Changing CLUB loss and Tensor stacking

DPI/models.py

@@ -194,6 +194,27 @@ class TransitionModel(nn.Module):
         prior = {"mean": state_prior_mean, "std": state_prior_std, "sample": sample_state_prior, "history": history, "distribution": state_prior_dist}
         return prior
     
+    def stack_states(self, states, dim=0):
+        
+        s = dict(
+            mean = torch.stack([state['mean'] for state in states], dim=dim),
+            std  = torch.stack([state['std'] for state in states], dim=dim),
+            sample = torch.stack([state['sample'] for state in states], dim=dim),
+            history = torch.stack([state['history'] for state in states], dim=dim),)
+        dist = dict(distribution = [state['distribution'] for state in states])
+        s.update(dist)
+        return s
+    
+    def imagine_rollout(self, state, action, history, horizon):
+        imagined_priors = []
+        for i in range(horizon):
+            prior = self.imagine_step(state, action, history)
+            state = prior["sample"]
+            history = prior["history"]
+            imagined_priors.append(prior)
+        imagined_priors = self.stack_states(imagined_priors, dim=0)
+        return imagined_priors
+    
     def reparemeterize(self, mean, std):
         eps = torch.randn_like(std)
         return mean + eps * std
@@ -227,40 +248,6 @@ class TanhBijector(torch.distributions.Transform):
         return 2.0 * (torch.log(torch.tensor([2.0])) - x - F.softplus(-2.0 * x))
      
 
-class CLUBSample(nn.Module):  # Sampled version of the CLUB estimator
-    def __init__(self, x_dim, y_dim, hidden_size):
-        super(CLUBSample, self).__init__()
-        self.p_mu = nn.Sequential(
-                    nn.Linear(x_dim, hidden_size//2),
-                    nn.ReLU(),
-                    nn.Linear(hidden_size//2, hidden_size//2),
-                    nn.ReLU(),
-                    nn.Linear(hidden_size//2, y_dim)
-                    )
-
-        self.p_logvar = nn.Sequential(
-                        nn.Linear(x_dim, hidden_size//2),
-                        nn.ReLU(),
-                        nn.Linear(hidden_size//2, hidden_size//2),
-                        nn.ReLU(),
-                        nn.Linear(hidden_size//2, y_dim),
-                        nn.Tanh()
-                        )
-
-    def get_mu_logvar(self, x_samples):
-        mu = self.p_mu(x_samples)
-        logvar = self.p_logvar(x_samples)
-        return mu, logvar
-    
-    def loglikeli(self, x_samples, y_samples):
-        mu, logvar = self.get_mu_logvar(x_samples)
-        return (-(mu - y_samples)**2 /logvar.exp()-logvar).sum(dim=1).mean(dim=0)
-
-    def forward(self, x_samples, y_samples):
-        mu, logvar = self.get_mu_logvar(x_samples)
-        return - self.loglikeli(x_samples, y_samples)
-    
-
 class ProjectionHead(nn.Module):
     def __init__(self, state_size, action_size, hidden_size):
         super(ProjectionHead, self).__init__()
@@ -295,3 +282,43 @@ class ContrastiveHead(nn.Module):
         logits = logits - torch.max(logits, 1)[0][:, None]
         logits = logits * self.temperature
         return logits
+    
+
+class CLUBSample(nn.Module):  # Sampled version of the CLUB estimator
+    def __init__(self, last_states, current_states, negative_current_states, predicted_current_states):
+        super(CLUBSample, self).__init__()
+        self.last_states = last_states
+        self.current_states = current_states
+        self.negative_current_states = negative_current_states
+        self.predicted_current_states = predicted_current_states
+
+    def get_mu_var_samples(self, state_dict):
+        dist = state_dict["distribution"]
+        sample = dist.sample()  # Use state_dict["sample"] if you want to use the same sample for all the losses
+        mu = dist.mean
+        var = dist.variance
+        return mu, var, sample
+        
+    def loglikeli(self):
+        _, _, pred_sample = self.get_mu_var_samples(self.predicted_current_states)
+        mu_curr, var_curr, _ = self.get_mu_var_samples(self.current_states)
+        logvar_curr = torch.log(var_curr)    
+        return (-(mu_curr - pred_sample)**2 /var_curr-logvar_curr).sum(dim=1).mean(dim=0)
+
+    def forward(self):
+        _, _, pred_sample = self.get_mu_var_samples(self.predicted_current_states)
+        mu_curr, var_curr, _ = self.get_mu_var_samples(self.current_states)
+        mu_neg, var_neg, _ = self.get_mu_var_samples(self.negative_current_states)
+        
+        pos = (-(mu_curr - pred_sample)**2 /var_curr).sum(dim=1).mean(dim=0)
+        neg = (-(mu_neg - pred_sample)**2 /var_neg).sum(dim=1).mean(dim=0)
+        upper_bound = pos - neg
+
+        return upper_bound/2
+
+    def learning_loss(self):
+        return - self.loglikeli()
+    
+
+if "__name__ == __main__":
+    pass

