Merge pull request #10 from praveen-palanisamy/master

Added opensim-rl environment and a sample configuration and options for a continuous DQN agent to learn in that environment

README.md

@@ -108,6 +108,7 @@ NOTE: we follow the exact code structure as [pytorch-dnc](https://github.com/jin
 - [Visdom](https://github.com/facebookresearch/visdom)
 - [OpenAI Gym >=v0.9.0 (for lower versoins, just need to change into the available games, e.g. change PongDeterministic-v4 to PongDeterministic-v3)](https://github.com/openai/gym)
 - [mujoco-py (Optional: for training continuous version of a3c)](https://github.com/openai/mujoco-py)
+- [opensim-rl (Optional: for training in Opensim-rl environment) (https://github.com/stanfordnmbl/osim-rl)
 *******
 
 

core/agents/dqn.py

@@ -150,7 +150,7 @@ def _epsilon_greedy(self, q_values_ts):
             self.eps = self.eps_eval
         # choose action
         if np.random.uniform() < self.eps:  # then we choose a random action
-            action = random.randrange(self.action_dim)
+            action = np.random.rand(self.action_dim).tolist()
         else:                               # then we choose the greedy action
             if self.use_cuda:
                 action = np.argmax(q_values_ts.cpu().numpy())
@@ -164,7 +164,7 @@ def _forward(self, observation):
         state_ts = torch.from_numpy(np.array(state)).unsqueeze(0).type(self.dtype)
         q_values_ts = self.model(Variable(state_ts, volatile=True)).data # NOTE: only doing inference here, so volatile=True
         if self.training and self.step < self.learn_start:  # then we don't do any learning, just accumulate experiences into replay memory
-            action = random.randrange(self.action_dim)      # thus we only randomly sample actions here, since the model hasn't been updated at all till now
+            action = np.random.rand(self.action_dim).tolist()      # thus we only randomly sample actions here, since the model hasn't been updated at all till now
         else:
             action = self._epsilon_greedy(q_values_ts)
 

opensim.py

@@ -0,0 +1,75 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import numpy as np
+from copy import deepcopy
+from gym.spaces.box import Box
+import inspect
+
+from utils.helpers import Experience            # NOTE: here state0 is always "None"
+from utils.helpers import preprocessAtari, rgb2gray, rgb2y, scale
+from core.env import Env
+
+class OpenSim(Env):  # low dimensional observations
+    """ Class to setup the OpenSim-RL environment (https://github.com/praveen-palanisamy/pytorch-rl.git) Where the agent has to learn to run! Continuous (18 dim) action space."""
+    def __init__(self, args, env_ind=0):
+        super(OpenSim, self).__init__(args, env_ind)
+
+        assert self.env_type == "opensim"
+        try: from osim.env import RunEnv 
+        except ImportError as e: self.logger.warning("WARNING: opensim not found")
+
+        self.env = RunEnv(visualize= True)
+        #self.env.seed(self.seed)    # NOTE: so each env would be different
+
+        # action space setup
+        self.actions     = range(self.action_dim)
+        self.logger.warning("Action Space: %s", self.env.action_space)
+
+        # state space setup
+        self.logger.warning("State  Space: %s", self.state_shape)
+
+        # continuous space
+        #if args.agent_type == "a3c":
+        self.enable_continuous = True #args.enable_continuous
+
+    def _preprocessState(self, state):    # NOTE: here no preprecessing is needed
+        return state
+
+    @property
+    def action_dim(self):
+        return self.env.action_space.shape[0]
+
+    @property
+    def state_shape(self):
+        return self.env.observation_space.shape[0]
+
+    def render(self):
+        #if self.mode == 2:
+        #    frame = self.env.render(mode='rgb_array')
+        #    frame_name = self.img_dir + "frame_%04d.jpg" % self.frame_ind
+        #    self.imsave(frame_name, frame)
+        #    self.logger.warning("Saved  Frame    @ Step: " + str(self.frame_ind) + " To: " + frame_name)
+        #    self.frame_ind += 1
+        #    return frame
+        #else:
+        #    return self.env.render()
+        return
+
+
+    def visual(self):
+        pass
+
+    def sample_random_action(self):
+        return self.env.action_space.sample()
+
+    def reset(self):
+        self._reset_experience()
+        self.exp_state1 = self.env.reset()
+        return self._get_experience()
+
+    def step(self, action):
+        self.exp_action = action
+        if self.enable_continuous:
+            self.exp_state1, self.exp_reward, self.exp_terminal1, _ = self.env.step(self.exp_action)
+        return self._get_experience()

