NeurIPS2018-AI-for-Prosthetics-Challenge training code (#40)

* NeurIPS2018-AI-for-Prosthetics-Challenge training code

* remove model_zoo, provide download link

* remove model_zoo, provide download link

* add restore_from_one_head api, refine README, fix logger bug

* fix test bug

* fix rpm bug, refine ddpg train script

* fix rpm bug, refine Readme

README.md

@@ -61,13 +61,13 @@ three steps to build an agent
 
 model = AtariModel(img_shape=(32, 32), action_dim=4)
 algorithm = DQN(model)
-agent = AtariAgent(aglrotihm)
+agent = AtariAgent(algorithm)
 ```
 
 # Install:
 ### Dependencies
 - Python 2.7 or 3.5+. 
-- PaddlePaddle >=1.0 (We try to make our repository always compatible with newest version PaddlePaddle)  
+- PaddlePaddle >=1.2.1 (We try to make our repository always compatible with newest version PaddlePaddle)  
 
 
 ```

examples/DDPG/train.py

@@ -19,12 +19,10 @@ import time
 from mujoco_agent import MujocoAgent
 from mujoco_model import MujocoModel
 from parl.algorithms import DDPG
-from parl.utils import logger, action_mapping
-from replay_memory import ReplayMemory
+from parl.utils import logger, action_mapping, ReplayMemory
 
 MAX_EPISODES = 5000
-TEST_EVERY_EPISODES = 50
-MAX_STEPS_EACH_EPISODE = 1000
+TEST_EVERY_EPISODES = 20
 ACTOR_LR = 1e-4
 CRITIC_LR = 1e-3
 GAMMA = 0.99
@@ -39,7 +37,7 @@ ENV_SEED = 1
 def run_train_episode(env, agent, rpm):
     obs = env.reset()
     total_reward = 0
-    for j in range(MAX_STEPS_EACH_EPISODE):
+    while True:
         batch_obs = np.expand_dims(obs, axis=0)
         action = agent.predict(batch_obs.astype('float32'))
         action = np.squeeze(action)
@@ -70,7 +68,7 @@ def run_train_episode(env, agent, rpm):
 def run_evaluate_episode(env, agent):
     obs = env.reset()
     total_reward = 0
-    for j in range(MAX_STEPS_EACH_EPISODE):
+    while True:
         batch_obs = np.expand_dims(obs, axis=0)
         action = agent.predict(batch_obs.astype('float32'))
         action = np.squeeze(action)
@@ -112,7 +110,8 @@ def main():
         logger.info('Episode: {} Reward: {}'.format(i, train_reward))
         if (i + 1) % TEST_EVERY_EPISODES == 0:
             evaluate_reward = run_evaluate_episode(env, agent)
-            logger.info('Evaluate Reward: {}'.format(evaluate_reward))
+            logger.info('Episode {}, Evaluate reward: {}'.format(
+                i, evaluate_reward))
 
 
 if __name__ == '__main__':

examples/DQN/train.py

@@ -64,6 +64,7 @@ def run_train_episode(env, agent, rpm):
         state = next_state
         if isOver:
             break
+    if all_cost:
         logger.info('[Train]total_reward: {}, mean_cost: {}'.format(
             total_reward, np.mean(all_cost)))
     return total_reward, step
@@ -122,7 +123,7 @@ def main():
         if total_step // args.test_every_steps == test_flag:
             pbar.write("testing")
             eval_rewards = []
-            for _ in tqdm(range(30), desc='eval agent'):
+            for _ in tqdm(range(3), desc='eval agent'):
                 eval_reward = run_evaluate_episode(test_env, agent)
                 eval_rewards.append(eval_reward)
             test_flag += 1

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/README.md

-## The Winning Solution for the NeurIPS 2018: AI for Prosthetics Challenge
+# The Winning Solution for the NeurIPS 2018: AI for Prosthetics Challenge
 
-This folder will contains the code used to train the winning models for the [NeurIPS 2018: AI for Prosthetics Challenge](https://www.crowdai.org/challenges/neurips-2018-ai-for-prosthetics-challenge) along with the resulting models. (Codes of training part is organizing, but the resulting models is available now.)
-
-### Dependencies
+This folder contains the code used to train the winning models for the [NeurIPS 2018: AI for Prosthetics Challenge](https://www.crowdai.org/challenges/neurips-2018-ai-for-prosthetics-challenge) along with the resulting models.
+## Dependencies
 - python3.6
-- [paddlepaddle>=1.2.0](https://github.com/PaddlePaddle/Paddle)
+- [paddlepaddle>=1.2.1](https://github.com/PaddlePaddle/Paddle)
 - [osim-rl](https://github.com/stanfordnmbl/osim-rl)
+- [grpcio==1.12.1](https://grpc.io/docs/quickstart/python.html)
+- tqdm
+- tensorflow (To use tensorboard)
+
+## Result
 
-### Start Testing best models
+| Avg reward of all episodes | Avg reward of complete episodes | Falldown % | Evaluate episodes |
+|----------------------------|---------------------------------|------------|-------------------|
+| 9968.5404                  | 9980.3952                       | 0.0026     | 500 CPUs * 10 episodes              |
+
+## Start test our submit models
 - How to Run
+
 ```bash
 # cd current directory
-# install best models file (saved_model.tar.gz) 
+# cd final_submit/
+# download submit models file (saved_model.tar.gz)
 tar zxvf saved_model.tar.gz
 python test.py
 ```
-> You can install models file from [Baidu Pan](https://pan.baidu.com/s/1NN1auY2eDblGzUiqR8Bfqw) or [Google Drive](https://drive.google.com/open?id=1DQHrwtXzgFbl9dE7jGOe9ZbY0G9-qfq3)
+> You can download models file from [Baidu Pan](https://pan.baidu.com/s/1NN1auY2eDblGzUiqR8Bfqw) or [Google Drive](https://drive.google.com/open?id=1DQHrwtXzgFbl9dE7jGOe9ZbY0G9-qfq3)
+
+
+## Start train
+
+### Stage I: Curriculum learning
+
+#### 1. Run Fastest
+
+```bash
+# server
+python simulator_server.py --port [PORT] --ensemble_num 1 
+
+# client (Suggest: 200+ clients)
+python simulator_client.py --port [PORT] --ip [IP] --reward_type RunFastest
+```
+
+#### 2. target speed 3.0 m/s
 
-- More arguments
 ```bash
-# Run with GPU
-python test.py --use_cuda 
+# server
+python simulator_server.py --port [PORT] --ensemble_num 1 --warm_start_batchs 1000 \
+           --restore_model_path [RunFastest model]
+
+# client (Suggest: 200+ clients)
+python simulator_client.py --port [PORT] --ip [IP] --reward_type FixedTargetSpeed --target_v 3.0 \
+           --act_penalty_lowerbound 1.5 
+```
 
-# Visulize the game
-python test.py --vis
+#### 3. target speed 2.0 m/s
 
-# Set the random seed 
-python test.py --seed 1024
+```bash
+# server
+python simulator_server.py --port [PORT] --ensemble_num 1 --warm_start_batchs 1000 \
+           --restore_model_path [FixedTargetSpeed 3.0m/s model]
 
-# Set the episode number to run
-python test.py --episode_num 2
+# client (Suggest: 200+ clients)
+python simulator_client.py --port [PORT] --ip [IP] --reward_type FixedTargetSpeed --target_v 2.0 \
+           --act_penalty_lowerbound 0.75 
 ```
 
-### Start Training
-- [ ] To be Done
+#### 4. target speed 1.25 m/s
+
+```bash
+# server
+python simulator_server.py --port [PORT] --ensemble_num 1 --warm_start_batchs 1000 \
+           --restore_model_path [FixedTargetSpeed 2.0m/s model]  
+
+# client (Suggest: 200+ clients)
+python simulator_client.py --port [PORT] --ip [IP] --reward_type FixedTargetSpeed --target_v 1.25 \
+           --act_penalty_lowerbound 0.6
+```
+
+### Stage II: Round2
+
+> You can download resulting 1.25m/s model in Stage I from [Baidu Pan](https://pan.baidu.com/s/1PVDgIe3NuLB-4qI5iSxtKA) or [Google Drive](https://drive.google.com/open?id=1jWzs3wvq7_ierIwGZXc-M92bv1X5eqs7)
+
+```bash
+# server
+python simulator_server.py --port [PORT] --ensemble_num 12 --warm_start_batchs 1000 \
+           --restore_model_path [FixedTargetSpeed 1.25m/s] --restore_from_one_head 
+
+# client (Suggest: 100+ clients)
+python simulator_client.py --port [PORT] --ip [IP] --reward_type Round2 --act_penalty_lowerbound 0.75 \
+           --act_penalty_coeff 7.0 --vel_penalty_coeff 20.0 --discrete_data --stage 3
+```
+
+> To get a higher score, you need train a seperate model for every stage (target_v change times), and fix trained model of previous stage. It's omitted here.
+
+### Test trained model
+
+```bash
+python test.py --restore_model_path [MODEL_PATH] --ensemble_num [ENSEMBLE_NUM]
+```

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/args.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+
+
+def get_server_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--port', required=True, type=int, help='server port')
+    parser.add_argument(
+        '--logdir', type=str, help='directory to save model/tensorboard data')
+    parser.add_argument(
+        '--restore_model_path',
+        type=str,
+        help='restore model path for warm start')
+    parser.add_argument(
+        '--restore_from_one_head',
+        action="store_true",
+        help=
+        'If set, will restore model from one head model. If ensemble_num > 1, will assign parameters of model0 to other models.'
+    )
+    parser.add_argument(
+        '--restore_rpm_path', type=str, help='restore rpm path for warm start')
+    parser.add_argument(
+        '--ensemble_num',
+        type=int,
+        required=True,
+        help='model number to ensemble')
+    parser.add_argument(
+        '--warm_start_batchs',
+        type=int,
+        default=100,
+        help='collect how many batch data to warm start')
+    args = parser.parse_args()
+    return args
+
+
+def get_client_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        '--stage',
+        default=0,
+        type=int,
+        help='''
+          stage number, which decides change times of target velocity. 
+          Eg: stage=0 will keep target_v 1.25m/s;
+              stage=3 will change target velocity 3 times, just like Round2 env.'''
+    )
+    parser.add_argument('--ident', type=int, required=False, help='worker id')
+    parser.add_argument('--ip', type=str, required=True, help='server ip')
+    parser.add_argument('--port', type=int, required=True, help='server port')
+    parser.add_argument(
+        '--target_v', type=float, help='target velocity for training')
+    parser.add_argument(
+        '--act_penalty_lowerbound',
+        type=float,
+        help='lower bound of action l2 norm penalty')
+    parser.add_argument(
+        '--act_penalty_coeff',
+        type=float,
+        default=5.0,
+        help='coefficient of action l2 norm penalty')
+    parser.add_argument(
+        '--vel_penalty_coeff',
+        type=float,
+        default=1.0,
+        help='coefficient of velocity gap penalty')
+    parser.add_argument(
+        '--discrete_data',
+        action="store_true",
+        help=
+        'if set, discrete target velocity in last stage (args.stage), make target velocity more uniform.'
+    )
+    parser.add_argument(
+        '--discrete_bin',
+        type=int,
+        default=10,
+        help='discrete target velocity in last stage to how many intervals')
+    parser.add_argument(
+        '--reward_type',
+        type=str,
+        help=
+        "Choose reward type, 'RunFastest' or 'FixedTargetSpeed' or 'Round2'")
+    parser.add_argument(
+        '--debug',
+        action="store_true",
+        help='if set, will print debug information')
+    args = parser.parse_args()
+
+    assert args.reward_type in ['RunFastest', 'FixedTargetSpeed', 'Round2']
+
+    return args

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/env_wrapper.py

@@ -13,41 +13,218 @@
 # limitations under the License.
 
 import abc
-import copy
 import gym
 import math
 import numpy as np
+import random
 from collections import OrderedDict
 from osim.env import ProstheticsEnv
 from parl.utils import logger
+from tqdm import tqdm
 
 MAXTIME_LIMIT = 1000
 ProstheticsEnv.time_limit = MAXTIME_LIMIT
 FRAME_SKIP = None
-FALL_PENALTY = 0
 