DPI/train.py

@@ -16,6 +16,8 @@ from logger import Logger
 from video import VideoRecorder
 from dmc2gym.wrappers import set_global_var
 
+import torchvision.transforms as T
+
 #from agent.baseline_agent import BaselineAgent
 #from agent.bisim_agent import BisimAgent
 #from agent.deepmdp_agent import DeepMDPAgent
@@ -31,7 +33,7 @@ def parse_args():
     parser.add_argument('--image_size', default=84, type=int)
     parser.add_argument('--channels', default=3, type=int)
     parser.add_argument('--action_repeat', default=1, type=int)
-    parser.add_argument('--frame_stack', default=4, type=int)
+    parser.add_argument('--frame_stack', default=3, type=int)
     parser.add_argument('--resource_files', type=str)
     parser.add_argument('--eval_resource_files', type=str)
     parser.add_argument('--img_source', default=None, type=str, choices=['color', 'noise', 'images', 'video', 'none'])
@@ -39,18 +41,18 @@ def parse_args():
     parser.add_argument('--high_noise', action='store_true')
     # replay buffer
     parser.add_argument('--replay_buffer_capacity', default=50000, type=int)   #50000
-    parser.add_argument('--episode_length', default=50, type=int)
+    parser.add_argument('--episode_length', default=51, type=int)
     # train
     parser.add_argument('--agent', default='dpi', type=str, choices=['baseline', 'bisim', 'deepmdp', 'db', 'dpi', 'rpc'])
-    parser.add_argument('--init_steps', default=1000, type=int)
+    parser.add_argument('--init_steps', default=10000, type=int)
-    parser.add_argument('--num_train_steps', default=1000, type=int)
+    parser.add_argument('--num_train_steps', default=10000, type=int)
-    parser.add_argument('--batch_size', default=200, type=int)  #512
+    parser.add_argument('--batch_size', default=20, type=int)  #512
     parser.add_argument('--state_size', default=256, type=int)
     parser.add_argument('--hidden_size', default=128, type=int)
     parser.add_argument('--history_size', default=128, type=int)
     parser.add_argument('--num-units', type=int, default=200, help='num hidden units for reward/value/discount models')
     parser.add_argument('--load_encoder', default=None, type=str)
-    parser.add_argument('--imagination_horizon', default=15, type=str)
+    parser.add_argument('--imagine_horizon', default=15, type=str)
     # eval
     parser.add_argument('--eval_freq', default=10, type=int)  # TODO: master had 10000
     parser.add_argument('--num_eval_episodes', default=20, type=int)
@@ -113,6 +115,7 @@ class DPI:
 
         # environment setup
         self.env = make_env(self.args)
+        #self.args.seed = np.random.randint(0, 1000)
         self.env.seed(self.args.seed)
 
         # noiseless environment setup
@@ -190,87 +193,123 @@ class DPI:
 
     def collect_sequences(self, episodes):
         obs = self.env.reset()
-        obs_clean = self.env_clean.reset()
+        #obs_clean = self.env_clean.reset()
         done = False
 
         #video = VideoRecorder(self.video_dir if args.save_video else None, resource_files=args.resource_files)
         for episode_count in tqdm.tqdm(range(episodes), desc='Collecting episodes'):
             if args.save_video:
                 self.env.video.init(enabled=True)
-                self.env_clean.video.init(enabled=True)
+                #self.env_clean.video.init(enabled=True)
 
             for i in range(self.args.episode_length):
+                
                 action = self.env.action_space.sample()
 
