Correct history with detach

DPI/models.py

@@ -19,7 +19,7 @@ class ObservationEncoder(nn.Module):
             input_channels = obs_shape[0] if i == 0 else output_channels
             output_channels = num_filters * (2 ** i)
             layers.append(nn.Conv2d(in_channels=input_channels, out_channels= output_channels, kernel_size=4, stride=2))
-            layers.append(nn.ReLU())
+            layers.append(nn.LeakyReLU())
 
         self.convs = nn.Sequential(*layers)
 
@@ -196,7 +196,8 @@ class TransitionModel(nn.Module):
 
     def imagine_step(self, prev_state, prev_action, prev_history):
         state_action = self.act_fn(self.fc_state_action(torch.cat([prev_state, prev_action], dim=-1)))
-        history = self.history_cell(torch.cat([state_action, prev_history], dim=-1), prev_history)
+        prev_hist = prev_history.detach()
+        history = self.history_cell(torch.cat([state_action, prev_hist], dim=-1), prev_hist)
 
         state_prior = self.fc_state_prior(torch.cat([history, prev_state, prev_action], dim=-1))
         state_prior_mean, state_prior_std = torch.chunk(state_prior, 2, dim=-1)

DPI/train.py

@@ -10,14 +10,17 @@ import dmc2gym
 import tqdm
 import wandb
 import utils
-from utils import ReplayBuffer, make_env, save_image
+from utils import ReplayBuffer, FreezeParameters, make_env, soft_update_params, save_image
-from models import ObservationEncoder, ObservationDecoder, TransitionModel, CLUBSample, Actor, ValueModel, RewardModel
+from models import ObservationEncoder, ObservationDecoder, TransitionModel, Actor, ValueModel, RewardModel, ProjectionHead, ContrastiveHead, CLUBSample
 from logger import Logger
 from video import VideoRecorder
 from dmc2gym.wrappers import set_global_var
 
+import torch.nn as nn
+import torch.nn.functional as F
 import torchvision.transforms as T
 
+
 #from agent.baseline_agent import BaselineAgent
 #from agent.bisim_agent import BisimAgent
 #from agent.deepmdp_agent import DeepMDPAgent
@@ -53,23 +56,27 @@ def parse_args():
     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('--imagine_horizon', default=15, type=str)
+    parser.add_argument('--grad_clip_norm', type=float, default=100.0, help='Gradient clipping norm')
     # eval
     parser.add_argument('--eval_freq', default=10, type=int)  # TODO: master had 10000
     parser.add_argument('--num_eval_episodes', default=20, type=int)
-    # critic
+    # value
-    parser.add_argument('--critic_lr', default=1e-3, type=float)
+    parser.add_argument('--value_lr', default=1e-4, type=float)
-    parser.add_argument('--critic_beta', default=0.9, type=float)
+    parser.add_argument('--value_beta', default=0.9, type=float)
-    parser.add_argument('--critic_tau', default=0.005, type=float)
+    parser.add_argument('--value_tau', default=0.005, type=float)
-    parser.add_argument('--critic_target_update_freq', default=2, type=int)
+    parser.add_argument('--value_target_update_freq', default=2, type=int)
+    # reward
+    parser.add_argument('--reward_lr', default=1e-4, type=float)
     # actor
-    parser.add_argument('--actor_lr', default=1e-3, type=float)
+    parser.add_argument('--actor_lr', default=1e-4, type=float)
     parser.add_argument('--actor_beta', default=0.9, type=float)
     parser.add_argument('--actor_log_std_min', default=-10, type=float)
     parser.add_argument('--actor_log_std_max', default=2, type=float)
     parser.add_argument('--actor_update_freq', default=2, type=int)
-    # encoder/decoder
+    # world/encoder/decoder
     parser.add_argument('--encoder_type', default='pixel', type=str, choices=['pixel', 'pixelCarla096', 'pixelCarla098', 'identity'])
     parser.add_argument('--encoder_feature_dim', default=50, type=int)
+    parser.add_argument('--world_model_lr', default=1e-3, type=float)
     parser.add_argument('--encoder_lr', default=1e-3, type=float)
     parser.add_argument('--encoder_tau', default=0.005, type=float)
     parser.add_argument('--encoder_stride', default=1, type=int)
@@ -79,6 +86,7 @@ def parse_args():
     parser.add_argument('--decoder_weight_lambda', default=0.0, type=float)
     parser.add_argument('--num_layers', default=4, type=int)
     parser.add_argument('--num_filters', default=32, type=int)
+    parser.add_argument('--aug', action='store_true')
     # sac
     parser.add_argument('--discount', default=0.99, type=float)
     parser.add_argument('--init_temperature', default=0.01, type=float)
@@ -154,6 +162,7 @@ class DPI:
         self.build_models(use_saved=False, saved_model_dir=self.model_dir)
 
     def build_models(self, use_saved, saved_model_dir=None):
+        # World Models
         self.obs_encoder = ObservationEncoder(
                             obs_shape=(self.args.frame_stack*self.args.channels,self.args.image_size,self.args.image_size), # (12,84,84)
                             state_size=self.args.state_size # 128
@@ -176,12 +185,14 @@ class DPI:
                             history_size=self.args.history_size, # 128
                             )
         