 
-class RemoteEnv(gym.Wrapper):
-    def __init__(self, env):
-        env.metadata = {}
-        env.action_space = None
-        env.observation_space = None
-        env.reward_range = None
+class CustomR2Env(gym.Wrapper):
+    """Customized target trajectory here, it support 3 ways currently
+        1.fixed_speed, e.g. reset(.., fixed_speed=1.25)
+        2.stage , e.g. reset(.., stage=1)
+        3.boundary, e.g. reset(.., boundary=True)
+    """
+
+    def __init__(self,
+                 env,
+                 time_limit=MAXTIME_LIMIT,
+                 discrete_data=False,
+                 discrete_bin=10):
+        logger.info("[CustomR2Env]type:{}, time_limit:{}".format(
+            type(env), time_limit))
+        assert isinstance(env, ProstheticsEnv), type(env)
         gym.Wrapper.__init__(self, env)
-        self.remote_env = env
-        self.first_time = True
-
-    def step(self, act):
-        return self.remote_env.env_step(act.tolist())
-
-    def reset(self):
-        if self.first_time:
-            self.first_time = False
-            return self.remote_env.env_create()
-        obs = self.remote_env.env_reset()
-        if not obs:
-            return None
-        return obs
+
+        self.env.time_limit = time_limit
+        self.env.spec.timestep_limit = time_limit
+        self.time_limit = time_limit
+
+        # boundary flag
+        self._generate_boundary_target_flag = True
+
+        self.discrete_data = discrete_data
+        self.discrete_bin = discrete_bin
+
+    def rect(self, row):
+        r = row[0]
+        theta = row[1]
+        x = r * math.cos(theta)
+        y = 0
+        z = r * math.sin(theta)
+        return np.array([x, y, z])
+
+    def _generate_boundary_table(self, ):
+        possible_combine = [(math.pi / 8, 0.5), (math.pi / 8, -0.5),
+                            (-math.pi / 8, 0.5), (-math.pi / 8, -0.5)]
+        self._boundary_table = []
+        for a in possible_combine:
+            for b in possible_combine:
+                for c in possible_combine:
+                    self._boundary_table.append([a, b, c])
+        assert len(self._boundary_table) == 64
+
+    def generate_boundary_target(self, poisson_lambda=300):
+        if self._generate_boundary_target_flag == True:
+            self._generate_boundary_target_flag = False
+            self._generate_boundary_table()
+            self._boundary_index = 0
+        nsteps = self.time_limit + 1
+        velocity = np.zeros(nsteps)
+        heading = np.zeros(nsteps)
+
+        velocity[0] = 1.25
+        heading[0] = 0
+        trajectory = self._boundary_table[self._boundary_index]
+
+        change = np.cumsum(np.random.poisson(poisson_lambda, 10))
+        target_change_times = 0
+        for i in range(1, nsteps):
+            velocity[i] = velocity[i - 1]
+            heading[i] = heading[i - 1]
+            if i in change:
+                velocity[i] += trajectory[target_change_times][1]
+                heading[i] += trajectory[target_change_times][0]
+                # trajectory has length 3, the target_change_times should not be large than 2
+                target_change_times = min(2, target_change_times + 1)
+
+        self._boundary_index = (self._boundary_index + 1) % 64
+
+    def _generate_target_vel(self, stage, change_num):
+        target_vels = None
+        if stage == 0:
+            target_vels = [1.25]
+        elif stage == 1:
+            assert change_num >= 1
+            interval = 1.0 / self.discrete_bin
+            discrete_id = np.random.randint(self.discrete_bin)
+
+            min_vel = 0.75 + discrete_id * interval
+            max_vel = 0.75 + (discrete_id + 1) * interval
+
+            target_vels = [1.25]
+            for i in range(change_num):
+                if i == 0:
+                    target_vels.append(random.uniform(min_vel, max_vel))
+                else:
+                    target_vels.append(target_vels[-1] +
+                                       random.uniform(-0.5, 0.5))
+        elif stage == 2:
+            assert change_num >= 2
+            interval = 2.0 / self.discrete_bin
+            discrete_id = np.random.randint(self.discrete_bin)
+            min_vel = 0.25 + discrete_id * interval
+            max_vel = 0.25 + (discrete_id + 1) * interval
+            while True:
+                target_vels = [1.25]
+                for i in range(change_num):
+                    target_vels.append(target_vels[-1] +
+                                       random.uniform(-0.5, 0.5))
+                if target_vels[2] >= min_vel and target_vels[2] <= max_vel:
+                    break
+        elif stage == 3:
+            assert change_num >= 3
+            interval = 3.0 / self.discrete_bin
+            discrete_id = np.random.randint(self.discrete_bin)
+            min_vel = -0.25 + discrete_id * interval
+            max_vel = -0.25 + (discrete_id + 1) * interval
+            while True:
+                target_vels = [1.25]
+                for i in range(change_num):
+                    target_vels.append(target_vels[-1] +
+                                       random.uniform(-0.5, 0.5))
+                if target_vels[3] >= min_vel and target_vels[3] <= max_vel:
+                    break
+        else:
+            raise NotImplemented
+        logger.info('[CustomR2Env] stage: {}, target_vels: {}'.format(
+            stage, target_vels))
+        return target_vels
+
+    def generate_stage_targets(self, poisson_lambda=300, stage=None):
+        nsteps = self.time_limit + 1
+        velocity = np.zeros(nsteps)
+        heading = np.zeros(nsteps)
+
+        velocity[0] = 1.25
+        heading[0] = 0
+
+        change = np.cumsum(np.random.poisson(poisson_lambda, 10))
+        if stage == 0:
+            change = []
+        elif stage == 1:
+            change = change[:1]
+        elif stage == 2:
+            change = change[:2]
+        elif stage == 3:
+            if change[3] <= 1000:
+                change = change[:4]
+            else:
+                change = change[:3]
+        else:
+            raise NotImplemented
+
+        if self.discrete_data:
+            target_vels = self._generate_target_vel(
+                stage=stage, change_num=len(change))
+
+        change_cnt = 0
+        for i in range(1, nsteps):
+            velocity[i] = velocity[i - 1]
+            heading[i] = heading[i - 1]
+
+            if i in change:
+                change_cnt += 1
+                if self.discrete_data:
+                    velocity[i] = target_vels[change_cnt]
+                else:
+                    velocity[i] += random.choice([-1, 1]) * random.uniform(
+                        -0.5, 0.5)
+                heading[i] += random.choice([-1, 1]) * random.uniform(
+                    -math.pi / 8, math.pi / 8)
+
+        trajectory_polar = np.vstack((velocity, heading)).transpose()
+        targets = np.apply_along_axis(self.rect, 1, trajectory_polar)
+        return targets
+
+    def reset(self, **kwargs):
+        fixed_speed = None
+        if 'fixed_speed' in kwargs:
+            fixed_speed = kwargs.pop('fixed_speed', None)
+        stage = None
+        if 'stage' in kwargs:
+            stage = kwargs.pop('stage', None)
+        boundary = None
+        if 'boundary' in kwargs:
+            boundary = kwargs.pop('boundary', None)
+        _ = self.env.reset(**kwargs)
+        if fixed_speed is not None:
+            targets = np.zeros([self.time_limit + 1, 3], dtype=np.float32)
+            targets[:, 0] = fixed_speed
+            self.env.targets = targets
+        elif stage is not None:
+            self.env.targets = self.generate_stage_targets(stage=stage)
+        elif boundary is not None:
+            self.generate_boundary_target()
+        else:
+            # generate new target
+            self.env.generate_new_targets(
+                poisson_lambda=int(self.time_limit * (300 / 1000)))
+        if 'project' in kwargs:
+            if kwargs.get('project') == True:
+                return self.env.get_observation()
+        return self.env.get_state_desc()
+
+    def step(self, action, **kwargs):
+        return self.env.step(action, **kwargs)
 
 
 def calc_vel_diff(state_desc):
@@ -85,6 +262,7 @@ class ActionScale(gym.Wrapper):
 
 class FrameSkip(gym.Wrapper):
     def __init__(self, env, k):
+        logger.info("[FrameSkip]type:{}".format(type(env)))
         gym.Wrapper.__init__(self, env)
         self.frame_skip = k
         global FRAME_SKIP
@@ -103,13 +281,13 @@ class FrameSkip(gym.Wrapper):
                 if 'reward' in key:
                     # to assure that we don't igonre other reward
                     # if new reward was added, consider its logic here
-                    assert (key == 'shaping_reward') or (key == 'r2_reward')
+                    assert (key == 'shaping_reward') or (
+                        key == 'r2_reward') or (key == 'x_offset_reward')
                     merge_info[key] = merge_info.get(key, 0.0) + info[key]
                 else:
                     merge_info[key] = info[key]
 
             if info['target_changed']:
-                #merge_info['shaping_reward'] += info['shaping_reward'] * (self.frame_skip - k - 1)
                 logger.warn("[FrameSkip] early break since target was changed")
                 break
 
@@ -128,6 +306,8 @@ class RewardShaping(gym.Wrapper):
 
     def __init__(self, env):
         logger.info("[RewardShaping]type:{}".format(type(env)))
+        assert (isinstance(env, ProstheticsEnv)
+                or isinstance(env, CustomR2Env)), type(env)
 
         self.step_count = 0
         self.pre_state_desc = None
@@ -150,25 +330,7 @@ class RewardShaping(gym.Wrapper):
         self.step_count += 1
         obs, r, done, info = self.env.step(action, **kwargs)
         info = self.reward_shaping(obs, r, done, action)
-        if info['target_vel'] > 2.75:
-            rate = math.sqrt((2.75**2) / (info['target_vel']**2))
-            logger.warn('Changing targets, origin targets: {}'.format(
-                obs['target_vel']))
-            obs['target_vel'][0] = obs['target_vel'][0] * rate
-            obs['target_vel'][2] = obs['target_vel'][2] * rate
-            logger.warn('Changing targets, new targets: {}'.format(
-                obs['target_vel']))
-            info['target_vel'] = 2.75
-        if info['target_vel'] < -0.25:
-            rate = math.sqrt(((-0.25)**2) / (info['target_vel']**2))
-            logger.warn('Changing targets, origin targets: {}'.format(
-                obs['target_vel']))
-            obs['target_vel'][0] = obs['target_vel'][0] * rate
-            obs['target_vel'][2] = obs['target_vel'][2] * rate
-            logger.warn('Changing targets, new targets: {}'.format(
-                obs['target_vel']))
-            info['target_vel'] = -0.25
-
+        #logger.info('Step {}: target_vel: {}'.format(self.step_count, obs['target_vel']))
         delta = 0
         if self.last_target_vel is not None:
             delta = np.absolute(
@@ -188,9 +350,7 @@ class RewardShaping(gym.Wrapper):
         timeout = False
         if self.step_count >= MAXTIME_LIMIT:
             timeout = True
-        if done and not timeout:
-            # penalty for falling down
-            info['shaping_reward'] += FALL_PENALTY
+
         info['timeout'] = timeout
         self.pre_state_desc = obs
         return obs, r, done, info
@@ -205,31 +365,182 @@ class RewardShaping(gym.Wrapper):
         return obs
 
 
-class ForwardReward(RewardShaping):
-    """ A reward shaping wraper"""
+class TestReward(RewardShaping):
+    """ Reward shaping wrapper for test"""
+
+    def __init__(self, env):
+        RewardShaping.__init__(self, env)
+
+    def reward_shaping(self, state_desc, r2_reward, done, action):
+        return {'shaping_reward': 0}
+
+
+class RunFastestReward(RewardShaping):
+    """ Reward shaping wrapper for fixed target speed"""
 
     def __init__(self, env):
         RewardShaping.__init__(self, env)
 
     def reward_shaping(self, state_desc, r2_reward, done, action):
-        target_vel = math.sqrt(state_desc["target_vel"][0]**2 +
-                               state_desc["target_vel"][2]**2)
-        if state_desc["target_vel"][0] < 0:
-            target_vel = -target_vel
+        if self.pre_state_desc is None:
+            x_offset = 0
+        else:
+            x_offset = state_desc["body_pos"]["pelvis"][
+                0] - self.pre_state_desc["body_pos"]["pelvis"][0]
+
+        ret_r = 0
+        if self.pre_state_desc is not None:
+            l_foot_reward = state_desc["body_pos"]["tibia_l"][
+                0] - self.pre_state_desc["body_pos"]["tibia_l"][0]
+            r_foot_reward = state_desc["body_pos"]["pros_tibia_r"][
+                0] - self.pre_state_desc["body_pos"]["pros_tibia_r"][0]
+            ret_r = max(l_foot_reward, r_foot_reward)
+
+        # penalty
+        headx = state_desc['body_pos']['head'][0]
+        px = state_desc['body_pos']['pelvis'][0]
+        headz = state_desc['body_pos']['head'][2]
+        pz = state_desc['body_pos']['pelvis'][2]
+        kneer = state_desc['joint_pos']['knee_r'][-1]
+        kneel = state_desc['joint_pos']['knee_l'][-1]
+        lean_x = min(0.3, max(0, px - headx - 0.15)) * 0.05
+        lean_z = min(0.3, max(0, pz - headz - 0.15)) * 0.05
+        joint = sum([max(0, k - 0.1) for k in [kneer, kneel]]) * 0.03
+        penalty = lean_x + lean_z + joint
+
+        ret_r -= penalty * 0.15
+
+        cur_vel_x = state_desc['body_vel']['pelvis'][0]
+        cur_vel_z = state_desc['body_vel']['pelvis'][2]
+        scalar_vel = math.sqrt(cur_vel_z**2 + cur_vel_x**2)
+
+        info = {
+            'shaping_reward': ret_r,
+            'r2_reward': r2_reward,
+            'x_offset_reward': x_offset,
+            'scalar_vel': scalar_vel,
+            'mean_action_l2_penalty': 0,
+        }
+        return info
+
+
+class FixedTargetSpeedReward(RewardShaping):
+    """ Reward shaping wrapper for fixed target speed"""
+
+    def __init__(self, env, target_v, act_penalty_lowerbound,
+                 act_penalty_coeff, vel_penalty_coeff):
+        RewardShaping.__init__(self, env)
+
+        assert target_v is not None
+        assert act_penalty_lowerbound is not None
+        assert act_penalty_coeff is not None
+        assert vel_penalty_coeff is not None
+
+        self.target_v = target_v
+        self.act_penalty_lowerbound = act_penalty_lowerbound
+        self.act_penalty_coeff = act_penalty_coeff
+        self.vel_penalty_coeff = vel_penalty_coeff
+
+    def reward_shaping(self, state_desc, r2_reward, done, action):
+        if self.pre_state_desc is None:
+            x_offset = 0
+        else:
+            x_offset = state_desc["body_pos"]["pelvis"][
+                0] - self.pre_state_desc["body_pos"]["pelvis"][0]
+
+        # Reward for not falling
+        ret_r = 36
+
+        vel_penalty = ((state_desc["body_vel"]["pelvis"][0] - self.target_v)**2
+                       + (state_desc["body_vel"]["pelvis"][2] - 0)**2)
+
+        origin_action_l2_penalty = np.linalg.norm(action)
+        action_l2_penalty = max(self.act_penalty_lowerbound,
+                                origin_action_l2_penalty)
+
+        ret_r = ret_r - vel_penalty * self.vel_penalty_coeff - action_l2_penalty * self.act_penalty_coeff
+
+        cur_vel_x = state_desc['body_vel']['pelvis'][0]
+        cur_vel_z = state_desc['body_vel']['pelvis'][2]
+        scalar_vel = math.sqrt(cur_vel_z**2 + cur_vel_x**2)
+
+        info = {
+            'shaping_reward': ret_r,
+            'r2_reward': r2_reward,
+            'x_offset_reward': x_offset,
+            'scalar_vel': scalar_vel,
+            'mean_action_l2_penalty': origin_action_l2_penalty,
+        }
+        return info
+
+
+class Round2Reward(RewardShaping):
+    """ Reward shaping wrapper for fixed target speed"""
+
+    def __init__(self, env, act_penalty_lowerbound, act_penalty_coeff,
+                 vel_penalty_coeff):
+        RewardShaping.__init__(self, env)
+
+        assert act_penalty_lowerbound is not None
+        assert act_penalty_coeff is not None
+        assert vel_penalty_coeff is not None
+
+        self.act_penalty_lowerbound = act_penalty_lowerbound
+        self.act_penalty_coeff = act_penalty_coeff
+        self.vel_penalty_coeff = vel_penalty_coeff
+
+    def reward_shaping(self, state_desc, r2_reward, done, action):
+        if self.pre_state_desc is None:
+            x_offset = 0
+        else:
+            x_offset = state_desc["body_pos"]["pelvis"][
+                0] - self.pre_state_desc["body_pos"]["pelvis"][0]
+
+        # Reward for not falling
+        ret_r = 10
+
+        # Small penalty for too much activation (cost of transport)
+        muscle_activations = []
+        for muscle in sorted(state_desc["muscles"].keys()):
+            muscle_activations += [state_desc["muscles"][muscle]["activation"]]
+        muscle_penalty = np.sum(np.array(muscle_activations)**2) * 0.001
+
+        vel_penalty = (
+            (state_desc["target_vel"][0] - state_desc["body_vel"]["pelvis"][0])
+            **2 + (state_desc["target_vel"][2] -
+                   state_desc["body_vel"]["pelvis"][2])**2)
+
+        origin_action_l2_penalty = np.linalg.norm(action)
+        action_l2_penalty = max(self.act_penalty_lowerbound,
+                                origin_action_l2_penalty)
+
+        if self.step_count < 60 or (
+                self.step_count - self.last_target_change_step < 60):
+            ret_r = ret_r - vel_penalty * self.vel_penalty_coeff
+        else:
+            ret_r = ret_r - vel_penalty * self.vel_penalty_coeff - action_l2_penalty * self.act_penalty_coeff
+
+        ret_r -= muscle_penalty
+
+        cur_vel_x = state_desc['body_vel']['pelvis'][0]
+        cur_vel_z = state_desc['body_vel']['pelvis'][2]
+        scalar_vel = math.sqrt(cur_vel_z**2 + cur_vel_x**2)
 
         info = {
-            'shaping_reward': r2_reward,
-            'target_vel': target_vel,
+            'shaping_reward': ret_r,
             'r2_reward': r2_reward,
+            'x_offset_reward': x_offset,
+            'scalar_vel': scalar_vel,
+            'mean_action_l2_penalty': origin_action_l2_penalty,
         }
         return info
 
 
 class ObsTranformerBase(gym.Wrapper):
     def __init__(self, env):
+        logger.info("[ObsTranformerBase]type:{}".format(type(env)))
         gym.Wrapper.__init__(self, env)
         self.step_fea = MAXTIME_LIMIT
-        self.raw_obs = None
         global FRAME_SKIP
         self.frame_skip = int(FRAME_SKIP)
 