utils/factory.py

@@ -6,10 +6,12 @@
 from core.envs.atari_ram import AtariRamEnv
 from core.envs.atari import AtariEnv
 from core.envs.lab import LabEnv
+from core.envs.opensim import OpenSim
 EnvDict = {"gym":       GymEnv,                 # classic control games from openai w/ low-level   input
            "atari-ram": AtariRamEnv,            # atari integrations from openai, with low-level   input
            "atari":     AtariEnv,               # atari integrations from openai, with pixel-level input
-           "lab":       LabEnv}
+           "lab":       LabEnv,
+	   "opensim": OpenSim}
 
 from core.models.empty import EmptyModel
 from core.models.dqn_mlp import DQNMlpModel
@@ -20,6 +22,7 @@
 from core.models.acer_cnn_dis import ACERCnnDisModel
 ModelDict = {"empty":        EmptyModel,        # contains nothing, only should be used w/ EmptyAgent
              "dqn-mlp":      DQNMlpModel,       # for dqn low-level    input
+             "dqn-mlp-con":  DQNMlpModel,       # for dqn low-level    input
              "dqn-cnn":      DQNCnnModel,       # for dqn pixel-level  input
              "a3c-mlp-con":  A3CMlpConModel,    # for a3c low-level    input (NOTE: continuous must end in "-con")
              "a3c-cnn-dis":  A3CCnnDisModel,    # for a3c pixel-level  input

utils/options.py

@@ -22,19 +22,20 @@
 [ "dqn",      "atari",     "BreakoutDeterministic-v4", "dqn-cnn",      "sequential"],  # 4
 [ "a3c",      "atari",     "PongDeterministic-v4",     "a3c-cnn-dis",  "none"      ],  # 5
 [ "a3c",      "gym",       "InvertedPendulum-v1",      "a3c-mlp-con",  "none"      ],  # 6
-[ "acer",     "gym",       "MountainCar-v0",           "acer-mlp-dis", "episodic"  ]   # 7  # NOTE: acer under testing
+[ "acer",     "gym",       "MountainCar-v0",           "acer-mlp-dis", "episodic"  ],   # 7  # NOTE: acer under testing
+[ "dqn",      "opensim",     "opensim", "dqn-mlp-con",      "sequential"]  # 8
 ]
 
 class Params(object):   # NOTE: shared across all modules
     def __init__(self):
         self.verbose     = 0            # 0(warning) | 1(info) | 2(debug)
 
         # training signature
-        self.machine     = "aisdaim"    # "machine_id"
-        self.timestamp   = "17082400"   # "yymmdd##"
+        self.machine     = "hpc011"    # "machine_id"
+        self.timestamp   = "1"   # "yymmdd##"
         # training configuration
         self.mode        = 1            # 1(train) | 2(test model_file)
-        self.config      = 7
+        self.config      = 8
 
         self.seed        = 123
         self.render      = False        # whether render the window from the original envs or not
@@ -53,7 +54,7 @@ def __init__(self):
                 self.hidden_dim     = 16
             else:
                 self.hist_len       = 4
-                self.hidden_dim     = 256
+                self.hidden_dim     = 512#256
 
             self.use_cuda           = torch.cuda.is_available()
             self.dtype              = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.FloatTensor
@@ -139,8 +140,12 @@ def __init__(self):
             self.wid_state = 80
             self.preprocess_mode = 3  # 0(nothing) | 1(rgb2gray) | 2(rgb2y) | 3(crop&resize depth)
             self.img_encoding_type = "passthrough"
+
+        elif self.env_type == "opensim":
+            pass
+
         else:
-            assert False, "env_type must be: gym | atari-ram | atari | lab"
+            assert False, "env_type must be: gym | atari-ram | atari | lab | opensim"
 
 class ModelParams(Params):  # settings for network architecture
     def __init__(self):
@@ -228,6 +233,31 @@ def __init__(self):
             self.action_repetition   = 4
             self.memory_interval     = 1
             self.train_interval      = 4
+        elif self.agent_type == "dqn" and self.env_type == "opensim":
+            self.steps               = 50000000 # max #iterations
+            self.early_stop          = None     # max #steps per episode
+            self.gamma               = 0.99
+            self.clip_grad           = 40.#np.inf
+            self.lr                  = 0.00025
+            self.lr_decay            = False
+            self.weight_decay        = 0.
+            self.eval_freq           = 250000#12500    # NOTE: here means every this many steps
+            self.eval_steps          = 125000#2500
+            self.prog_freq           = 10000#self.eval_freq
+            self.test_nepisodes      = 1
+
+            self.learn_start         = 50000    # start update params after this many steps
+            self.batch_size          = 32
+            self.valid_size          = 500
+            self.eps_start           = 1
+            self.eps_end             = 0.1
+            self.eps_eval            = 0.#0.05
+            self.eps_decay           = 1000000
+            self.target_model_update = 10000
+            self.action_repetition   = 4
+            self.memory_interval     = 1
+            self.train_interval      = 4
+
         elif self.agent_type == "a3c":
             self.steps               = 20000000 # max #iterations
             self.early_stop          = None     # max #steps per episode