group in self.param_groups:
+            group.setdefault("momentum", 0)
+            group.setdefault("centered", False)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+                grad = p.grad
+                if grad.is_sparse:
+                    raise RuntimeError("RMSpropTF does not support sparse gradients")
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state["step"] = 0
+                    # PyTorch initialized to zeros here
+                    state["square_avg"] = torch.ones_like(p, memory_format=torch.preserve_format)
+                    if group["momentum"] > 0:
+                        state["momentum_buffer"] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    if group["centered"]:
+                        state["grad_avg"] = torch.zeros_like(p, memory_format=torch.preserve_format)
+
+                square_avg = state["square_avg"]
+                alpha = group["alpha"]
+
+                state["step"] += 1
+
+                if group["weight_decay"] != 0:
+                    grad = grad.add(p, alpha=group["weight_decay"])
+
+                square_avg.mul_(alpha).addcmul_(grad, grad, value=1 - alpha)
+
+                if group["centered"]:
+                    grad_avg = state["grad_avg"]
+                    grad_avg.mul_(alpha).add_(grad, alpha=1 - alpha)
+                    # PyTorch added epsilon after square root
+                    # avg = square_avg.addcmul(grad_avg, grad_avg, value=-1).sqrt_().add_(group['eps'])
+                    avg = square_avg.addcmul(grad_avg, grad_avg, value=-1).add_(group["eps"]).sqrt_()
+                else:
+                    # PyTorch added epsilon after square root
+                    # avg = square_avg.sqrt().add_(group['eps'])
+                    avg = square_avg.add(group["eps"]).sqrt_()
+
+                if group["momentum"] > 0:
+                    buf = state["momentum_buffer"]
+                    buf.mul_(group["momentum"]).addcdiv_(grad, avg)
+                    p.add_(buf, alpha=-group["lr"])
+                else:
+                    p.addcdiv_(grad, avg, value=-group["lr"])
+
+        return loss

stable_baselines3/common/torch_layers.py

@@ -74,7 +74,7 @@ def __init__(self, observation_space: gym.spaces.Box, features_dim: int = 512):
             nn.ReLU(),
             nn.Conv2d(32, 64, kernel_size=4, stride=2, padding=0),
             nn.ReLU(),
-            nn.Conv2d(64, 32, kernel_size=3, stride=1, padding=0),
+            nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=0),
             nn.ReLU(),
             nn.Flatten(),
         )