@@ -268,19 +579,13 @@ class ObsTranformerBase(gym.Wrapper):
 
         self.step_fea -= FRAME_SKIP
 
-        self.raw_obs = copy.deepcopy(obs)
         obs = self.get_observation(obs)
-        self.raw_obs['step_count'] = MAXTIME_LIMIT - self.step_fea
         return obs, r, done, info
 
     def reset(self, **kwargs):
         obs = self.env.reset(**kwargs)
-        if obs is None:
-            return None
         self.step_fea = MAXTIME_LIMIT
-        self.raw_obs = copy.deepcopy(obs)
         obs = self.get_observation(obs)
-        self.raw_obs['step_count'] = MAXTIME_LIMIT - self.step_fea
         return obs
 
 
@@ -421,7 +726,6 @@ class PelvisBasedObs(ObsTranformerBase):
         res = np.append(res, feet_dis)
         remaining_time = (self.step_fea -
                           (MAXTIME_LIMIT / 2.0)) / (MAXTIME_LIMIT / 2.0) * -1.0
-        #logger.info('remaining_time fea: {}'.format(remaining_time))
         res = np.append(res, remaining_time)
 
         # target driven
@@ -450,9 +754,10 @@ if __name__ == '__main__':
 
     env = ProstheticsEnv(visualize=False)
     env.change_model(model='3D', difficulty=1, prosthetic=True)
-    env = ForwardReward(env)
+    env = CustomR2Env(env)
+    env = RunFastestReward(env)
     env = FrameSkip(env, 4)
     env = ActionScale(env)
     env = PelvisBasedObs(env)
     for i in range(64):
-        observation = env.reset(project=False)
+        observation = env.reset(project=False, stage=0)

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/final_submit/env_wrapper.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import abc
+import copy
+import gym
+import math
+import numpy as np
+from collections import OrderedDict
+from osim.env import ProstheticsEnv
+from parl.utils import logger
+
+MAXTIME_LIMIT = 1000
+ProstheticsEnv.time_limit = MAXTIME_LIMIT
+FRAME_SKIP = None
+FALL_PENALTY = 0
+
+
+class RemoteEnv(gym.Wrapper):
+    def __init__(self, env):
+        env.metadata = {}
+        env.action_space = None
+        env.observation_space = None
+        env.reward_range = None
+        gym.Wrapper.__init__(self, env)
+        self.remote_env = env
+        self.first_time = True
+
+    def step(self, act):
+        return self.remote_env.env_step(act.tolist())
+
+    def reset(self):
+        if self.first_time:
+            self.first_time = False
+            return self.remote_env.env_create()
+        obs = self.remote_env.env_reset()
+        if not obs:
+            return None
+        return obs
+
+
+def calc_vel_diff(state_desc):
+    cur_vel_x = state_desc['body_vel']['pelvis'][0]
+    cur_vel_z = state_desc['body_vel']['pelvis'][2]
+    target_vel_x = state_desc['target_vel'][0]
+    target_vel_z = state_desc['target_vel'][2]
+    diff_vel_x = cur_vel_x - target_vel_x
+    diff_vel_z = cur_vel_z - target_vel_z
+
+    cur_vel = (cur_vel_x**2 + cur_vel_z**2)**0.5
+    target_vel = (target_vel_x**2 + target_vel_z**2)**0.5
+    diff_vel = cur_vel - target_vel
+
+    target_theta = math.atan(-1.0 * target_vel_z / target_vel_x)
+    # alone y axis
+    cur_theta = state_desc['body_pos_rot']['pelvis'][1]
+    diff_theta = cur_theta - target_theta
+
+    return cur_vel_x, cur_vel_z, diff_vel_x, diff_vel_z, diff_vel, diff_theta
+
+
+class ActionScale(gym.Wrapper):
+    def __init__(self, env):
+        gym.Wrapper.__init__(self, env)
+
+    def step(self, action, **kwargs):
+        action = (np.copy(action) + 1.0) * 0.5
+        action = np.clip(action, 0.0, 1.0)
+        return self.env.step(action, **kwargs)
+
+    def reset(self, **kwargs):
+        return self.env.reset(**kwargs)
+
+
+class FrameSkip(gym.Wrapper):
+    def __init__(self, env, k):
+        gym.Wrapper.__init__(self, env)
+        self.frame_skip = k
+        global FRAME_SKIP
+        FRAME_SKIP = k
+        self.frame_count = 0
+
+    def step(self, action, **kwargs):
+        r = 0.0
+        merge_info = {}
+        for k in range(self.frame_skip):
+            self.frame_count += 1
+            obs, reward, done, info = self.env.step(action, **kwargs)
+            r += reward
+
+            for key in info.keys():
+                if 'reward' in key:
+                    # to assure that we don't igonre other reward
+                    # if new reward was added, consider its logic here
+                    assert (key == 'shaping_reward') or (key == 'r2_reward')
+                    merge_info[key] = merge_info.get(key, 0.0) + info[key]
+                else:
+                    merge_info[key] = info[key]
+
+            if info['target_changed']:
+                #merge_info['shaping_reward'] += info['shaping_reward'] * (self.frame_skip - k - 1)
+                logger.warn("[FrameSkip] early break since target was changed")
+                break
+
+            if done:
+                break
+        merge_info['frame_count'] = self.frame_count
+        return obs, r, done, merge_info
+
+    def reset(self, **kwargs):
+        self.frame_count = 0
+        return self.env.reset(**kwargs)
+
+
+class RewardShaping(gym.Wrapper):
+    """ A wrapper for reward shaping, note this wrapper must be the first wrapper """
+
+    def __init__(self, env):
+        logger.info("[RewardShaping]type:{}".format(type(env)))
+
+        self.step_count = 0
+        self.pre_state_desc = None
+        self.last_target_vel = None
+        self.last_target_change_step = 0
+        self.target_change_times = 0
+        gym.Wrapper.__init__(self, env)
+
+    @abc.abstractmethod
+    def reward_shaping(self, state_desc, reward, done, action):
+        """define your own reward computation function
+    Args:
+        state_desc(dict): state description for current model
+        reward(scalar): generic reward generated by env
+        done(bool): generic done flag generated by env
+    """
+        pass
+
+    def step(self, action, **kwargs):
+        self.step_count += 1
+        obs, r, done, info = self.env.step(action, **kwargs)
+        info = self.reward_shaping(obs, r, done, action)
+        if info['target_vel'] > 2.75:
+            rate = math.sqrt((2.75**2) / (info['target_vel']**2))
+            logger.warn('Changing targets, origin targets: {}'.format(
+                obs['target_vel']))
+            obs['target_vel'][0] = obs['target_vel'][0] * rate
+            obs['target_vel'][2] = obs['target_vel'][2] * rate
+            logger.warn('Changing targets, new targets: {}'.format(
+                obs['target_vel']))
+            info['target_vel'] = 2.75
+        if info['target_vel'] < -0.25:
+            rate = math.sqrt(((-0.25)**2) / (info['target_vel']**2))
+            logger.warn('Changing targets, origin targets: {}'.format(
+                obs['target_vel']))
+            obs['target_vel'][0] = obs['target_vel'][0] * rate
+            obs['target_vel'][2] = obs['target_vel'][2] * rate
+            logger.warn('Changing targets, new targets: {}'.format(
+                obs['target_vel']))
+            info['target_vel'] = -0.25
+
+        delta = 0
+        if self.last_target_vel is not None:
+            delta = np.absolute(
+                np.array(self.last_target_vel) - np.array(obs['target_vel']))
+        if (self.last_target_vel is None) or np.all(delta < 1e-5):
+            info['target_changed'] = False
+        else:
+            info['target_changed'] = True
+            logger.info("[env_wrapper] target_changed, vx:{}   vz:{}".format(
+                obs['target_vel'][0], obs['target_vel'][2]))
+            self.last_target_change_step = self.step_count
+            self.target_change_times += 1
+        info['target_change_times'] = self.target_change_times
+        self.last_target_vel = obs['target_vel']
+
+        assert 'shaping_reward' in info
+        timeout = False
+        if self.step_count >= MAXTIME_LIMIT:
+            timeout = True
+        if done and not timeout:
+            # penalty for falling down
+            info['shaping_reward'] += FALL_PENALTY
+        info['timeout'] = timeout
+        self.pre_state_desc = obs
+        return obs, r, done, info
+
+    def reset(self, **kwargs):
+        self.step_count = 0
+        self.last_target_vel = None
+        self.last_target_change_step = 0
+        self.target_change_times = 0
+        obs = self.env.reset(**kwargs)
+        self.pre_state_desc = obs
+        return obs
+
+
+class ForwardReward(RewardShaping):
+    """ A reward shaping wraper"""
+
+    def __init__(self, env):
+        RewardShaping.__init__(self, env)
+
+    def reward_shaping(self, state_desc, r2_reward, done, action):
+        target_vel = math.sqrt(state_desc["target_vel"][0]**2 +
+                               state_desc["target_vel"][2]**2)
+        if state_desc["target_vel"][0] < 0:
+            target_vel = -target_vel
+
+        info = {
+            'shaping_reward': r2_reward,
+            'target_vel': target_vel,
+            'r2_reward': r2_reward,
+        }
+        return info
+
+
+class ObsTranformerBase(gym.Wrapper):
+    def __init__(self, env):
+        gym.Wrapper.__init__(self, env)
+        self.step_fea = MAXTIME_LIMIT
+        self.raw_obs = None
+        global FRAME_SKIP
+        self.frame_skip = int(FRAME_SKIP)
+
+    def get_observation(self, state_desc):
+        obs = self._get_observation(state_desc)
+        if not isinstance(self, PelvisBasedObs):
+            cur_vel_x, cur_vel_z, diff_vel_x, diff_vel_z, diff_vel, diff_theta = calc_vel_diff(
+                state_desc)
+            obs = np.append(obs, [
+                cur_vel_x, cur_vel_z, diff_vel_x, diff_vel_z, diff_vel,
+                diff_theta
+            ])
+        else:
+            pass
+        return obs
+
+    @abc.abstractmethod
+    def _get_observation(self, state_desc):
+        pass
+
+    def feature_normalize(self, obs, mean, std, duplicate_id):
+        scaler_len = mean.shape[0]
+        assert obs.shape[0] >= scaler_len
+        obs[:scaler_len] = (obs[:scaler_len] - mean) / std
+        final_obs = []
+        for i in range(obs.shape[0]):
+            if i not in duplicate_id:
+                final_obs.append(obs[i])
+        return np.array(final_obs)
+
+    def step(self, action, **kwargs):
+        obs, r, done, info = self.env.step(action, **kwargs)
+        if info['target_changed']:
+            # reset step_fea when change target
+            self.step_fea = MAXTIME_LIMIT
+
+        self.step_fea -= FRAME_SKIP
+
+        self.raw_obs = copy.deepcopy(obs)
+        obs = self.get_observation(obs)
+        self.raw_obs['step_count'] = MAXTIME_LIMIT - self.step_fea
+        return obs, r, done, info
+
+    def reset(self, **kwargs):
+        obs = self.env.reset(**kwargs)
+        if obs is None:
+            return None
+        self.step_fea = MAXTIME_LIMIT
+        self.raw_obs = copy.deepcopy(obs)
+        obs = self.get_observation(obs)
+        self.raw_obs['step_count'] = MAXTIME_LIMIT - self.step_fea
+        return obs
+
+
+class PelvisBasedObs(ObsTranformerBase):
+    def __init__(self, env):
+        ObsTranformerBase.__init__(self, env)
+        data = np.load('./pelvisBasedObs_scaler.npz')
+        self.mean, self.std, self.duplicate_id = data['mean'], data[
+            'std'], data['duplicate_id']
+        self.duplicate_id = self.duplicate_id.astype(np.int32).tolist()
+
+    def get_core_matrix(self, yaw):
+        core_matrix = np.zeros(shape=(3, 3))
+        core_matrix[0][0] = math.cos(yaw)
+        core_matrix[0][2] = -1.0 * math.sin(yaw)
+        core_matrix[1][1] = 1
+        core_matrix[2][0] = math.sin(yaw)
+        core_matrix[2][2] = math.cos(yaw)
+        return core_matrix
+
+    def _get_observation(self, state_desc):
+        o = OrderedDict()
+        for body_part in [
+                'pelvis', 'femur_r', 'pros_tibia_r', 'pros_foot_r', 'femur_l',
+                'tibia_l', 'talus_l', 'calcn_l', 'toes_l', 'torso', 'head'
+        ]:
+            # position
+            o[body_part + '_x'] = state_desc['body_pos'][body_part][0]
+            o[body_part + '_y'] = state_desc['body_pos'][body_part][1]
+            o[body_part + '_z'] = state_desc['body_pos'][body_part][2]
+            # velocity
+            o[body_part + '_v_x'] = state_desc["body_vel"][body_part][0]
+            o[body_part + '_v_y'] = state_desc["body_vel"][body_part][1]
+            o[body_part + '_v_z'] = state_desc["body_vel"][body_part][2]
+
+            o[body_part + '_x_r'] = state_desc["body_pos_rot"][body_part][0]
+            o[body_part + '_y_r'] = state_desc["body_pos_rot"][body_part][1]
+            o[body_part + '_z_r'] = state_desc["body_pos_rot"][body_part][2]
+
+            o[body_part + '_v_x_r'] = state_desc["body_vel_rot"][body_part][0]
+            o[body_part + '_v_y_r'] = state_desc["body_vel_rot"][body_part][1]
+            o[body_part + '_v_z_r'] = state_desc["body_vel_rot"][body_part][2]
+
+        for joint in [
+                'hip_r', 'knee_r', 'ankle_r', 'hip_l', 'knee_l', 'ankle_l',
+                'back'
+        ]:
+            if 'hip' not in joint:
+                o[joint + '_joint_pos'] = state_desc['joint_pos'][joint][0]
+                o[joint + '_joint_vel'] = state_desc['joint_vel'][joint][0]
+            else:
+                for i in range(3):
+                    o[joint + '_joint_pos_' +
+                      str(i)] = state_desc['joint_pos'][joint][i]
+                    o[joint + '_joint_vel_' +
+                      str(i)] = state_desc['joint_vel'][joint][i]
+
+        # In NIPS2017, only use activation
+        for muscle in sorted(state_desc["muscles"].keys()):
+            activation = state_desc["muscles"][muscle]["activation"]
+            if isinstance(activation, float):
+                activation = [activation]
+            for i, val in enumerate(activation):
+                o[muscle + '_activation_' + str(i)] = activation[i]
+
+            fiber_length = state_desc["muscles"][muscle]["fiber_length"]
+            if isinstance(fiber_length, float):
+                fiber_length = [fiber_length]
+            for i, val in enumerate(fiber_length):
+                o[muscle + '_fiber_length_' + str(i)] = fiber_length[i]
+
+            fiber_velocity = state_desc["muscles"][muscle]["fiber_velocity"]
+            if isinstance(fiber_velocity, float):
+                fiber_velocity = [fiber_velocity]
+            for i, val in enumerate(fiber_velocity):
+                o[muscle + '_fiber_velocity_' + str(i)] = fiber_velocity[i]
+
+        # z axis of mass have some problem now, delete it later
+        o['mass_x'] = state_desc["misc"]["mass_center_pos"][0]
+        o['mass_y'] = state_desc["misc"]["mass_center_pos"][1]
+        o['mass_z'] = state_desc["misc"]["mass_center_pos"][2]
+
+        o['mass_v_x'] = state_desc["misc"]["mass_center_vel"][0]
+        o['mass_v_y'] = state_desc["misc"]["mass_center_vel"][1]
+        o['mass_v_z'] = state_desc["misc"]["mass_center_vel"][2]
+        for key in ['talus_l_y', 'toes_l_y']:
+            o['touch_indicator_' + key] = np.clip(0.05 - o[key] * 10 + 0.5, 0.,
+                                                  1.)
+            o['touch_indicator_2_' + key] = np.clip(0.1 - o[key] * 10 + 0.5,
+                                                    0., 1.)
+
+        # Tranformer
+        core_matrix = self.get_core_matrix(o['pelvis_y_r'])
+        pelvis_pos = np.array([o['pelvis_x'], o['pelvis_y'],
+                               o['pelvis_z']]).reshape((3, 1))
+        pelvis_vel = np.array(
+            [o['pelvis_v_x'], o['pelvis_v_y'], o['pelvis_v_z']]).reshape((3,
+                                                                          1))
+        for body_part in [
+                'mass', 'femur_r', 'pros_tibia_r', 'pros_foot_r', 'femur_l',
+                'tibia_l', 'talus_l', 'calcn_l', 'toes_l', 'torso', 'head'
+        ]:
+            # rotation
+            if body_part != 'mass':
+                o[body_part + '_y_r'] -= o['pelvis_y_r']
+                o[body_part + '_v_y_r'] -= o['pelvis_v_y_r']
+            # position/velocity
+            global_pos = []
+            global_vel = []
+            for each in ['_x', '_y', '_z']:
+                global_pos.append(o[body_part + each])
+                global_vel.append(o[body_part + '_v' + each])
+            global_pos = np.array(global_pos).reshape((3, 1))
+            global_vel = np.array(global_vel).reshape((3, 1))
+            pelvis_rel_pos = core_matrix.dot(global_pos - pelvis_pos)
+            w = o['pelvis_v_y_r']
+            offset = np.array(
+                [-w * pelvis_rel_pos[2], 0, w * pelvis_rel_pos[0]])
+            pelvis_rel_vel = core_matrix.dot(global_vel - pelvis_vel) + offset
+            for i, each in enumerate(['_x', '_y', '_z']):
+                o[body_part + each] = pelvis_rel_pos[i][0]
+                o[body_part + '_v' + each] = pelvis_rel_vel[i][0]
+
+        for key in ['pelvis_x', 'pelvis_z', 'pelvis_y_r']:
+            del o[key]
+
+        current_v = np.array(state_desc['body_vel']['pelvis']).reshape((3, 1))
+        pelvis_current_v = core_matrix.dot(current_v)
+        o['pelvis_v_x'] = pelvis_current_v[0]
+        o['pelvis_v_z'] = pelvis_current_v[2]
+
+        res = np.array(list(o.values()))
+        res = self.feature_normalize(
+            res, mean=self.mean, std=self.std, duplicate_id=self.duplicate_id)
+
+        feet_dis = ((o['tibia_l_x'] - o['pros_tibia_r_x'])**2 +
+                    (o['tibia_l_z'] - o['pros_tibia_r_z'])**2)**0.5
+        res = np.append(res, feet_dis)
+        remaining_time = (self.step_fea -
+                          (MAXTIME_LIMIT / 2.0)) / (MAXTIME_LIMIT / 2.0) * -1.0
+        #logger.info('remaining_time fea: {}'.format(remaining_time))
+        res = np.append(res, remaining_time)
+
+        # target driven
+        target_v = np.array(state_desc['target_vel']).reshape((3, 1))
+        pelvis_target_v = core_matrix.dot(target_v)
+        diff_vel_x = pelvis_target_v[0] - pelvis_current_v[0]
+        diff_vel_z = pelvis_target_v[2] - pelvis_current_v[2]
+        diff_vel = np.sqrt(pelvis_target_v[0] ** 2 + pelvis_target_v[2] ** 2) - \
+                   np.sqrt(pelvis_current_v[0] ** 2 + pelvis_current_v[2] ** 2)
+
+        target_vel_x = target_v[0]
+        target_vel_z = target_v[2]
+        target_theta = math.atan(-1.0 * target_vel_z / target_vel_x)
+        current_theta = state_desc['body_pos_rot']['pelvis'][1]
+        diff_theta = target_theta - current_theta
+        res = np.append(res, [
+            diff_vel_x[0] / 3.0, diff_vel_z[0] / 3.0, diff_vel[0] / 3.0,
+            diff_theta / (np.pi * 3 / 8)
+        ])
+
+        return res
+
+
+if __name__ == '__main__':
+    from osim.env import ProstheticsEnv
+
+    env = ProstheticsEnv(visualize=False)
+    env.change_model(model='3D', difficulty=1, prosthetic=True)
+    env = ForwardReward(env)
+    env = FrameSkip(env, 4)
+    env = ActionScale(env)
+    env = PelvisBasedObs(env)
+    for i in range(64):
+        observation = env.reset(project=False)

