mindspore: maddpg

xuance/mindspore/learners/multi_agent_rl/iddpg_learner.py

@@ -8,7 +8,6 @@
 from xuance.mindspore.utils import clip_grads
 from xuance.common import List
 from argparse import Namespace
-from operator import itemgetter
 
 
 class IDDPG_Learner(LearnerMAS):
@@ -34,6 +33,7 @@ def __init__(self,
         self.gamma = config.gamma
         self.tau = config.tau
         self.mse_loss = MSELoss()
+        # Get gradient function
         self.grad_fn_actor = {key: ms.value_and_grad(self.forward_fn_actor, None,
                                                      self.optimizer[key]['actor'].parameters, has_aux=True)
                               for key in self.model_keys}
@@ -109,9 +109,9 @@ def update(self, sample):
             info.update({
                 f"{key}/learning_rate_actor": learning_rate_actor.asnumpy(),
                 f"{key}/learning_rate_critic": learning_rate_critic.asnumpy(),
-                f"{key}/loss_actor": loss_a.numpy(),
-                f"{key}/loss_critic": loss_c.numpy(),
-                f"{key}/predictQ": q_eval_a.mean().numpy()
+                f"{key}/loss_actor": loss_a.asnumpy(),
+                f"{key}/loss_critic": loss_c.asnumpy(),
+                f"{key}/predictQ": q_eval_a.mean().asnumpy()
             })
 
         self.policy.soft_update(self.tau)

xuance/mindspore/learners/multi_agent_rl/maddpg_learner.py

@@ -5,92 +5,139 @@
 Implementation: MindSpore
 Trick: Parameter sharing for all agents, with agents' one-hot IDs as actor-critic's inputs.
 """
-from xuance.mindspore import ms, Module, Tensor, optim
+from mindspore.nn import MSELoss
+from xuance.mindspore import ms, Module, Tensor, optim, ops
 from xuance.mindspore.learners import LearnerMAS
+from xuance.mindspore.utils import clip_grads
 from xuance.common import List
 from argparse import Namespace
+from operator import itemgetter
 