-                next_obs, _, done, _ = self.env.step(action)
+                next_obs, rew, done, _ = self.env.step(action)
-                next_obs_clean, _, done, _ = self.env_clean.step(action)
+                #next_obs_clean, _, done, _ = self.env_clean.step(action)
 
                 self.data_buffer.add(obs, action, next_obs, episode_count+1, done)
-                self.data_buffer_clean.add(obs_clean, action, next_obs_clean, episode_count+1, done)
+                #self.data_buffer_clean.add(obs_clean, action, next_obs_clean, episode_count+1, done)
                 
                 if args.save_video:
-                    self.env.video.record(self.env_clean)
+                    self.env.video.record(self.env)
-                    self.env_clean.video.record(self.env_clean)
+                    #self.env_clean.video.record(self.env_clean)
                 
-                if done:
+                if done or i == self.args.episode_length-1:
                     obs = self.env.reset()
-                    obs_clean = self.env_clean.reset()
+                    #obs_clean = self.env_clean.reset()
                     done=False
                 else:
                     obs = next_obs
-                    obs_clean = next_obs_clean
+                    #obs_clean = next_obs_clean
             if args.save_video:
                 self.env.video.save('noisy/%d.mp4' % episode_count)
-                self.env_clean.video.save('clean/%d.mp4' % episode_count)
+                #self.env_clean.video.save('clean/%d.mp4' % episode_count)
         print("Collected {} random episodes".format(episode_count+1))
 
     def train(self):
         # collect experience
         self.collect_sequences(self.args.batch_size)
 
-        # Group observations and next_observations by steps
+        # Group observations and next_observations by steps from past to present
-        observations = torch.Tensor(self.data_buffer.group_steps(self.data_buffer,"observations")).float()
+        last_observations = torch.Tensor(self.data_buffer.group_steps(self.data_buffer,"observations")).float()[:self.args.episode_length-1]
-        next_observations = torch.Tensor(self.data_buffer.group_steps(self.data_buffer,"next_observations")).float()
+        current_observations = torch.Tensor(self.data_buffer.group_steps(self.data_buffer,"next_observations")).float()[:self.args.episode_length-1]
-        actions = torch.Tensor(self.data_buffer.group_steps(self.data_buffer,"actions",obs=False)).float()
+        next_observations = torch.Tensor(self.data_buffer.group_steps(self.data_buffer,"next_observations")).float()[1:]
+        actions = torch.Tensor(self.data_buffer.group_steps(self.data_buffer,"actions",obs=False)).float()[:self.args.episode_length-1]
+        next_actions = torch.Tensor(self.data_buffer.group_steps(self.data_buffer,"actions",obs=False)).float()[1:]
 
         # Initialize transition model states
         self.transition_model.init_states(self.args.batch_size, device="cpu") # (N,128) 
         self.history = self.transition_model.prev_history # (N,128)
 
         # Train encoder
-        previous_information_loss = 0
+        total_ub_loss = 0
-        previous_encoder_loss = 0
+        total_encoder_loss = 0
-        for i in range(self.args.episode_length):
+        for i in range(self.args.episode_length-1):
-            # Encode observations and next_observations
+            if i > 0:
-            self.states_dist = self.obs_encoder(observations[i])
+                # Encode observations and next_observations
-            self.next_states_dist = self.obs_encoder(next_observations[i])
+                self.last_states_dict = self.obs_encoder(last_observations[i])
+                self.current_states_dict = self.obs_encoder(current_observations[i])
+                self.next_states_dict = self.obs_encoder_momentum(next_observations[i])
+                self.action = actions[i] # (N,6)
+                history = self.transition_model.prev_history
                 
-            # Sample states and next_states
+                # Encode negative observations
-            self.states = self.states_dist["sample"] # (N,128)
+                idx = torch.randperm(current_observations[i].shape[0]) # random permutation on batch
-            self.next_states = self.next_states_dist["sample"] # (N,128)
+                random_time_index = torch.randint(0, self.args.episode_length-2, (1,)).item() # random time index
-            self.actions = actions[i] # (N,6)
+                negative_current_observations = current_observations[random_time_index][idx]
+                self.negative_current_states_dict = self.obs_encoder(negative_current_observations)
 
-            # Calculate upper bound loss
+                # Predict current state from past state with transition model
-            past_latent_loss = previous_information_loss + self._upper_bound_minimization(self.states, self.next_states)
+                last_states_sample = self.last_states_dict["sample"]
+                predicted_current_state_dict = self.transition_model.imagine_step(last_states_sample, self.action, self.history)
+                self.history = predicted_current_state_dict["history"]
                 