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/final_submit/test.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import numpy as np
+import time
+from env_wrapper import FrameSkip, ActionScale, PelvisBasedObs, ForwardReward
+from osim.env import ProstheticsEnv
+from parl.utils import logger
+from submit_model import SubmitModel
+
+
+def play_multi_episode(submit_model, episode_num=2, vis=False, seed=0):
+    np.random.seed(seed)
+    env = ProstheticsEnv(visualize=vis)
+    env.change_model(model='3D', difficulty=1, prosthetic=True, seed=seed)
+    env = ForwardReward(env)
+    env = FrameSkip(env, 4)
+    env = ActionScale(env)
+    env = PelvisBasedObs(env)
+    all_reward = []
+    all_shaping_reward = 0
+    last_frames_count = 0
+
+    for e in range(episode_num):
+        t = time.time()
+        episode_reward = 0.0
+        episode_shaping_reward = 0.0
+        observation = env.reset(project=False)
+        target_change_times = 0
+        step = 0
+        loss = []
+        while True:
+            step += 1
+            action = submit_model.pred_batch(observation, target_change_times)
+            observation, reward, done, info = env.step(action, project=False)
+            step_frames = info['frame_count'] - last_frames_count
+            last_frames_count = info['frame_count']
+            episode_reward += reward
+            # we pacle it here to drop the first step after changing
+            if target_change_times >= 1:
+                loss.append(10 * step_frames - reward)
+            if info['target_changed']:
+                target_change_times = min(target_change_times + 1, 3)
+            logger.info("[step/{}]reward:{}  info:{}".format(
+                step, reward, info))
+            episode_shaping_reward += info['shaping_reward']
+            if done:
+                break
+        all_reward.append(episode_reward)
+        all_shaping_reward += episode_shaping_reward
+        t = time.time() - t
+        logger.info(
+            "[episode/{}] time: {} episode_reward:{} change_loss:{} after_change_loss:{} mean_reward:{}"
+            .format(e, t, episode_reward, np.sum(loss[:15]), np.sum(loss[15:]),
+                    np.mean(all_reward)))
+    logger.info("Mean reward:{}".format(np.mean(all_reward)))
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        '--use_cuda', action="store_true", help='If set, will run in gpu 0')
+    parser.add_argument(
+        '--vis', action="store_true", help='If set, will visualize.')
+    parser.add_argument('--seed', type=int, default=0, help='Random seed.')
+    parser.add_argument(
+        '--episode_num', type=int, default=1, help='Episode number to run.')
+    args = parser.parse_args()
+
+    submit_model = SubmitModel(use_cuda=args.use_cuda)
+
+    play_multi_episode(
+        submit_model,
+        episode_num=args.episode_num,
+        vis=args.vis,
+        seed=args.seed)

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/multi_head_ddpg.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import parl.layers as layers
+from copy import deepcopy
+from paddle import fluid
+from parl.framework.algorithm_base import Algorithm
+
+__all__ = ['MultiHeadDDPG']
+
+
+class MultiHeadDDPG(Algorithm):
+    def __init__(self, models, hyperparas):
+        """ model: should implement the function get_actor_params()
+        """
+        self.models = models
+        self.target_models = []
+        for model in models:
+            target_model = deepcopy(model)
+            self.target_models.append(target_model)
+
+        # fetch hyper parameters
+        self.gamma = hyperparas['gamma']
+        self.tau = hyperparas['tau']
+        self.ensemble_num = hyperparas['ensemble_num']
+
+    def define_predict(self, obs, model_id):
+        """ use actor model of self.models[model_id] to predict the action
+        """
+        return self.models[model_id].policy(obs)
+
+    def define_ensemble_predict(self, obs):
+        """ ensemble predict:
+        1. For actions of all actors, each critic will score them
+           and normalize its scores;
+        2. For each actor, will calculate its score by 
+           average scores given by all critics
+        3. choose action of the actor whose score is best
+        """
+        actor_outputs = []
+        for i in range(self.ensemble_num):
+            actor_outputs.append(self.models[i].policy(obs))
+        batch_actions = layers.concat(actor_outputs, axis=0)
+        batch_obs = layers.expand(obs, expand_times=[self.ensemble_num, 1])
+
+        critic_outputs = []
+        for i in range(self.ensemble_num):
+            critic_output = self.models[i].value(batch_obs, batch_actions)
+            critic_output = layers.unsqueeze(critic_output, axes=[1])
+            critic_outputs.append(critic_output)
+        score_matrix = layers.concat(critic_outputs, axis=1)
+
+        # Normalize scores given by each critic
+        sum_critic_score = layers.reduce_sum(
+            score_matrix, dim=0, keep_dim=True)
+        sum_critic_score = layers.expand(
+            sum_critic_score, expand_times=[self.ensemble_num, 1])
+        norm_score_matrix = score_matrix / sum_critic_score
+
+        actions_mean_score = layers.reduce_mean(
+            norm_score_matrix, dim=1, keep_dim=True)
+        best_score_id = layers.argmax(actions_mean_score, axis=0)
+        best_score_id = layers.cast(best_score_id, dtype='int32')
+        ensemble_predict_action = layers.gather(batch_actions, best_score_id)
+        return ensemble_predict_action
+
+    def define_learn(self, obs, action, reward, next_obs, terminal, actor_lr,
+                     critic_lr, model_id):
+        """ update actor and critic model of self.models[model_id] with DDPG algorithm
+        """
+        actor_cost = self._actor_learn(obs, actor_lr, model_id)
+        critic_cost = self._critic_learn(obs, action, reward, next_obs,
+                                         terminal, critic_lr, model_id)
+        return actor_cost, critic_cost
+
+    def _actor_learn(self, obs, actor_lr, model_id):
+        action = self.models[model_id].policy(obs)
+        Q = self.models[model_id].value(obs, action)
+        cost = layers.reduce_mean(-1.0 * Q)
+        optimizer = fluid.optimizer.AdamOptimizer(actor_lr)
+        optimizer.minimize(
+            cost, parameter_list=self.models[model_id].get_actor_params())
+        return cost
+
+    def _critic_learn(self, obs, action, reward, next_obs, terminal, critic_lr,
+                      model_id):
+        next_action = self.target_models[model_id].policy(next_obs)
+        next_Q = self.target_models[model_id].value(next_obs, next_action)
+
+        terminal = layers.cast(terminal, dtype='float32')
+        target_Q = reward + (1.0 - terminal) * self.gamma * next_Q
+        target_Q.stop_gradient = True
+
+        Q = self.models[model_id].value(obs, action)
+        cost = layers.square_error_cost(Q, target_Q)
+        cost = layers.reduce_mean(cost)
+        optimizer = fluid.optimizer.AdamOptimizer(critic_lr)
+        optimizer.minimize(cost)
+        return cost
+
+    def sync_target(self,
+                    gpu_id,
+                    model_id,
+                    decay=None,
+                    share_vars_parallel_executor=None):
+        if decay is None:
+            decay = 1.0 - self.tau
+        self.models[model_id].sync_params_to(
+            self.target_models[model_id],
+            gpu_id=gpu_id,
+            decay=decay,
+            share_vars_parallel_executor=share_vars_parallel_executor)