 
 class MADDPG_Learner(LearnerMAS):
-    class ActorNetWithLossCell(Module):
-        def __init__(self, backbone, n_agents):
-            super(MADDPG_Learner.ActorNetWithLossCell, self).__init__()
-            self._backbone = backbone
-            self._mean = ms.ops.ReduceMean(keep_dims=True)
-            self.n_agents = n_agents
-
-        def construct(self, bs, o, ids, agt_mask):
-            _, actions_eval = self._backbone(o, ids)
-            loss_a = -(self._backbone.critic(o, actions_eval, ids) * agt_mask).sum() / agt_mask.sum()
-            return loss_a
-
-    class CriticNetWithLossCell(Module):
-        def __init__(self, backbone):
-            super(MADDPG_Learner.CriticNetWithLossCell, self).__init__()
-            self._backbone = backbone
-            self._loss = nn.MSELoss()
-
-        def construct(self, o, a_n, ids, agt_mask, tar_q):
-            q_eval = self._backbone.critic(o, a_n, ids)
-            td_error = (q_eval - tar_q) * agt_mask
-            loss_c = (td_error ** 2).sum() / agt_mask.sum()
-            return loss_c
-
     def __init__(self,
                  config: Namespace,
                  model_keys: List[str],
                  agent_keys: List[str],
                  policy: Module):
-        self.gamma = gamma
-        self.tau = config.tau
-        self.sync_frequency = sync_frequency
-        self.mse_loss = nn.MSELoss()
         super(MADDPG_Learner, self).__init__(config, model_keys, agent_keys, policy)
         self.optimizer = {
-            'actor': optimizer[0],
-            'critic': optimizer[1]
-        }
+            key: {
+                'actor': optim.Adam(params=self.policy.parameters_actor[key], lr=self.config.learning_rate_actor,
+                                    eps=1e-5),
+                'critic': optim.Adam(params=self.policy.parameters_critic[key], lr=self.config.learning_rate_critic,
+                                     eps=1e-5)}
+            for key in self.model_keys}
         self.scheduler = {
-            'actor': scheduler[0],
-            'critic': scheduler[1]
-        }
-        # define mindspore trainers
-        self.actor_loss_net = self.ActorNetWithLossCell(policy, self.n_agents)
-        self.actor_train = nn.TrainOneStepCell(self.actor_loss_net, self.optimizer['actor'])
-        self.actor_train.set_train()
-        self.critic_loss_net = self.CriticNetWithLossCell(policy)
-        self.critic_train = nn.TrainOneStepCell(self.critic_loss_net, self.optimizer['critic'])
-        self.critic_train.set_train()
+            key: {'actor': optim.lr_scheduler.LinearLR(self.optimizer[key]['actor'], start_factor=1.0,
+                                                       end_factor=0.5, total_iters=self.config.running_steps),
+                  'critic': optim.lr_scheduler.LinearLR(self.optimizer[key]['critic'], start_factor=1.0,
+                                                        end_factor=0.5, total_iters=self.config.running_steps)}
+            for key in self.model_keys}
+        self.gamma = config.gamma
+        self.tau = config.tau
+        self.mse_loss = MSELoss()
+        # Get gradient function
+        self.grad_fn_actor = {key: ms.value_and_grad(self.forward_fn_actor, None,
+                                                     self.optimizer[key]['actor'].parameters, has_aux=True)
+                              for key in self.model_keys}
+        self.grad_fn_critic = {key: ms.value_and_grad(self.forward_fn_critic, None,
+                                                      self.optimizer[key]['critic'].parameters, has_aux=True)
+                               for key in self.model_keys}
+        self.policy.set_train()
+
+    def forward_fn_actor(self, batch_size, obs, obs_joint, actions, ids, mask_values, agent_key):
+        _, actions_eval = self.policy(observation=obs, agent_ids=ids)
+        if self.use_parameter_sharing:
+            act_eval = actions_eval[agent_key].reshape(batch_size, self.n_agents, -1).reshape(batch_size, -1)
+        else:
+            a_joint = {k: actions_eval[k] if k == agent_key else actions[k] for k in self.agent_keys}
+            act_eval = ops.cat(itemgetter(*self.agent_keys)(a_joint), axis=-1).reshape(batch_size, -1)
+        _, q_policy = self.policy.Qpolicy(joint_observation=obs_joint, joint_actions=act_eval,
+                                          agent_ids=ids, agent_key=agent_key)
+        q_policy_i = q_policy[agent_key].reshape(-1)
+        loss_a = -(q_policy_i * mask_values).sum() / mask_values.sum()
+        return loss_a, q_policy_i
+
+    def forward_fn_critic(self, obs_joint, actions_joint, ids, mask_values, q_target, agent_key):
+        _, q_eval = self.policy.Qpolicy(joint_observation=obs_joint, joint_actions=actions_joint, agent_ids=ids,
+                                        agent_key=agent_key)
+        q_eval_a = q_eval[agent_key].reshape(-1)
+        td_error = (q_eval_a - ops.stop_gradient(q_target)) * mask_values
+        loss_c = (td_error ** 2).sum() / mask_values.sum()
+        return loss_c, q_eval_a
 
     def update(self, sample):
         self.iterations += 1
-        obs = Tensor(sample['obs'])
-        actions = Tensor(sample['actions'])
-        obs_next = Tensor(sample['obs_next'])
-        rewards = Tensor(sample['rewards'])
-        terminals = Tensor(sample['terminals']).view(-1, self.n_agents, 1)
-        agent_mask = Tensor(sample['agent_mask']).view(-1, self.n_agents, 1)
-        batch_size = obs.shape[0]
-        IDs = ops.broadcast_to(self.expand_dims(self.eye(self.n_agents, self.n_agents, ms.float32), 0),
-                               (batch_size, -1, -1))
-        # calculate the loss and train
-        actions_next = self.policy.target_actor(obs_next, IDs)
-        q_next = self.policy.target_critic(obs_next, actions_next, IDs)
-        q_target = rewards + (1 - terminals) * self.args.gamma * q_next
+        info = {}
 