-        self.action_model = Actor(
+        # Actor Model
+        self.actor_model = Actor(
                             state_size=self.args.state_size, # 128
                             hidden_size=self.args.hidden_size, # 256, 
                             action_size=self.env.action_space.shape[0], # 6
                             )
         
+        # Value Models
         self.value_model = ValueModel(
                             state_size=self.args.state_size, # 128
                             hidden_size=self.args.hidden_size, # 256
@@ -197,12 +208,38 @@ class DPI:
                             hidden_size=self.args.hidden_size, # 256
                             )
         
-        # model parameters
+        # Contrastive Models
-        self.model_parameters = list(self.obs_encoder.parameters()) + list(self.obs_encoder_momentum.parameters()) + \
+        self.prjoection_head = ProjectionHead(
-                                list(self.obs_decoder.parameters()) + list(self.transition_model.parameters())
+                            state_size=self.args.state_size, # 128
+                            action_size=self.env.action_space.shape[0], # 6
+                            hidden_size=self.args.hidden_size, # 256
+                            )
         
-        # optimizer
+        self.prjoection_head_momentum = ProjectionHead(
-        self.optimizer = torch.optim.Adam(self.model_parameters, lr=self.args.encoder_lr)
+                            state_size=self.args.state_size, # 128
+                            action_size=self.env.action_space.shape[0], # 6
+                            hidden_size=self.args.hidden_size, # 256
+                            )
+        
+        self.contrastive_head = ContrastiveHead(
+                            hidden_size=self.args.hidden_size, # 256
+                            )
+
+
+        # model parameters
+        self.world_model_parameters = list(self.obs_encoder.parameters()) + list(self.obs_decoder.parameters()) + \
+                                list(self.value_model.parameters())  + list(self.transition_model.parameters()) + \
+                                list(self.prjoection_head.parameters())
+
+        # optimizers
+        self.world_model_opt = torch.optim.Adam(self.world_model_parameters, self.args.world_model_lr)
+        self.value_opt = torch.optim.Adam(self.value_model.parameters(), self.args.value_lr)
+        self.actor_opt = torch.optim.Adam(self.actor_model.parameters(), self.args.actor_lr)
+
+        # Create Modules
+        self.world_model_modules = [self.obs_encoder, self.obs_decoder, self.value_model, self.transition_model, self.prjoection_head]
+        self.value_modules = [self.value_model]
+        self.actor_modules = [self.actor_model]
 
         if use_saved:
             self._use_saved_models(saved_model_dir)
@@ -214,6 +251,8 @@ class DPI:
 
     def collect_sequences(self, episodes):
         obs = self.env.reset()
+        self.ob_mean = np.mean(obs, 0).astype(np.float32)
+        self.ob_std = np.std(obs, 0).mean().astype(np.float32)
         #obs_clean = self.env_clean.reset()
         done = False
 
@@ -265,14 +304,15 @@ class DPI:
         self.history = self.transition_model.prev_history # (N,128)
 
         # Train encoder
-        total_ub_loss = 0
+        step = 0
-        total_encoder_loss = 0
+        total_steps = 10000
+        while step < total_steps:
             for i in range(self.args.episode_length-1):
                 if i > 0:
                     # Encode observations and next_observations
-                self.last_states_dict = self.obs_encoder(last_observations[i])
+                    self.last_states_dict = self.get_features(last_observations[i])
-                self.current_states_dict = self.obs_encoder(current_observations[i])
+                    self.current_states_dict = self.get_features(current_observations[i])
-                self.next_states_dict = self.obs_encoder_momentum(next_observations[i])
+                    self.next_states_dict = self.get_features(next_observations[i], momentum=True)
                     self.action = actions[i] # (N,6)
                     history = self.transition_model.prev_history
                     
@@ -287,6 +327,8 @@ class DPI:
                     predicted_current_state_dict = self.transition_model.imagine_step(last_states_sample, self.action, self.history)
                     self.history = predicted_current_state_dict["history"]
 
+
+                    
                     # Calculate upper bound loss
                     ub_loss = self._upper_bound_minimization(self.last_states_dict, 
                                                             self.current_states_dict,
@@ -298,12 +340,45 @@ class DPI:
                     encoder_loss = self._past_encoder_loss(self.current_states_dict,
                                                         predicted_current_state_dict)
 
-                total_ub_loss += ub_loss
+                    #total_ub_loss += ub_loss
-                total_encoder_loss += encoder_loss
+                    #total_encoder_loss += encoder_loss
                     