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/opensim_agent.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import parl.layers as layers
+import re
+from paddle import fluid
+from paddle.fluid.executor import _fetch_var
+from parl.framework.agent_base import Agent
+from parl.utils import logger
+
+
+class OpenSimAgent(Agent):
+    def __init__(self, algorithm, obs_dim, act_dim, ensemble_num):
+        self.obs_dim = obs_dim
+        self.act_dim = act_dim
+        self.ensemble_num = ensemble_num
+        super(OpenSimAgent, self).__init__(algorithm)
+
+        # Use ParallelExecutor to make program running faster
+        use_cuda = True if self.gpu_id >= 0 else False
+        self.learn_pe = []
+        self.pred_pe = []
+
+        exec_strategy = fluid.ExecutionStrategy()
+        exec_strategy.use_experimental_executor = True
+        exec_strategy.num_threads = 4
+        build_strategy = fluid.BuildStrategy()
+        build_strategy.remove_unnecessary_lock = True
+
+        for i in range(self.ensemble_num):
+            with fluid.scope_guard(fluid.global_scope().new_scope()):
+                pe = fluid.ParallelExecutor(
+                    use_cuda=use_cuda,
+                    main_program=self.learn_programs[i],
+                    exec_strategy=exec_strategy,
+                    build_strategy=build_strategy)
+                self.learn_pe.append(pe)
+
+            with fluid.scope_guard(fluid.global_scope().new_scope()):
+                pe = fluid.ParallelExecutor(
+                    use_cuda=use_cuda,
+                    main_program=self.predict_programs[i],
+                    exec_strategy=exec_strategy,
+                    build_strategy=build_strategy)
+                self.pred_pe.append(pe)
+
+            # Attention: In the beginning, sync target model totally.
+            self.alg.sync_target(
+                gpu_id=self.gpu_id,
+                model_id=i,
+                decay=1.0,
+                share_vars_parallel_executor=self.learn_pe[i])
+            # Do cache, will create ParallelExecutor of sync params in advance
+            # If not, there are some issues when ensemble_num > 1
+            self.alg.sync_target(
+                gpu_id=self.gpu_id,
+                model_id=i,
+                share_vars_parallel_executor=self.learn_pe[i])
+
+        with fluid.scope_guard(fluid.global_scope().new_scope()):
+            self.ensemble_predict_pe = fluid.ParallelExecutor(
+                use_cuda=use_cuda,
+                main_program=self.ensemble_predict_program,
+                exec_strategy=exec_strategy,
+                build_strategy=build_strategy)
+
+    def build_program(self):
+        self.predict_programs = []
+        self.predict_outputs = []
+        self.learn_programs = []
+        self.learn_programs_output = []
+        for i in range(self.ensemble_num):
+            predict_program = fluid.Program()
+            with fluid.program_guard(predict_program):
+                obs = layers.data(
+                    name='obs', shape=[self.obs_dim], dtype='float32')
+                act = self.alg.define_predict(obs, model_id=i)
+            self.predict_programs.append(predict_program)
+            self.predict_outputs.append([act.name])
+
+            learn_program = fluid.Program()
+            with fluid.program_guard(learn_program):
+                obs = layers.data(
+                    name='obs', shape=[self.obs_dim], dtype='float32')
+                act = layers.data(
+                    name='act', shape=[self.act_dim], dtype='float32')
+                reward = layers.data(name='reward', shape=[], dtype='float32')
+                next_obs = layers.data(
+                    name='next_obs', shape=[self.obs_dim], dtype='float32')
+                terminal = layers.data(name='terminal', shape=[], dtype='bool')
+                actor_lr = layers.data(
+                    name='actor_lr',
+                    shape=[1],
+                    dtype='float32',
+                    append_batch_size=False)
+                critic_lr = layers.data(
+                    name='critic_lr',
+                    shape=[1],
+                    dtype='float32',
+                    append_batch_size=False)
+                actor_loss, critic_loss = self.alg.define_learn(
+                    obs,
+                    act,
+                    reward,
+                    next_obs,
+                    terminal,
+                    actor_lr,
+                    critic_lr,
+                    model_id=i)
+            self.learn_programs.append(learn_program)
+            self.learn_programs_output.append([critic_loss.name])
+
+        self.ensemble_predict_program = fluid.Program()
+        with fluid.program_guard(self.ensemble_predict_program):
+            obs = layers.data(
+                name='obs', shape=[self.obs_dim], dtype='float32')
+            act = self.alg.define_ensemble_predict(obs)
+        self.ensemble_predict_output = [act.name]
+
+    def predict(self, obs, model_id):
+        feed = {'obs': obs}
+        feed = [feed]
+        act = self.pred_pe[model_id].run(
+            feed=feed, fetch_list=self.predict_outputs[model_id])[0]
+        return act
+
+    def ensemble_predict(self, obs):
+        feed = {'obs': obs}
+        feed = [feed]
+        act = self.ensemble_predict_pe.run(
+            feed=feed, fetch_list=self.ensemble_predict_output)[0]
+        return act
+
+    def learn(self, obs, act, reward, next_obs, terminal, actor_lr, critic_lr,
+              model_id):
+        feed = {
+            'obs': obs,
+            'act': act,
+            'reward': reward,
+            'next_obs': next_obs,
+            'terminal': terminal,
+            'actor_lr': actor_lr,
+            'critic_lr': critic_lr
+        }
+
+        feed = [feed]
+        critic_loss = self.learn_pe[model_id].run(
+            feed=feed, fetch_list=self.learn_programs_output[model_id])[0]
+        self.alg.sync_target(
+            gpu_id=self.gpu_id,
+            model_id=model_id,
+            share_vars_parallel_executor=self.learn_pe[model_id])
+        return critic_loss
+
+    def save_params(self, dirname):
+        for i in range(self.ensemble_num):
+            fluid.io.save_params(
+                executor=self.fluid_executor,
+                dirname=dirname,
+                main_program=self.learn_programs[i])
+
+    def load_params(self, dirname, from_one_head):
+        if from_one_head:
+            logger.info('[From one head, extend to multi head:]')
+            # load model 0
+            fluid.io.load_params(
+                executor=self.fluid_executor,
+                dirname=dirname,
+                main_program=self.learn_programs[0])
+
+            # sync identity params of model/target_model 0 to other models/target_models
+            for i in range(1, self.ensemble_num):
+                params = list(
+                    filter(
+                        lambda x: 'identity' in x.name and '@GRAD' not in x.name,
+                        self.learn_programs[i].list_vars()))
+                for param in params:
+                    param_var = _fetch_var(param.name, return_numpy=False)
+
+                    model0_name = re.sub(r"identity_\d+", "identity_0",
+                                         param.name)
+                    model0_value = _fetch_var(model0_name, return_numpy=True)
+                    logger.info('{} -> {}'.format(model0_name, param.name))
+                    param_var.set(model0_value, self.place)
+
+            # sync share params of target_model 0 to other target models
+            # After deepcopy, shapre params between target models  is different
+            for i in range(1, self.ensemble_num):
+                params = list(
+                    filter(
+                        lambda x: 'shared' in x.name and 'PARL_target' in x.name and '@GRAD' not in x.name,
+                        self.learn_programs[i].list_vars()))
+                for param in params:
+                    param_var = _fetch_var(param.name, return_numpy=False)
+
+                    model0_name = re.sub(r"_\d+$", "_0", param.name)
+                    model0_value = _fetch_var(model0_name, return_numpy=True)
+                    logger.info('{} -> {}'.format(model0_name, param.name))
+                    param_var.set(model0_value, self.place)
+
+        else:
+            for i in range(self.ensemble_num):
+                fluid.io.load_params(
+                    executor=self.fluid_executor,
+                    dirname=dirname,
+                    main_program=self.learn_programs[i])

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/opensim_model.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import parl.layers as layers
+from paddle import fluid
+from paddle.fluid.param_attr import ParamAttr
+from parl.framework.model_base import Model
+
+
+class OpenSimModel(Model):
+    def __init__(self, obs_dim, vel_obs_dim, act_dim, model_id=0, shared=True):
+        self.actor_model = ActorModel(obs_dim, vel_obs_dim, act_dim, model_id,
+                                      shared)
+        self.critic_model = CriticModel(obs_dim, vel_obs_dim, act_dim,
+                                        model_id, shared)
+
+    def policy(self, obs):
+        return self.actor_model.policy(obs)
+
+    def value(self, obs, action):
+        return self.critic_model.value(obs, action)
+
+    def get_actor_params(self):
+        return self.actor_model.parameter_names
+
+
+class ActorModel(Model):
+    def __init__(self, obs_dim, vel_obs_dim, act_dim, model_id, shared):
+        hid0_size = 800
+        hid1_size = 400
+        hid2_size = 200
+        vel_hid0_size = 200
+        vel_hid1_size = 400
+
+        self.obs_dim = obs_dim
+        self.vel_obs_dim = vel_obs_dim
+
+        # bottom layers
+        if shared:
+            scope_name = 'policy_shared'
+        else:
+            scope_name = 'policy_identity_{}'.format(model_id)
+
+        self.fc0 = layers.fc(
+            size=hid0_size,
+            act='tanh',
+            param_attr=ParamAttr(name='{}/h0/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/h0/b'.format(scope_name)))
+        self.fc1 = layers.fc(
+            size=hid1_size,
+            act='tanh',
+            param_attr=ParamAttr(name='{}/h1/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/h1/b'.format(scope_name)))
+        self.vel_fc0 = layers.fc(
+            size=vel_hid0_size,
+            act='tanh',
+            param_attr=ParamAttr(name='{}/vel_h0/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/vel_h0/b'.format(scope_name)))
+        self.vel_fc1 = layers.fc(
+            size=vel_hid1_size,
+            act='tanh',
+            param_attr=ParamAttr(name='{}/vel_h1/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/vel_h1/b'.format(scope_name)))
+
+        # top layers
+        scope_name = 'policy_identity_{}'.format(model_id)
+
+        self.fc2 = layers.fc(
+            size=hid2_size,
+            act='tanh',
+            param_attr=ParamAttr(name='{}/h2/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/h2/b'.format(scope_name)))
+        self.fc3 = layers.fc(
+            size=act_dim,
+            act='tanh',
+            param_attr=ParamAttr(name='{}/means/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/means/b'.format(scope_name)))
+
+    def policy(self, obs):
+        real_obs = layers.slice(
+            obs, axes=[1], starts=[0], ends=[-self.vel_obs_dim])
+        # target related fetures
+        vel_obs = layers.slice(
+            obs, axes=[1], starts=[-self.vel_obs_dim], ends=[self.obs_dim])
+
+        hid0 = self.fc0(real_obs)
+        hid1 = self.fc1(hid0)
+        vel_hid0 = self.vel_fc0(vel_obs)
+        vel_hid1 = self.vel_fc1(vel_hid0)
+        concat = layers.concat([hid1, vel_hid1], axis=1)
+        hid2 = self.fc2(concat)
+        means = self.fc3(hid2)
+        return means
+
+
+class CriticModel(Model):
+    def __init__(self, obs_dim, vel_obs_dim, act_dim, model_id, shared):
+        super(CriticModel, self).__init__()
+        hid0_size = 800
+        hid1_size = 400
+        vel_hid0_size = 200
+        vel_hid1_size = 400
+
+        self.obs_dim = obs_dim
+        self.vel_obs_dim = vel_obs_dim
+
+        # buttom layers
+        if shared:
+            scope_name = 'critic_shared'
+        else:
+            scope_name = 'critic_identity_{}'.format(model_id)
+
+        self.fc0 = layers.fc(
+            size=hid0_size,
+            act='selu',
+            param_attr=ParamAttr(name='{}/w1/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/w1/b'.format(scope_name)))
+        self.fc1 = layers.fc(
+            size=hid1_size,
+            act='selu',
+            param_attr=ParamAttr(name='{}/h1/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/h1/b'.format(scope_name)))
+        self.vel_fc0 = layers.fc(
+            size=vel_hid0_size,
+            act='selu',
+            param_attr=ParamAttr(name='{}/vel_h0/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/vel_h0/b'.format(scope_name)))
+        self.vel_fc1 = layers.fc(
+            size=vel_hid1_size,
+            act='selu',
+            param_attr=ParamAttr(name='{}/vel_h1/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/vel_h1/b'.format(scope_name)))
+        self.act_fc0 = layers.fc(
+            size=hid1_size,
+            act='selu',
+            param_attr=ParamAttr(name='{}/a1/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/a1/b'.format(scope_name)))
+
+        # top layers
+        scope_name = 'critic_identity_{}'.format(model_id)
+
+        self.fc2 = layers.fc(
+            size=hid1_size,
+            act='selu',
+            param_attr=ParamAttr(name='{}/h3/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/h3/b'.format(scope_name)))
+        self.fc3 = layers.fc(
+            size=1,
+            act='selu',
+            param_attr=ParamAttr(name='{}/value/W'.format(scope_name)),
+            bias_attr=ParamAttr(name='{}/value/b'.format(scope_name)))
+
+    def value(self, obs, action):
+        real_obs = layers.slice(
+            obs, axes=[1], starts=[0], ends=[-self.vel_obs_dim])
+        # target related fetures
+        vel_obs = layers.slice(
+            obs, axes=[1], starts=[-self.vel_obs_dim], ends=[self.obs_dim])
+
+        hid0 = self.fc0(real_obs)
+        hid1 = self.fc1(hid0)
+        vel_hid0 = self.vel_fc0(vel_obs)
+        vel_hid1 = self.vel_fc1(vel_hid0)
+        a1 = self.act_fc0(action)
+        concat = layers.concat([hid1, a1, vel_hid1], axis=1)
+        hid2 = self.fc2(concat)
+        Q = self.fc3(hid2)
+        Q = layers.squeeze(Q, axes=[1])
+        return Q

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/simulator_client.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import grpc
+import json
+import numpy as np
+import simulator_pb2
+import simulator_pb2_grpc
+from args import get_client_args
+from env_wrapper import FrameSkip, ActionScale, PelvisBasedObs, MAXTIME_LIMIT, CustomR2Env, RunFastestReward, FixedTargetSpeedReward, Round2Reward
+from osim.env import ProstheticsEnv
+from parl.utils import logger
+
+ProstheticsEnv.time_limit = MAXTIME_LIMIT
+
+
+class Worker(object):
+    def __init__(self, server_ip='localhost', server_port=5007):
+        if args.ident is not None:
+            self.worker_id = args.ident
+        else:
+            self.worker_id = np.random.randint(int(1e18))
+
+        self.address = '{}:{}'.format(server_ip, server_port)
+
+        random_seed = int(self.worker_id % int(1e9))
+        np.random.seed(random_seed)
+
+        env = ProstheticsEnv(visualize=False, seed=random_seed)
+        env.change_model(
+            model='3D', difficulty=1, prosthetic=True, seed=random_seed)
+        env.spec.timestep_limit = MAXTIME_LIMIT
+        env = CustomR2Env(env)
+
+        if args.reward_type == 'RunFastest':
+            env = RunFastestReward(env)
+        elif args.reward_type == 'FixedTargetSpeed':
+            env = FixedTargetSpeedReward(
+                env, args.target_v, args.act_penalty_lowerbound,
+                args.act_penalty_coeff, args.vel_penalty_coeff)
+        elif args.reward_type == 'Round2':
+            env = Round2Reward(env, args.act_penalty_lowerbound,
+                               args.act_penalty_coeff, args.vel_penalty_coeff)
+        else:
+            assert False, 'Not supported reward type!'
+
+        env = FrameSkip(env, 4)
+        env = ActionScale(env)
+        self.env = PelvisBasedObs(env)
+
+    def run(self):
+        observation = self.env.reset(project=False, stage=args.stage)
+        reward = 0
+        done = False
+        info = {'shaping_reward': 0.0}
+        info['first'] = True
+        with grpc.insecure_channel(self.address) as channel:
+            stub = simulator_pb2_grpc.SimulatorStub(channel)
+            while True:
+                response = stub.Send(
+                    simulator_pb2.Request(
+                        observation=observation,
+                        reward=reward,
+                        done=done,
+                        info=json.dumps(info),
+                        id=self.worker_id))
+
+                extra = json.loads(response.extra)
+
+                if 'reset' in extra and extra['reset']:
+                    logger.info('Server require to reset!')
+                    observation = self.env.reset(
+                        project=False, stage=args.stage)
+                    reward = 0
+                    done = False
+                    info = {'shaping_reward': 0.0}
+                    continue
+
+                if 'shutdown' in extra and extra['shutdown']:
+                    break
+
+                action = np.array(response.action)
+                next_observation, reward, done, info = self.env.step(
+                    action, project=False)
+
+                # debug info
+                if args.debug:
+                    logger.info("dim:{} obs:{}   act:{}   reward:{}  done:{}  info:{}".format(\
+                        len(observation), np.sum(observation), np.sum(action), reward, done, info))
+                observation = next_observation
+                if done:
+                    observation = self.env.reset(
+                        project=False, stage=args.stage)
+
+                    # the last observation should be used to compute append_value in simulator_server
+                    info['last_obs'] = next_observation.tolist()
+
+
+if __name__ == '__main__':
+    args = get_client_args()
+
+    worker = Worker(server_ip=args.ip, server_port=args.port)
+
+    worker.run()