-        # calculate the loss and train
-        loss_a = self.actor_train(batch_size, obs, IDs, agent_mask)
-        loss_c = self.critic_train(obs, actions, IDs, agent_mask, q_target)
-        self.policy.soft_update(self.tau)
+        # prepare training data
+        sample_Tensor = self.build_training_data(sample,
+                                                 use_parameter_sharing=self.use_parameter_sharing,
+                                                 use_actions_mask=False)
+        batch_size = sample_Tensor['batch_size']
+        obs = sample_Tensor['obs']
+        actions = sample_Tensor['actions']
+        obs_next = sample_Tensor['obs_next']
+        rewards = sample_Tensor['rewards']
+        terminals = sample_Tensor['terminals']
+        agent_mask = sample_Tensor['agent_mask']
+        IDs = sample_Tensor['agent_ids']
+        if self.use_parameter_sharing:
+            key = self.model_keys[0]
+            bs = batch_size * self.n_agents
+            obs_joint = obs[key].reshape(batch_size, -1)
+            next_obs_joint = obs_next[key].reshape(batch_size, -1)
+            actions_joint = actions[key].reshape(batch_size, -1)
+            rewards[key] = rewards[key].reshape(batch_size * self.n_agents)
+            terminals[key] = terminals[key].reshape(batch_size * self.n_agents)
+        else:
+            bs = batch_size
+            obs_joint = ops.cat(itemgetter(*self.agent_keys)(obs), axis=-1).reshape(batch_size, -1)
+            next_obs_joint = ops.cat(itemgetter(*self.agent_keys)(obs_next), axis=-1).reshape(batch_size, -1)
+            actions_joint = ops.cat(itemgetter(*self.agent_keys)(actions), axis=-1).reshape(batch_size, -1)
+
+        # get actions
+        _, actions_next = self.policy.Atarget(next_observation=obs_next, agent_ids=IDs)
+        # get values
+        if self.use_parameter_sharing:
+            key = self.model_keys[0]
+            actions_next_joint = actions_next[key].reshape(batch_size, self.n_agents, -1).reshape(batch_size, -1)
+        else:
+            actions_next_joint = ops.cat(itemgetter(*self.model_keys)(actions_next), -1).reshape(batch_size, -1)
+        _, q_next = self.policy.Qtarget(joint_observation=next_obs_joint, joint_actions=actions_next_joint,
+                                        agent_ids=IDs)
 
-        learning_rate_actor = self.scheduler['actor'](self.iterations).asnumpy()
-        learning_rate_critic = self.scheduler['critic'](self.iterations).asnumpy()
+        for key in self.model_keys:
+            mask_values = agent_mask[key]
+            # updata critic
+            q_next_i = q_next[key].reshape(bs)
+            q_target = rewards[key] + (1 - terminals[key]) * self.gamma * q_next_i
+            (loss_c, q_eval_a), grads_critic = self.grad_fn_critic[key](obs_joint, actions_joint, IDs, mask_values,
+                                                                        q_target, key)
+            if self.use_grad_clip:
+                grads_critic = clip_grads(grads_critic, Tensor(-self.grad_clip_norm), Tensor(self.grad_clip_norm))
+            self.optimizer[key]['critic'](grads_critic)
 
-        info = {
-            "learning_rate_actor": learning_rate_actor,
-            "learning_rate_critic": learning_rate_critic,
-            "loss_actor": loss_a.asnumpy(),
-            "loss_critic": loss_c.asnumpy()
-        }
+            # update actor
+            (loss_a, _), grads_actor = self.grad_fn_actor[key](batch_size, obs, obs_joint, actions, IDs, mask_values,
+                                                               key)
+            if self.use_grad_clip:
+                grads_actor = clip_grads(grads_actor, Tensor(-self.grad_clip_norm), Tensor(self.grad_clip_norm))
+            self.optimizer[key]['actor'](grads_actor)
 
+            self.scheduler[key]['actor'].step()
+            self.scheduler[key]['critic'].step()
+            learning_rate_actor = self.scheduler[key]['actor'].get_last_lr()[0]
+            learning_rate_critic = self.scheduler[key]['critic'].get_last_lr()[0]
+
+            info.update({
+                f"{key}/learning_rate_actor": learning_rate_actor.asnumpy(),
+                f"{key}/learning_rate_critic": learning_rate_critic.asnumpy(),
+                f"{key}/loss_actor": loss_a.asnumpy(),
+                f"{key}/loss_critic": loss_c.asnumpy(),
+                f"{key}/predictQ": q_eval_a.mean().asnumpy()
+            })
+
+        self.policy.soft_update(self.tau)
         return info