-            # Calculate encoder loss
+                # Calculate upper bound loss
-            past_encoder_loss = previous_encoder_loss + self._past_encoder_loss(self.states, self.next_states, 
+                ub_loss = self._upper_bound_minimization(self.last_states_dict, 
-                                                        self.states_dist, self.next_states_dist,
+                                                         self.current_states_dict,
-                                                        self.actions, self.history, i)
+                                                         self.negative_current_states_dict,
+                                                         predicted_current_state_dict
+                                                         )
+                
+                # Calculate encoder loss
+                encoder_loss = self._past_encoder_loss(self.current_states_dict,
+                                                       predicted_current_state_dict)
+
+                total_ub_loss += ub_loss
+                total_encoder_loss += encoder_loss
+
+                imagine_horizon = np.minimum(self.args.imagine_horizon, self.args.episode_length-1-i)
+                imagined_rollout = self.transition_model.imagine_rollout(self.current_states_dict["sample"], self.action, self.history, imagine_horizon)
+                print(imagine_horizon)
+                #exit()
+
+                #print(total_ub_loss, total_encoder_loss)
 
 
-            print(past_encoder_loss, past_latent_loss)
 
-            previous_information_loss = past_latent_loss
+    def _upper_bound_minimization(self, last_states, current_states, negative_current_states, predicted_current_states):
-            previous_encoder_loss = past_encoder_loss
+        club_sample = CLUBSample(last_states,
-
+                                 current_states,
-    def _upper_bound_minimization(self, states, next_states):
+                                 negative_current_states,
-        club_sample = CLUBSample(self.args.state_size,
+                                 predicted_current_states)
-                                 self.args.state_size,
+        club_loss = club_sample()
-                                 self.args.hidden_size)
-        club_loss = club_sample(states, next_states)
         return club_loss
     
+    def _past_encoder_loss(self, curr_states_dict, predicted_curr_states_dict):
+        # current state distribution
+        curr_states_dist = curr_states_dict["distribution"]
+
+        # predicted current state distribution
+        predicted_curr_states_dist = predicted_curr_states_dict["distribution"]
+
+        # KL divergence loss
+        loss = torch.distributions.kl.kl_divergence(curr_states_dist, predicted_curr_states_dist).mean()
+
+        return loss
+
+    """
     def _past_encoder_loss(self, states, next_states, states_dist, next_states_dist, actions, history, step):
         # Imagine next state
         if step == 0:
@@ -287,6 +326,21 @@ class DPI:
         loss = torch.distributions.kl.kl_divergence(imagined_next_states_dist, next_states_dist["distribution"]).mean()
 
         return loss
+    """
+    
+    def get_features(self, x, momentum=False):
+        if self.aug:
+            x = T.RandomCrop((80, 80))(x)  # (None,80,80,4)
+            x = T.functional.pad(x, (4, 4, 4, 4), "symmetric")  # (None,88,88,4)
+            x = T.RandomCrop((84, 84))(x)  # (None,84,84,4)
+
+        with torch.no_grad():
+            x = (x.float() - self.ob_mean) / self.ob_std
+            if momentum:
+                x = self.obs_encoder(x).detach()
+            else:
+                x = self.obs_encoder_momentum(x)
+        return x
 
 if __name__ == '__main__':
     args = parse_args()

DPI/utils.py

@@ -161,11 +161,11 @@ class ReplayBuffer:
         non_zero_indices = np.nonzero(buffer.episode_count)[0]
         variable = variable[non_zero_indices]
         if obs:
-            variable = variable.reshape(self.args.episode_length, self.args.batch_size,
+            variable = variable.reshape(self.args.batch_size, self.args.episode_length,
-                                                        self.args.frame_stack*self.args.channels,
+                                        self.args.frame_stack*self.args.channels,
-                                                        self.args.image_size,self.args.image_size)
+                                        self.args.image_size,self.args.image_size).transpose(1, 0, 2, 3, 4)
         else:
-            variable = variable.reshape(self.args.episode_length, self.args.batch_size,-1)
+            variable = variable.reshape(self.args.batch_size, self.args.episode_length, -1).transpose(1, 0, 2)
         return variable
 
     def transform_grouped_steps(self, variable):