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/simulator_pb2.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Generated by the protocol buffer compiler.  DO NOT EDIT!
+# source: simulator.proto
+
+import sys
+_b = sys.version_info[0] < 3 and (lambda x: x) or (
+    lambda x: x.encode('latin1'))
+from google.protobuf import descriptor as _descriptor
+from google.protobuf import message as _message
+from google.protobuf import reflection as _reflection
+from google.protobuf import symbol_database as _symbol_database
+from google.protobuf import descriptor_pb2
+# @@protoc_insertion_point(imports)
+
+_sym_db = _symbol_database.Default()
+
+DESCRIPTOR = _descriptor.FileDescriptor(
+    name='simulator.proto',
+    package='simulator',
+    syntax='proto3',
+    serialized_pb=_b(
+        '\n\x0fsimulator.proto\x12\tsimulator\"V\n\x07Request\x12\x13\n\x0bobservation\x18\x01 \x03(\x01\x12\x0e\n\x06reward\x18\x02 \x01(\x01\x12\x0c\n\x04\x64one\x18\x03 \x01(\x08\x12\x0c\n\x04info\x18\x04 \x01(\t\x12\n\n\x02id\x18\x05 \x01(\x03\"&\n\x05Reply\x12\x0e\n\x06\x61\x63tion\x18\x01 \x03(\x01\x12\r\n\x05\x65xtra\x18\x02 \x01(\t2;\n\tSimulator\x12.\n\x04Send\x12\x12.simulator.Request\x1a\x10.simulator.Reply\"\x00\x62\x06proto3'
+    ))
+
+_REQUEST = _descriptor.Descriptor(
+    name='Request',
+    full_name='simulator.Request',
+    filename=None,
+    file=DESCRIPTOR,
+    containing_type=None,
+    fields=[
+        _descriptor.FieldDescriptor(
+            name='observation',
+            full_name='simulator.Request.observation',
+            index=0,
+            number=1,
+            type=1,
+            cpp_type=5,
+            label=3,
+            has_default_value=False,
+            default_value=[],
+            message_type=None,
+            enum_type=None,
+            containing_type=None,
+            is_extension=False,
+            extension_scope=None,
+            options=None,
+            file=DESCRIPTOR),
+        _descriptor.FieldDescriptor(
+            name='reward',
+            full_name='simulator.Request.reward',
+            index=1,
+            number=2,
+            type=1,
+            cpp_type=5,
+            label=1,
+            has_default_value=False,
+            default_value=float(0),
+            message_type=None,
+            enum_type=None,
+            containing_type=None,
+            is_extension=False,
+            extension_scope=None,
+            options=None,
+            file=DESCRIPTOR),
+        _descriptor.FieldDescriptor(
+            name='done',
+            full_name='simulator.Request.done',
+            index=2,
+            number=3,
+            type=8,
+            cpp_type=7,
+            label=1,
+            has_default_value=False,
+            default_value=False,
+            message_type=None,
+            enum_type=None,
+            containing_type=None,
+            is_extension=False,
+            extension_scope=None,
+            options=None,
+            file=DESCRIPTOR),
+        _descriptor.FieldDescriptor(
+            name='info',
+            full_name='simulator.Request.info',
+            index=3,
+            number=4,
+            type=9,
+            cpp_type=9,
+            label=1,
+            has_default_value=False,
+            default_value=_b("").decode('utf-8'),
+            message_type=None,
+            enum_type=None,
+            containing_type=None,
+            is_extension=False,
+            extension_scope=None,
+            options=None,
+            file=DESCRIPTOR),
+        _descriptor.FieldDescriptor(
+            name='id',
+            full_name='simulator.Request.id',
+            index=4,
+            number=5,
+            type=3,
+            cpp_type=2,
+            label=1,
+            has_default_value=False,
+            default_value=0,
+            message_type=None,
+            enum_type=None,
+            containing_type=None,
+            is_extension=False,
+            extension_scope=None,
+            options=None,
+            file=DESCRIPTOR),
+    ],
+    extensions=[],
+    nested_types=[],
+    enum_types=[],
+    options=None,
+    is_extendable=False,
+    syntax='proto3',
+    extension_ranges=[],
+    oneofs=[],
+    serialized_start=30,
+    serialized_end=116,
+)
+
+_REPLY = _descriptor.Descriptor(
+    name='Reply',
+    full_name='simulator.Reply',
+    filename=None,
+    file=DESCRIPTOR,
+    containing_type=None,
+    fields=[
+        _descriptor.FieldDescriptor(
+            name='action',
+            full_name='simulator.Reply.action',
+            index=0,
+            number=1,
+            type=1,
+            cpp_type=5,
+            label=3,
+            has_default_value=False,
+            default_value=[],
+            message_type=None,
+            enum_type=None,
+            containing_type=None,
+            is_extension=False,
+            extension_scope=None,
+            options=None,
+            file=DESCRIPTOR),
+        _descriptor.FieldDescriptor(
+            name='extra',
+            full_name='simulator.Reply.extra',
+            index=1,
+            number=2,
+            type=9,
+            cpp_type=9,
+            label=1,
+            has_default_value=False,
+            default_value=_b("").decode('utf-8'),
+            message_type=None,
+            enum_type=None,
+            containing_type=None,
+            is_extension=False,
+            extension_scope=None,
+            options=None,
+            file=DESCRIPTOR),
+    ],
+    extensions=[],
+    nested_types=[],
+    enum_types=[],
+    options=None,
+    is_extendable=False,
+    syntax='proto3',
+    extension_ranges=[],
+    oneofs=[],
+    serialized_start=118,
+    serialized_end=156,
+)
+
+DESCRIPTOR.message_types_by_name['Request'] = _REQUEST
+DESCRIPTOR.message_types_by_name['Reply'] = _REPLY
+_sym_db.RegisterFileDescriptor(DESCRIPTOR)
+
+Request = _reflection.GeneratedProtocolMessageType(
+    'Request',
+    (_message.Message, ),
+    dict(
+        DESCRIPTOR=_REQUEST,
+        __module__='simulator_pb2'
+        # @@protoc_insertion_point(class_scope:simulator.Request)
+    ))
+_sym_db.RegisterMessage(Request)
+
+Reply = _reflection.GeneratedProtocolMessageType(
+    'Reply',
+    (_message.Message, ),
+    dict(
+        DESCRIPTOR=_REPLY,
+        __module__='simulator_pb2'
+        # @@protoc_insertion_point(class_scope:simulator.Reply)
+    ))
+_sym_db.RegisterMessage(Reply)
+
+_SIMULATOR = _descriptor.ServiceDescriptor(
+    name='Simulator',
+    full_name='simulator.Simulator',
+    file=DESCRIPTOR,
+    index=0,
+    options=None,
+    serialized_start=158,
+    serialized_end=217,
+    methods=[
+        _descriptor.MethodDescriptor(
+            name='Send',
+            full_name='simulator.Simulator.Send',
+            index=0,
+            containing_service=None,
+            input_type=_REQUEST,
+            output_type=_REPLY,
+            options=None,
+        ),
+    ])
+_sym_db.RegisterServiceDescriptor(_SIMULATOR)
+
+DESCRIPTOR.services_by_name['Simulator'] = _SIMULATOR
+
+# @@protoc_insertion_point(module_scope)

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/simulator_pb2_grpc.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Generated by the gRPC Python protocol compiler plugin. DO NOT EDIT!
+import grpc
+
+import simulator_pb2 as simulator__pb2
+
+
+class SimulatorStub(object):
+    """The greeting service definition.
+  """
+
+    def __init__(self, channel):
+        """Constructor.
+
+    Args:
+      channel: A grpc.Channel.
+    """
+        self.Send = channel.unary_unary(
+            '/simulator.Simulator/Send',
+            request_serializer=simulator__pb2.Request.SerializeToString,
+            response_deserializer=simulator__pb2.Reply.FromString,
+        )
+
+
+class SimulatorServicer(object):
+    """The greeting service definition.
+  """
+
+    def Send(self, request, context):
+        """Request Action
+    """
+        context.set_code(grpc.StatusCode.UNIMPLEMENTED)
+        context.set_details('Method not implemented!')
+        raise NotImplementedError('Method not implemented!')
+
+
+def add_SimulatorServicer_to_server(servicer, server):
+    rpc_method_handlers = {
+        'Send':
+        grpc.unary_unary_rpc_method_handler(
+            servicer.Send,
+            request_deserializer=simulator__pb2.Request.FromString,
+            response_serializer=simulator__pb2.Reply.SerializeToString,
+        ),
+    }
+    generic_handler = grpc.method_handlers_generic_handler(
+        'simulator.Simulator', rpc_method_handlers)
+    server.add_generic_rpc_handlers((generic_handler, ))

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/simulator_server.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import grpc
+import json
+import numpy as np
+import os
+import queue
+import simulator_pb2
+import simulator_pb2_grpc
+import six
+import time
+import threading
+from args import get_server_args
+from collections import defaultdict
+from concurrent import futures
+from multi_head_ddpg import MultiHeadDDPG
+from opensim_agent import OpenSimAgent
+from opensim_model import OpenSimModel
+from parl.utils import logger, ReplayMemory
+from utils import calc_indicators, ScalarsManager, TransitionExperience
+
+ACT_DIM = 19
+VEL_DIM = 4
+OBS_DIM = 185 + VEL_DIM
+GAMMA = 0.96
+TAU = 0.001
+ACTOR_LR = 3e-5
+CRITIC_LR = 3e-5
+TRAIN_TIMES = 100
+BATCH_SIZE = 128
+NOISE_DECAY = 0.999998
+_ONE_DAY_IN_SECONDS = 60 * 60 * 24
+
+
+class SimulatorServer(simulator_pb2_grpc.SimulatorServicer):
+    class ClientState(object):
+        def __init__(self):
+            self.memory = []  # list of Experience
+            self.ident = None
+            self.model_idx = np.random.randint(args.ensemble_num)
+            self.last_target_changed = 0
+            self.target_change_times = 0
+
+        def reset(self):
+            self.last_target_changed = 0
+            self.memory = []
+            self.model_idx = np.random.randint(args.ensemble_num)
+            self.target_change_times = 0
+
+        def update_last_target_changed(self):
+            self.last_target_changed = len(self.memory)
+
+    def __init__(self):
+        self.rpm = ReplayMemory(int(2e6), OBS_DIM, ACT_DIM)
+
+        # Need acquire lock when model learning or predicting
+        self.locks = []
+        for i in range(args.ensemble_num):
+            self.locks.append(threading.Lock())
+
+        models = []
+        for i in range(args.ensemble_num):
+            models.append(OpenSimModel(OBS_DIM, VEL_DIM, ACT_DIM, model_id=i))
+
+        hyperparas = {
+            'gamma': GAMMA,
+            'tau': TAU,
+            'ensemble_num': args.ensemble_num
+        }
+        alg = MultiHeadDDPG(models, hyperparas)
+
+        self.agent = OpenSimAgent(alg, OBS_DIM, ACT_DIM, args.ensemble_num)
+
+        self.scalars_manager = ScalarsManager(logger.get_dir())
+
+        # add lock when appending data to rpm or writing scalars to tensorboard
+        self.MEMORY_LOCK = threading.Lock()
+
+        self.clients = defaultdict(self.ClientState)
+
+        self.ready_client_queue = queue.Queue()
+
+        self.noiselevel = 0.5
+        self.global_step = 0
+
+        # thread to keep training
+        t = threading.Thread(target=self.keep_training)
+        t.start()
+
+    def _new_ready_client(self):
+        """ The client is ready to start new episode,
+        but blocking until training thread call client_ready_event.set()
+    """
+        client_ready_event = threading.Event()
+        self.ready_client_queue.put(client_ready_event)
+        logger.info(
+            "[new_ready_client] approximate size of ready clients:{}".format(
+                self.ready_client_queue.qsize()))
+        client_ready_event.wait()
+
+    def Send(self, request, context):
+        """ Implement Send function in SimulatorServicer
+        Everytime a request comming, will create a new thread to handle
+    """
+        ident, obs, reward, done, info = request.id, request.observation, request.reward, request.done, request.info
+        client = self.clients[ident]
+        info = json.loads(info)
+
+        if 'first' in info:
+            # Waiting training thread to allow start new episode
+            self._new_ready_client()
+
+        obs = np.array(obs, dtype=np.float32)
+        self._process_msg(ident, obs, reward, done, info)
+
+        if done:
+            # Waiting training thread to allow start new episode
+            self._new_ready_client()
+
+        action = self.pred_batch(obs, client.model_idx)
+        step = len(client.memory) - client.last_target_changed
+
+        # whether to add noise depends on the ensemble_num
+        if args.ensemble_num == 1:
+            current_noise = self.noiselevel * (0.98**(step - 1))
+            noise = np.zeros((ACT_DIM, ), dtype=np.float32)
+            if ident % 3 == 0:
+                if step % 5 == 0:
+                    noise = np.random.randn(ACT_DIM) * current_noise
+            elif ident % 3 == 1:
+                if step % 5 == 0:
+                    noise = np.random.randn(ACT_DIM) * current_noise * 2
+            action += noise
+        action = np.clip(action, -1, 1)
+        client.memory[-1].action = action
+        extra_info = {}
+        return simulator_pb2.Reply(action=action, extra=json.dumps(extra_info))
+
+    def _process_msg(self, ident, obs, reward, done, info):
+        client = self.clients[ident]
+        reward_scale = (1 - GAMMA)
+        info['shaping_reward'] *= reward_scale
+        if len(client.memory) > 0:
+            client.memory[-1].reward = reward
+            info['target_change_times'] = client.target_change_times
+            client.memory[-1].info = info
+            if info['target_changed']:
+                client.target_change_times = min(
+                    client.target_change_times + 1, 3)
+                # re-sample model_idx after target was changed
+                client.model_idx = np.random.randint(args.ensemble_num)
+            if done:
+                assert 'last_obs' in info
+                self._parse_memory(client, ident, info['last_obs'])
+        client.memory.append(
+            TransitionExperience(obs=obs, action=None, reward=None, info=None))
+        if 'target_changed' in info and info['target_changed']:
+            client.update_last_target_changed()
+        return False
+
+    def _parse_memory(self, client, ident, last_obs):
+        mem = client.memory
+        n = len(mem)
+
+        # debug info
+        if ident == 1:
+            for i, exp in enumerate(mem):
+                logger.info(
+                    "[step:{}] obs:{} action:{} reward:{} shaping_reward:{}".
+                    format(i, np.sum(mem[i].obs), np.sum(mem[i].action),
+                           mem[i].reward, mem[i].info['shaping_reward']))
+
+        episode_rpm = []
+        for i in range(n - 1):
+            if not mem[i].info['target_changed']:
+                episode_rpm.append([
+                    mem[i].obs, mem[i].action, mem[i].info['shaping_reward'],
+                    mem[i + 1].obs, False, mem[i].info['target_change_times']
+                ])
+        if not mem[-1].info['target_changed']:
+            episode_rpm.append([
+                mem[-1].obs, mem[-1].action, mem[-1].info['shaping_reward'],
+                last_obs, not mem[-1].info['timeout'],
+                mem[i].info['target_change_times']
+            ])
+
+        indicators_dict = calc_indicators(mem)
+        indicators_dict['free_client_num'] = self.ready_client_queue.qsize()
+        indicators_dict['noiselevel'] = self.noiselevel
+
+        with self.MEMORY_LOCK:
+            self.add_episode_rpm(episode_rpm)
+            self.scalars_manager.record(indicators_dict, self.global_step)
+            self.global_step += 1
+            if self.global_step >= 50:
+                self.noiselevel = self.noiselevel * NOISE_DECAY
+
+        client.reset()
+
+    def learn(self):
+        result_q = queue.Queue()
+        th_list = []
+        for j in range(args.ensemble_num):
+            t = threading.Thread(
+                target=self.train_single_model, args=(j, result_q))
+            th_list.append(t)
+        start_time = time.time()
+        for t in th_list:
+            t.start()
+        for t in th_list:
+            t.join()
+
+        logger.info("[learn] {} heads, time consuming:{}".format(
+            args.ensemble_num,
+            time.time() - start_time))
+        for t in th_list:
+            result = result_q.get()
+            for critic_loss in result:
+                self.scalars_manager.feed_critic_loss(critic_loss)
+
+    def train_single_model(self, model_idx, result_q):
+        logger.info("[train_single_model] model_idx:{}".format(model_idx))
+        critic_loss_list = []
+        lock = self.locks[model_idx]
+        memory = self.rpm
+
+        actor_lr = ACTOR_LR * (1.0 - 0.05 * model_idx)
+        critic_lr = CRITIC_LR * (1.0 + 0.1 * model_idx)
+
+        for T in range(TRAIN_TIMES):
+            [states, actions, rewards, new_states,
+             dones] = memory.sample_batch(BATCH_SIZE)
+            lock.acquire()
+            critic_loss = self.agent.learn(states, actions, rewards,
+                                           new_states, dones, actor_lr,
+                                           critic_lr, model_idx)
+            lock.release()
+            critic_loss_list.append(critic_loss)
+        result_q.put(critic_loss_list)
+
+    def keep_training(self):
+        episode_count = 1000000
+        for T in range(episode_count):
+            if self.rpm.size() > BATCH_SIZE * args.warm_start_batchs:
+                self.learn()
+                logger.info(
+                    "[keep_training/{}] trying to acq a new env".format(T))
+
+            # Keep training and predicting balance
+            # After training, waiting for a ready client, and set the client start new episode
+            ready_client_event = self.ready_client_queue.get()
+            ready_client_event.set()
+
+            if np.mod(T, 100) == 0:
+                logger.info("saving models")
+                self.save(T)
+            if np.mod(T, 10000) == 0:
+                logger.info("saving rpm")
+                self.save_rpm()
+
+    def save_rpm(self):
+        save_path = os.path.join(logger.get_dir(), "rpm.npz")
+        self.rpm.save(save_path)
+
+    def restore_rpm(self, rpm_dir):
+        self.rpm.load(rpm_dir)
+
+    def save(self, T):
+        save_path = os.path.join(logger.get_dir(),
+                                 'model_every_100_episodes/step-{}'.format(T))
+        self.agent.save_params(save_path)
+
+    def restore(self, model_path, restore_from_one_head):
+        logger.info('restore model from {}'.format(model_path))
+        self.agent.load_params(model_path, restore_from_one_head)
+
+    def add_episode_rpm(self, episode_rpm):
+        for x in episode_rpm:
+            self.rpm.append(
+                obs=x[0], act=x[1], reward=x[2], next_obs=x[3], terminal=x[4])
+
+    def pred_batch(self, obs, model_idx=None):
+        assert model_idx is not None
+        batch_obs = np.expand_dims(obs, axis=0)
+        self.locks[model_idx].acquire()
+        action = self.agent.predict(batch_obs, model_idx)
+        self.locks[model_idx].release()
+        action = np.squeeze(action, axis=0)
+        return action
+
+
+class SimulatorHandler(threading.Thread):
+    def __init__(self, simulator_server):
+        threading.Thread.__init__(self)
+        self.server = grpc.server(futures.ThreadPoolExecutor(max_workers=400))
+        simulator_pb2_grpc.add_SimulatorServicer_to_server(
+            simulator_server, self.server)
+        self.server.add_insecure_port('[::]:{}'.format(args.port))
+
+    def run(self):
+        self.server.start()
+        try:
+            while True:
+                time.sleep(_ONE_DAY_IN_SECONDS)
+        except KeyboardInterrupt:
+            self.server.stop(0)
+
+
+if __name__ == '__main__':
+    args = get_server_args()
+
+    if args.logdir is not None:
+        logger.set_dir(args.logdir)
+
+    simulator_server = SimulatorServer()
+
+    if args.restore_rpm_path is not None:
+        simulator_server.restore_rpm(args.restore_rpm_path)
+    if args.restore_model_path is not None:
+        simulator_server.restore(args.restore_model_path,
+                                 args.restore_from_one_head)
+
+    simulator_hanlder = SimulatorHandler(simulator_server=simulator_server)
+    simulator_hanlder.run()