xuance/mindspore/policies/deterministic_marl.py

@@ -872,10 +872,10 @@ def Qpolicy(self, joint_observation: Tensor, joint_actions: Tensor,
         for key in agent_list:
             if self.use_parameter_sharing:
                 if self.use_rnn:
-                    joint_rep_out = outputs[key]['state'].unsqueeze(1).expand(-1, self.n_agents, -1, -1)
+                    joint_rep_out = outputs[key]['state'].unsqueeze(1).broadcast_to((-1, self.n_agents, -1, -1))
                     joint_rep_out = joint_rep_out.reshape(bs, seq_len, -1)
                 else:
-                    joint_rep_out = outputs[key]['state'].unsqueeze(1).expand(-1, self.n_agents, -1)
+                    joint_rep_out = outputs[key]['state'].unsqueeze(1).broadcast_to((-1, self.n_agents, -1))
                     joint_rep_out = joint_rep_out.reshape(bs, -1)
                 critic_in = ops.cat([joint_rep_out, agent_ids], axis=-1)
             else:
@@ -921,10 +921,10 @@ def Qtarget(self, joint_observation: Tensor, joint_actions: Tensor,
         for key in agent_list:
             if self.use_parameter_sharing:
                 if self.use_rnn:
-                    joint_rep_out = outputs[key]['state'].unsqueeze(1).expand(-1, self.n_agents, -1, -1)
+                    joint_rep_out = outputs[key]['state'].unsqueeze(1).broadcast_to((-1, self.n_agents, -1, -1))
                     joint_rep_out = joint_rep_out.reshape(bs, seq_len, -1)
                 else:
-                    joint_rep_out = outputs[key]['state'].unsqueeze(1).expand(-1, self.n_agents, -1)
+                    joint_rep_out = outputs[key]['state'].unsqueeze(1).broadcast_to((-1, self.n_agents, -1))
                     joint_rep_out = joint_rep_out.reshape(bs, -1)
                 critic_in = ops.cat([joint_rep_out, agent_ids], axis=-1)
             else:
@@ -1031,13 +1031,13 @@ def Qpolicy(self, joint_observation: Tensor, joint_actions: Tensor,
         for key in agent_list:
             if self.use_parameter_sharing:
                 if self.use_rnn:
-                    joint_rep_out_A = outputs_A[key]['state'].unsqueeze(1).expand(-1, self.n_agents, -1, -1)
-                    joint_rep_out_B = outputs_B[key]['state'].unsqueeze(1).expand(-1, self.n_agents, -1, -1)
+                    joint_rep_out_A = outputs_A[key]['state'].unsqueeze(1).broadcast_to((-1, self.n_agents, -1, -1))
+                    joint_rep_out_B = outputs_B[key]['state'].unsqueeze(1).broadcast_to((-1, self.n_agents, -1, -1))
                     joint_rep_out_A = joint_rep_out_A.reshape(bs, seq_len, -1)
                     joint_rep_out_B = joint_rep_out_B.reshape(bs, seq_len, -1)
                 else:
-                    joint_rep_out_A = outputs_A[key]['state'].unsqueeze(1).expand(-1, self.n_agents, -1)
-                    joint_rep_out_B = outputs_B[key]['state'].unsqueeze(1).expand(-1, self.n_agents, -1)
+                    joint_rep_out_A = outputs_A[key]['state'].unsqueeze(1).broadcast_to((-1, self.n_agents, -1))
+                    joint_rep_out_B = outputs_B[key]['state'].unsqueeze(1).broadcast_to((-1, self.n_agents, -1))
                     joint_rep_out_A = joint_rep_out_A.reshape(bs, -1)
                     joint_rep_out_B = joint_rep_out_B.reshape(bs, -1)
                 critic_in_A = ops.cat([joint_rep_out_A, agent_ids], axis=-1)
@@ -1091,13 +1091,13 @@ def Qtarget(self, joint_observation: Tensor, joint_actions: Tensor,
         for key in agent_list:
             if self.use_parameter_sharing:
                 if self.use_rnn:
-                    joint_rep_out_A = outputs_A[key]['state'].unsqueeze(1).expand(-1, self.n_agents, -1, -1)
-                    joint_rep_out_B = outputs_B[key]['state'].unsqueeze(1).expand(-1, self.n_agents, -1, -1)
+                    joint_rep_out_A = outputs_A[key]['state'].unsqueeze(1).broadcast_to((-1, self.n_agents, -1, -1))
+                    joint_rep_out_B = outputs_B[key]['state'].unsqueeze(1).broadcast_to((-1, self.n_agents, -1, -1))
                     joint_rep_out_A = joint_rep_out_A.reshape(bs, seq_len, -1)
                     joint_rep_out_B = joint_rep_out_B.reshape(bs, seq_len, -1)
                 else:
-                    joint_rep_out_A = outputs_A[key]['state'].unsqueeze(1).expand(-1, self.n_agents, -1)
-                    joint_rep_out_B = outputs_B[key]['state'].unsqueeze(1).expand(-1, self.n_agents, -1)
+                    joint_rep_out_A = outputs_A[key]['state'].unsqueeze(1).broadcast_to((-1, self.n_agents, -1))
+                    joint_rep_out_B = outputs_B[key]['state'].unsqueeze(1).broadcast_to((-1, self.n_agents, -1))
                     joint_rep_out_A = joint_rep_out_A.reshape(bs, -1)
                     joint_rep_out_B = joint_rep_out_B.reshape(bs, -1)
                 critic_in_A = ops.cat([joint_rep_out_A, agent_ids], axis=-1)