+                    # contrastive projection
+                    vec_anchor = predicted_current_state_dict["sample"]
+                    vec_positive = self.next_states_dict["sample"].detach()
+                    z_anchor = self.prjoection_head(vec_anchor, self.action)
+                    z_positive = self.prjoection_head_momentum(vec_positive, next_actions[i]).detach()
+
+                    # contrastive loss
+                    logits = self.contrastive_head(z_anchor, z_positive)
+                    labels = labels = torch.arange(logits.shape[0]).long()
+                    lb_loss = F.cross_entropy(logits, labels)
+                    
+
+                    # update models
+                    world_model_loss = encoder_loss + 1e-1 * ub_loss + lb_loss #1e-1 * ub_loss + 1e-5 * encoder_loss + 1e-1 * lb_loss
+                    print("ub_loss: {:.4f}, encoder_loss: {:.4f}, lb_loss: {:.4f}".format(ub_loss, encoder_loss, lb_loss))
+                    print("world_model_loss: {:.4f}".format(world_model_loss))
+                    self.world_model_opt.zero_grad()
+                    world_model_loss.backward()
+                    nn.utils.clip_grad_norm_(self.world_model_parameters, self.args.grad_clip_norm)
+                    self.world_model_opt.step()
+
+                    
+                    # behaviour learning
+                    with FreezeParameters(self.world_model_modules):
                         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)
+                        imagined_rollout = self.transition_model.imagine_rollout(self.current_states_dict["sample"].detach(), 
+                                                                                self.action, self.history.detach(), 
+                                                                                imagine_horizon)
+                        print(imagined_rollout["sample"].shape, imagined_rollout["distribution"][0].sample().shape)
+                    #exit()
                     
+                    step += 1
+
+                    if step>total_steps:
+                        print("Training finished")
+                        break
                     #exit()
 
                     #print(total_ub_loss, total_encoder_loss)
@@ -315,7 +390,7 @@ class DPI:
                                  current_states,
                                  negative_current_states,
                                  predicted_current_states)
-        club_loss = club_sample()
+        club_loss = club_sample.loglikeli()
         return club_loss
     
     def _past_encoder_loss(self, curr_states_dict, predicted_curr_states_dict):
@@ -325,42 +400,27 @@ class DPI:
         # 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:
-            actions = torch.zeros(self.args.batch_size, self.env.action_space.shape[0]).float() # Zero action for first step
-            imagined_next_states = self.transition_model.imagine_step(states, actions, history)
-            self.history = imagined_next_states["history"]
-        else:
-            imagined_next_states = self.transition_model.imagine_step(states, actions, self.history) # (N,128)
-
-        # State Distribution
-        imagined_next_states_dist = imagined_next_states["distribution"]
-
-        # KL divergence loss
-        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:
+        import torchvision.transforms.functional as fn
+        x = x/255.0 - 0.5  # Preprocessing
+
+        if self.args.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)
+            else:
+                x = self.obs_encoder(x)
         return x
 
 if __name__ == '__main__':

DPI/utils.py

@@ -17,6 +17,7 @@ import dmc2gym
 
 import cv2
 from PIL import Image
+from typing import Iterable
 
 
 class eval_mode(object):
@@ -197,6 +198,12 @@ def make_env(args):
         )
     return env
 
+def soft_update_params(net, target_net, tau):
+    for param, target_param in zip(net.parameters(), target_net.parameters()):
+        target_param.data.copy_(
+            tau * param.data + (1 - tau) * target_param.data
+        )
+
 def save_image(array, filename):
     array = array.transpose(1, 2, 0)
     array = (array * 255).astype(np.uint8)
@@ -257,3 +264,39 @@ class CorruptVideos:
                     if delete:
                         self._delete_corrupt_video(filepath)
                         print(f"Deleted {filepath}")
+
+
+def get_parameters(modules: Iterable[nn.Module]):
+    """
+    Given a list of torch modules, returns a list of their parameters.
+    :param modules: iterable of modules
+    :returns: a list of parameters
+    """
+    model_parameters = []
+    for module in modules:
+        model_parameters += list(module.parameters())
+    return model_parameters
+
+class FreezeParameters:
+    def __init__(self, modules: Iterable[nn.Module]):
+        """
+        Context manager to locally freeze gradients.
+        In some cases with can speed up computation because gradients aren't calculated for these listed modules.
+        example:
+        ```
+        with FreezeParameters([module]):
+          output_tensor = module(input_tensor)
+        ```
+        :param modules: iterable of modules. used to call .parameters() to freeze gradients.
+        """
+        self.modules = modules
+        self.param_states = [p.requires_grad for p in get_parameters(self.modules)]
+
+    def __enter__(self):
+
+        for param in get_parameters(self.modules):
+            param.requires_grad = False
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        for i, param in enumerate(get_parameters(self.modules)):
+            param.requires_grad = self.param_states[i]