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/test.py

@@ -15,74 +15,82 @@
 import argparse
 import numpy as np
 import time
-from env_wrapper import FrameSkip, ActionScale, PelvisBasedObs, ForwardReward
+from env_wrapper import FrameSkip, ActionScale, PelvisBasedObs, TestReward
+from multi_head_ddpg import MultiHeadDDPG
+from opensim_agent import OpenSimAgent
+from opensim_model import OpenSimModel
 from osim.env import ProstheticsEnv
 from parl.utils import logger
-from submit_model import SubmitModel
+"""
+Test model with ensemble predict
+"""
 
 
-def play_multi_episode(submit_model, episode_num=2, vis=False, seed=0):
+def play_multi_episode(agent, episode_num=2, vis=False, seed=0):
     np.random.seed(seed)
     env = ProstheticsEnv(visualize=vis)
     env.change_model(model='3D', difficulty=1, prosthetic=True, seed=seed)
-    env = ForwardReward(env)
+    env = TestReward(env)
     env = FrameSkip(env, 4)
     env = ActionScale(env)
     env = PelvisBasedObs(env)
+
     all_reward = []
-    all_shaping_reward = 0
-    last_frames_count = 0
 
     for e in range(episode_num):
         t = time.time()
         episode_reward = 0.0
-        episode_shaping_reward = 0.0
-        observation = env.reset(project=False)
-        target_change_times = 0
+        obs = env.reset(project=False)
         step = 0
-        loss = []
         while True:
             step += 1
-            action = submit_model.pred_batch(observation, target_change_times)
-            observation, reward, done, info = env.step(action, project=False)
-            step_frames = info['frame_count'] - last_frames_count
-            last_frames_count = info['frame_count']
+
+            batch_obs = np.expand_dims(obs, axis=0)
+
+            action = agent.ensemble_predict(batch_obs)
+            action = np.squeeze(action, axis=0)
+            obs, reward, done, info = env.step(action, project=False)
             episode_reward += reward
-            # we pacle it here to drop the first step after changing
-            if target_change_times >= 1:
-                loss.append(10 * step_frames - reward)
-            if info['target_changed']:
-                target_change_times = min(target_change_times + 1, 3)
-            logger.info("[step/{}]reward:{}  info:{}".format(
-                step, reward, info))
-            episode_shaping_reward += info['shaping_reward']
+            logger.info("[step/{}]reward:{}".format(step, reward))
             if done:
                 break
         all_reward.append(episode_reward)
-        all_shaping_reward += episode_shaping_reward
         t = time.time() - t
         logger.info(
-            "[episode/{}] time: {} episode_reward:{} change_loss:{} after_change_loss:{} mean_reward:{}"
-            .format(e, t, episode_reward, np.sum(loss[:15]), np.sum(loss[15:]),
-                    np.mean(all_reward)))
+            "[episode/{}] time: {} episode_reward:{} mean_reward:{}".format(
+                e, t, episode_reward, np.mean(all_reward)))
     logger.info("Mean reward:{}".format(np.mean(all_reward)))
 
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
     parser.add_argument(
-        '--use_cuda', action="store_true", help='If set, will run in gpu 0')
+        '--restore_model_path', type=str, help='restore model path for test')
     parser.add_argument(
         '--vis', action="store_true", help='If set, will visualize.')
     parser.add_argument('--seed', type=int, default=0, help='Random seed.')
     parser.add_argument(
         '--episode_num', type=int, default=1, help='Episode number to run.')
+    parser.add_argument('--ensemble_num', type=int, help='ensemble_num')
     args = parser.parse_args()
 
-    submit_model = SubmitModel(use_cuda=args.use_cuda)
+    ACT_DIM = 19
+    VEL_DIM = 4
+    OBS_DIM = 185 + VEL_DIM
+    GAMMA = 0.96
+    TAU = 0.001
+    models = []
+    for i in range(args.ensemble_num):
+        models.append(OpenSimModel(OBS_DIM, VEL_DIM, ACT_DIM, model_id=i))
+    hyperparas = {
+        'gamma': GAMMA,
+        'tau': TAU,
+        'ensemble_num': args.ensemble_num
+    }
+    alg = MultiHeadDDPG(models, hyperparas)
+    agent = OpenSimAgent(alg, OBS_DIM, ACT_DIM, args.ensemble_num)
+
+    agent.load_params(args.restore_model_path)
 
     play_multi_episode(
-        submit_model,
-        episode_num=args.episode_num,
-        vis=args.vis,
-        seed=args.seed)
+        agent, episode_num=args.episode_num, vis=args.vis, seed=args.seed)

examples/NeurIPS2018-AI-for-Prosthetics-Challenge/utils.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import csv
+import numpy as np
+import tensorflow as tf
+import os
+import six
+
+
+class Summary(object):
+    """Logging in tensorboard without tensorflow ops.
+
+    Simple example on how to log scalars and images to tensorboard without tensor ops.
+    License: Copyleft
+
+    __author__ = "Michael Gygli"
+    """
+
+    def __init__(self, logdir):
+        """Creates a summary writer logging to logdir."""
+        self.writer = tf.summary.FileWriter(logdir)
+
+    def log_scalar(self, tag, value, step):
+        """Log a scalar variable.
+        Parameter
+        ----------
+        tag : basestring
+            Name of the scalar
+        value
+        step : int
+            training iteration
+        """
+        summary = tf.Summary(
+            value=[tf.Summary.Value(tag=tag, simple_value=value)])
+        self.writer.add_summary(summary, step)
+        self.writer.flush()
+
+
+class StatCounter(object):
+    """ A simple counter"""
+
+    def __init__(self, max_size=50):
+        self.reset()
+        self.max_size = max_size
+
+    def feed(self, v):
+        """
+        Args:
+            v(float or np.ndarray): has to be the same shape between calls.
+        """
+        self._values.append(v)
+        if len(self._values) > self.max_size:
+            self._values = self._values[-self.max_size:]
+
+    def reset(self):
+        self._values = []
+
+    @property
+    def count(self):
+        return len(self._values)
+
+    @property
+    def mean(self):
+        assert len(self._values)
+        return np.mean(self._values)
+
+    @property
+    def sum(self):
+        assert len(self._values)
+        return np.sum(self._values)
+
+    @property
+    def max(self):
+        assert len(self._values)
+        return max(self._values)
+
+    @property
+    def min(self):
+        assert len(self._values)
+        return min(self._values)
+
+    @property
+    def success_rate(self):
+        count = 0
+        for v in self._values:
+            if v > 35.0:
+                count += 1
+        return float(count) / len(self._values)
+
+
+def calc_indicators(mem):
+    START_STEPS = 15
+    n = len(mem)
+    episode_shaping_reward = np.sum(
+        [exp.info['shaping_reward'] for exp in mem])
+    episode_r2_reward = np.sum([
+        exp.info['r2_reward'] for exp in mem if exp.info['frame_count'] <= 1000
+    ])
+    x_offset_reward = np.sum(exp.info['x_offset_reward'] for exp in mem)
+
+    episode_length = mem[-1].info['frame_count']
+
+    scalar_vel = np.mean([exp.info['scalar_vel'] for exp in mem])
+    action_l2_penalty = np.mean(
+        [exp.info['mean_action_l2_penalty'] for exp in mem])
+
+    start_loss = 10 * START_STEPS * 4 - np.sum([exp.reward
+                                                for exp in mem][:START_STEPS])
+    all_start_loss = 0
+    for i in range(n):
+        if not mem[i].info['target_changed']:
+            frame_count = 4
+            if i - 1 >= 0:
+                frame_count = mem[i].info['frame_count'] - mem[
+                    i - 1].info['frame_count']
+            all_start_loss += 10.0 * frame_count - mem[i].reward
+        else:
+            break
+    start_other_loss = all_start_loss - start_loss
+    first_change_loss = 0
+    second_change_loss = 0
+    third_change_loss = 0
+    first_change_other_loss = 0
+    second_change_other_loss = 0
+    third_change_other_loss = 0
+    first_stage_vel = 0.0
+    second_stage_vel = 0.0
+    third_stage_vel = 0.0
+    other_loss = 0
+
+    change_loss = []
+    all_change_loss = []
+    change_vel = []
+
+    for i in range(n - 1):
+        if mem[i].info['target_changed']:
+            change_loss.append(0.0)
+            all_change_loss.append(0.0)
+            change_vel.append([])
+            for j in range(START_STEPS):
+                idx = i + 1 + j
+                if idx >= n or mem[idx].info['target_changed']:
+                    break
+                frame_count = 4
+                if idx - 1 >= 0:
+                    frame_count = mem[idx].info['frame_count'] - mem[
+                        idx - 1].info['frame_count']
+                change_loss[-1] += 10.0 * frame_count - mem[idx].reward
+            for j in range(n - i - 1):
+                idx = i + 1 + j
+                if idx >= n or mem[idx].info['target_changed']:
+                    break
+                if idx - 1 >= 0:
+                    frame_count = mem[idx].info['frame_count'] - mem[
+                        idx - 1].info['frame_count']
+                all_change_loss[-1] += 10.0 * frame_count - mem[idx].reward
+                change_vel[-1].append(mem[idx].info['scalar_vel'])
+
+    if len(change_loss) >= 1:
+        first_change_loss = change_loss[0]
+        first_change_other_loss = all_change_loss[0] - change_loss[0]
+    if len(change_loss) >= 2:
+        second_change_loss = change_loss[1]
+        second_change_other_loss = all_change_loss[1] - change_loss[1]
+    if len(change_loss) >= 3:
+        third_change_loss = change_loss[2]
+        third_change_other_loss = all_change_loss[2] - change_loss[2]
+    other_loss = 10 * mem[-1].info[
+        'frame_count'] - start_loss - first_change_loss - second_change_loss - third_change_loss - episode_r2_reward
+    if len(change_vel) >= 1:
+        first_stage_vel = np.mean(change_vel[0])
+    if len(change_vel) >= 2:
+        second_stage_vel = np.mean(change_vel[1])
+    if len(change_vel) >= 3:
+        third_stage_vel = np.mean(change_vel[2])
+
+    indicators_dict = {
+        'episode_shaping_reward': episode_shaping_reward,
+        'episode_r2_reward': episode_r2_reward,
+        'x_offset_reward': x_offset_reward,
+        'episode_length': episode_length,
+        'scalar_vel': scalar_vel,
+        'mean_action_l2_penalty': action_l2_penalty,
+        'start_loss': start_loss,
+        'first_change_loss': first_change_loss,
+        'second_change_loss': second_change_loss,
+        'third_change_loss': third_change_loss,
+        'start_other_loss': start_other_loss,
+        'first_change_other_loss': first_change_other_loss,
+        'second_change_other_loss': second_change_other_loss,
+        'third_change_other_loss': third_change_other_loss,
+        'other_loss': other_loss,
+        'first_stage_vel': first_stage_vel,
+        'second_stage_vel': second_stage_vel,
+        'third_stage_vel': third_stage_vel
+    }
+    return indicators_dict
+
+
+class ScalarsManager(object):
+    def __init__(self, logdir):
+        self.summary = Summary(logdir=logdir)
+
+        self.max_shaping_reward = 0
+        self.max_x_offset_reward = 0
+        self.max_r2_reward = 0
+
+        self.critic_loss_counter = StatCounter(max_size=500)
+
+        self.r2_reward_counter = StatCounter(max_size=500)
+        self.nofall_r2_reward_counter = StatCounter(max_size=500)
+        self.falldown_counter100 = StatCounter(max_size=100)
+
+        self.vel_keys = [
+            'scalar_vel', 'first_stage_vel', 'second_stage_vel',
+            'third_stage_vel'
+        ]
+        self.vel_counter = {}
+        for key in self.vel_keys:
+            self.vel_counter[key] = StatCounter(max_size=500)
+        self.reward_loss_keys = [
+            'start_loss', 'first_change_loss', 'second_change_loss',
+            'third_change_loss', 'start_other_loss', 'first_change_other_loss',
+            'second_change_other_loss', 'third_change_other_loss', 'other_loss'
+        ]
+        self.reward_loss_counter = {}
+        for key in self.reward_loss_keys:
+            self.reward_loss_counter[key] = StatCounter(max_size=500)
+
+    def feed_critic_loss(self, critic_loss):
+        self.critic_loss_counter.feed(critic_loss)
+
+    def record(self, record_dict, global_step):
+        self.max_shaping_reward = max(self.max_shaping_reward,
+                                      record_dict['episode_shaping_reward'])
+        self.max_x_offset_reward = max(self.max_x_offset_reward,
+                                       record_dict['x_offset_reward'])
+        self.max_r2_reward = max(self.max_r2_reward,
+                                 record_dict['episode_r2_reward'])
+
+        self.r2_reward_counter.feed(record_dict['episode_r2_reward'])
+        if record_dict['episode_length'] >= 1000:  # no falldown
+            self.nofall_r2_reward_counter.feed(
+                record_dict['episode_r2_reward'])
+            self.falldown_counter100.feed(0.0)
+        else:
+            self.falldown_counter100.feed(1.0)
+
+        for key in self.reward_loss_keys:
+            self.reward_loss_counter[key].feed(record_dict[key])
+        for key in self.vel_keys:
+            self.vel_counter[key].feed(record_dict[key])
+
+        self.summary.log_scalar('performance/falldown_rate',
+                                self.falldown_counter100.sum / 100.0,
+                                global_step)
+        self.summary.log_scalar('performance/max_r2_reward',
+                                self.max_r2_reward, global_step)
+        self.summary.log_scalar('performance/max_shaping_reward',
+                                self.max_shaping_reward, global_step)
+        self.summary.log_scalar('performance/max_x_offset_reward',
+                                self.max_x_offset_reward, global_step)
+        self.summary.log_scalar('performance/episode_r2_reward',
+                                record_dict['episode_r2_reward'], global_step)
+        self.summary.log_scalar('performance/episode_shaping_reward',
+                                record_dict['episode_shaping_reward'],
+                                global_step)
+        self.summary.log_scalar('performance/x_offset_reward',
+                                record_dict['x_offset_reward'], global_step)
+        self.summary.log_scalar('performance/episode_length',
+                                record_dict['episode_length'], global_step)
+        self.summary.log_scalar('performance/mean_action_l2_penalty',
+                                record_dict['mean_action_l2_penalty'],
+                                global_step)
+
+        self.summary.log_scalar('server/free_client_num',
+                                record_dict['free_client_num'], global_step)
+
+        self.summary.log_scalar('model/noiselevel', record_dict['noiselevel'],
+                                global_step)
+        if self.critic_loss_counter.count > 0:
+            mean_critic_loss = self.critic_loss_counter.mean
+            self.summary.log_scalar('model/critic_loss', mean_critic_loss,
+                                    global_step)
+
+        if self.r2_reward_counter.count > 400:
+            mean_r2_reward = self.r2_reward_counter.mean
+            self.summary.log_scalar('performance/recent_r2_reward',
+                                    mean_r2_reward, global_step)
+            mean_nofall_r2_reward = self.nofall_r2_reward_counter.mean
+            self.summary.log_scalar('performance/recent_nofall_r2_reward',
+                                    mean_nofall_r2_reward, global_step)
+
+            for key in self.vel_keys:
+                self.summary.log_scalar('scalar_vel/{}'.format(key),
+                                        self.vel_counter[key].mean,
+                                        global_step)
+
+            for key in self.reward_loss_keys:
+                if 'first' in key:
+                    self.summary.log_scalar('1_stage_loss_reward/' + key,
+                                            self.reward_loss_counter[key].mean,
+                                            global_step)
+                elif 'second' in key:
+                    self.summary.log_scalar('2_stage_loss_reward/' + key,
+                                            self.reward_loss_counter[key].mean,
+                                            global_step)
+                elif 'third' in key:
+                    self.summary.log_scalar('3_stage_loss_reward/' + key,
+                                            self.reward_loss_counter[key].mean,
+                                            global_step)
+                elif 'start' in key:
+                    self.summary.log_scalar('0_stage_loss_reward/' + key,
+                                            self.reward_loss_counter[key].mean,
+                                            global_step)
+                else:
+                    self.summary.log_scalar('loss_reward/' + key,
+                                            self.reward_loss_counter[key].mean,
+                                            global_step)
+            self.summary.log_scalar(
+                '0_stage_loss_reward/stage_loss',
+                self.reward_loss_counter['start_loss'].mean +
+                self.reward_loss_counter['start_other_loss'].mean, global_step)
+            self.summary.log_scalar(
+                '1_stage_loss_reward/stage_loss',
+                self.reward_loss_counter['first_change_loss'].mean +
+                self.reward_loss_counter['first_change_other_loss'].mean,
+                global_step)
+            self.summary.log_scalar(
+                '2_stage_loss_reward/stage_loss',
+                self.reward_loss_counter['second_change_loss'].mean +
+                self.reward_loss_counter['second_change_other_loss'].mean,
+                global_step)
+            self.summary.log_scalar(
+                '3_stage_loss_reward/stage_loss',
+                self.reward_loss_counter['third_change_loss'].mean +
+                self.reward_loss_counter['third_change_other_loss'].mean,
+                global_step)
+
+
+class TransitionExperience(object):
+    """ A transition of state, or experience"""
+
+    def __init__(self, obs, action, reward, info, **kwargs):
+        """ kwargs: whatever other attribute you want to save"""
+        self.obs = obs
+        self.action = action
+        self.reward = reward
+        self.info = info
+        for k, v in six.iteritems(kwargs):
+            setattr(self, k, v)

examples/PPO/train.py

@@ -25,12 +25,11 @@ from utils import *
 def run_train_episode(env, agent, scaler):
     obs = env.reset()
     observes, actions, rewards, unscaled_obs = [], [], [], []
-    done = False
     step = 0.0
     scale, offset = scaler.get()
     scale[-1] = 1.0  # don't scale time step feature
     offset[-1] = 0.0  # don't offset time step feature
-    while not done:
+    while True:
         obs = obs.reshape((1, -1))
         obs = np.append(obs, [[step]], axis=1)  # add time step feature
         unscaled_obs.append(obs)
@@ -50,6 +49,9 @@ def run_train_episode(env, agent, scaler):
         rewards.append(reward)
         step += 1e-3  # increment time step feature
 
+        if done:
+            break
+
     return (np.concatenate(observes), np.concatenate(actions),
             np.array(rewards, dtype='float32'), np.concatenate(unscaled_obs))
 
@@ -75,6 +77,7 @@ def run_evaluate_episode(env, agent, scaler):
         rewards.append(reward)
 
         step += 1e-3  # increment time step feature
+
         if done:
             break
     return np.sum(rewards)
@@ -153,7 +156,6 @@ def main():
 
 
 if __name__ == "__main__":
-
     parser = argparse.ArgumentParser()
     parser.add_argument(
         '--env',
@@ -186,6 +188,5 @@ if __name__ == "__main__":
         default='CLIP')
 
     args = parser.parse_args()
-    import time
-    logger.set_dir('./log_dir/{}_{}'.format(args.loss_type, time.time()))
+
     main()

parl/framework/agent_base.py

@@ -48,8 +48,9 @@ class Agent(object):
         if gpu_id is None:
             gpu_id = 0 if get_gpu_count() > 0 else -1
         self.gpu_id = gpu_id
-        place = fluid.CUDAPlace(gpu_id) if gpu_id >= 0 else fluid.CPUPlace()
-        self.fluid_executor = fluid.Executor(place)
+        self.place = fluid.CUDAPlace(
+            gpu_id) if gpu_id >= 0 else fluid.CPUPlace()
+        self.fluid_executor = fluid.Executor(self.place)
         self.fluid_executor.run(fluid.default_startup_program())
 
     def build_program(self):

parl/framework/model_base.py

@@ -27,13 +27,19 @@ class Network(object):
     A Network is an unordered set of LayerFuncs or Networks.
     """
 
-    def sync_params_to(self, target_net, gpu_id=0, decay=0.0):
+    def sync_params_to(self,
+                       target_net,
+                       gpu_id=0,
+                       decay=0.0,
+                       share_vars_parallel_executor=None):
         """
         Args:
             target_net: Network object deepcopy from source network
             gpu_id: gpu id of target_net 
             decay: Float. The decay to use. 
                    target_net_weights = decay * target_net_weights + (1 - decay) * source_net_weights
+            share_vars_parallel_executor: if not None, will use fluid.ParallelExecutor 
+                                          to run program instead of fluid.Executor
         """
         args_hash_id = hashlib.md5('{}_{}_{}'.format(
             id(target_net), gpu_id, decay).encode('utf-8')).hexdigest()
@@ -59,9 +65,6 @@ class Network(object):
 
             param_pairs = get_parameter_pairs(self, target_net)
 
-            place = fluid.CPUPlace() if gpu_id < 0 \
-                    else fluid.CUDAPlace(gpu_id)
-            self._cached_fluid_executor = fluid.Executor(place)
             self._cached_sync_params_program = fluid.Program()
 
             with fluid.program_guard(self._cached_sync_params_program):
@@ -71,7 +74,34 @@ class Network(object):
                     fluid.layers.assign(
                         decay * target_var + (1 - decay) * src_var, target_var)
 
+            if share_vars_parallel_executor is None:
+                # use fluid.Executor
+                place = fluid.CPUPlace() if gpu_id < 0 \
+                        else fluid.CUDAPlace(gpu_id)
+                self._cached_fluid_executor = fluid.Executor(place)
+            else:
+                # use fluid.ParallelExecutor
+                use_cuda = True if gpu_id >= 0 else False
+
+                # specify strategy to make ParallelExecutor run faster
+                exec_strategy = fluid.ExecutionStrategy()
+                exec_strategy.use_experimental_executor = True
+                exec_strategy.num_threads = 4
+                build_strategy = fluid.BuildStrategy()
+                build_strategy.remove_unnecessary_lock = True
+
+                with fluid.scope_guard(fluid.global_scope().new_scope()):
+                    self._cached_fluid_executor = fluid.ParallelExecutor(
+                        use_cuda=use_cuda,
+                        main_program=self._cached_sync_params_program,
+                        share_vars_from=share_vars_parallel_executor,
+                        exec_strategy=exec_strategy,
+                        build_strategy=build_strategy,
+                    )
+        if share_vars_parallel_executor is None:
             self._cached_fluid_executor.run(self._cached_sync_params_program)
+        else:
+            self._cached_fluid_executor.run(fetch_list=[])
 
     @property
     def parameter_names(self):

parl/utils/__init__.py

@@ -14,3 +14,4 @@
 
 from parl.utils.utils import *
 from parl.utils.gputils import *
+from parl.utils.replay_memory import *

parl/utils/logger.py

@@ -101,7 +101,10 @@ for level in _LOGGING_LEVEL:
 def _set_file(path):
     global _FILE_HANDLER
     if os.path.isfile(path):
+        try:
             os.remove(path)
+        except OSError:
+            pass
     hdl = logging.FileHandler(filename=path, encoding='utf-8', mode='w')
     hdl.setFormatter(_Formatter(datefmt='%m-%d %H:%M:%S'))
 

examples/DDPG/replay_memory.py → parl/utils/replay_memory.py

@@ -13,37 +13,74 @@
 # limitations under the License.
 
 import numpy as np
+from parl.utils import logger
+
+__all__ = ['ReplayMemory']
 
 
 class ReplayMemory(object):
     def __init__(self, max_size, obs_dim, act_dim):
-        self.max_size = max_size
-        self.obs_memory = np.zeros((max_size, obs_dim), dtype='float32')
-        self.act_memory = np.zeros((max_size, act_dim), dtype='float32')
-        self.reward_memory = np.zeros((max_size, ), dtype='float32')
-        self.next_obs_memory = np.zeros((max_size, obs_dim), dtype='float32')
-        self.terminal_memory = np.zeros((max_size, ), dtype='bool')
+        self.max_size = int(max_size)
+        self.obs_dim = obs_dim
+        self.act_dim = act_dim
+
+        self.obs = np.zeros((max_size, obs_dim), dtype='float32')
+        self.action = np.zeros((max_size, act_dim), dtype='float32')
+        self.reward = np.zeros((max_size, ), dtype='float32')
+        self.terminal = np.zeros((max_size, ), dtype='bool')
+        self.next_obs = np.zeros((max_size, obs_dim), dtype='float32')
+
         self._curr_size = 0
         self._curr_pos = 0
 
     def sample_batch(self, batch_size):
-        batch_idx = np.random.choice(self._curr_size, size=batch_size)
-        obs = self.obs_memory[batch_idx, :]
-        act = self.act_memory[batch_idx, :]
-        reward = self.reward_memory[batch_idx]
-        next_obs = self.next_obs_memory[batch_idx, :]
-        terminal = self.terminal_memory[batch_idx]
-        return obs, act, reward, next_obs, terminal
+        # index mapping to avoid sampling saving example
+        batch_idx = np.random.randint(
+            self._curr_size - 300 - 1, size=batch_size)
+        batch_idx = (self._curr_pos + 300 + batch_idx) % self._curr_size
+
+        obs = self.obs[batch_idx]
+        reward = self.reward[batch_idx]
+        action = self.action[batch_idx]
+        next_obs = self.next_obs[batch_idx]
+        terminal = self.terminal[batch_idx]
+        return obs, action, reward, next_obs, terminal
 
     def append(self, obs, act, reward, next_obs, terminal):
         if self._curr_size < self.max_size:
             self._curr_size += 1
-        self.obs_memory[self._curr_pos] = obs
-        self.act_memory[self._curr_pos] = act
-        self.reward_memory[self._curr_pos] = reward
-        self.next_obs_memory[self._curr_pos] = next_obs
-        self.terminal_memory[self._curr_pos] = terminal
+        self.obs[self._curr_pos] = obs
+        self.action[self._curr_pos] = act
+        self.reward[self._curr_pos] = reward
+        self.next_obs[self._curr_pos] = next_obs
+        self.terminal[self._curr_pos] = terminal
         self._curr_pos = (self._curr_pos + 1) % self.max_size
 
     def size(self):
         return self._curr_size
+
+    def save(self, pathname):
+        other = np.array([self._curr_size, self._curr_pos], dtype=np.int32)
+        np.savez(
+            pathname,
+            obs=self.obs,
+            action=self.action,
+            reward=self.reward,
+            terminal=self.terminal,
+            next_obs=self.next_obs,
+            other=other)
+
+    def load(self, pathname):
+        data = np.load(pathname)
+        other = data['other']
+        if int(other[0]) > self.max_size:
+            logger.warn('loading from a bigger size rpm!')
+        self._curr_size = min(int(other[0]), self.max_size)
+        self._curr_pos = min(int(other[1]), self.max_size - 1)
+
+        self.obs[:self._curr_size] = data['obs'][:self._curr_size]
+        self.action[:self._curr_size] = data['action'][:self._curr_size]
+        self.reward[:self._curr_size] = data['reward'][:self._curr_size]
+        self.terminal[:self._curr_size] = data['terminal'][:self._curr_size]
+        self.next_obs[:self._curr_size] = data['next_obs'][:self._curr_size]
+        logger.info("[load rpm]memory loade from {}".format(pathname))