forked from rahulkidambi/MobILE-NeurIPS2021
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrun.py
192 lines (167 loc) · 10.5 KB
/
run.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
import os
import gym
import time
import pdb
import numpy as np
import torch
from torch.utils.tensorboard import SummaryWriter
from mjrl.algos.npg_cg import NPG
from mjrl.algos.ppo_clip import PPO
from mjrl.algos.behavior_cloning import BC
from mjrl.policies.gaussian_mlp import MLP
from mjrl.baselines.mlp_baseline import MLPBaseline
from mjrl.utils.gym_env import GymEnv
from mjrl.utils.train_agent import train_agent
from mjrl.samplers.core import sample_paths, sample_data_batch
import mbil.gym_env
from mbil.utils import *
from mbil.gym_env import model_based_env
from mbil.dataset import OnlineDataset
from mbil.cost import RBFLinearCost
from mbil.dynamics_model import DynamicsEnsemble
def main():
args = get_args()
dirs, ids, ensemble_checkpoint, logger, writer, device = setup(args, ask_prompt=True)
# ======== Dataset Setup ==========
expert_db_path = os.path.join(args.data_path, 'expert_data', args.expert_db)
expert_state, expert_action, expert_next_state = get_db_mjrl(expert_db_path, args.num_trajs) # Expert DB
max_expert_norm = torch.max(torch.norm(expert_state, p=2, dim=1))
# Buffer storing online interactions to iteratively train dynamics models
online_dataset = OnlineDataset(args.env, args.buffer_size, device=device)
# ========= Create Model Ensemble =========
optim_args = {'optim': args.dynamics_optim, 'lr': args.dynamics_lr, 'momentum': args.dynamics_momentum, 'eps':1e-8}
model_ensemble = DynamicsEnsemble(args.env, args.n_models, online_dataset, hidden_sizes=args.dynamics_model_hidden, \
optim_args = optim_args, base_seed=args.seed, num_cpus=args.num_cpu)
# ======== ENV SETUP ========
logger.info(">>>>> Creating Environments")
inf_env = GymEnv(gym.make(args.env))
mb_env = GymEnv(model_based_env(gym.make(args.env), model_ensemble, init_state_buffer=expert_state.numpy(),\
norm_thresh = args.norm_thresh_coeff*max_expert_norm, device=device))
# ====== Cost Setup =======
# NOTE: (State) and (State, Next State) specific
cost_input = expert_state
input_type = 's'
if args.use_next_state:
cost_input = torch.cat([expert_state, expert_next_state], dim=1)
input_type = 'ss'
cost_function = RBFLinearCost(cost_input, update_type=args.update_type, feature_dim=args.feature_dim, \
input_type=input_type, bw_quantile=args.bw_quantile, lambda_b=args.lambda_b, lr=args.cost_lr, seed=args.seed)
# ============= INIT AGENT =============
policy = MLP(inf_env.spec, hidden_sizes=tuple(args.actor_model_hidden), seed=args.seed,
init_log_std=args.policy_init_log, min_log_std=args.policy_min_log)
baseline = MLPBaseline(inf_env.spec, reg_coef=args.vf_reg_coef, batch_size=args.vf_batch_size, \
hidden_sizes=tuple(args.critic_model_hidden), epochs=args.vf_iters, learn_rate=args.vf_lr)
# ============== Policy Gradient Init =============
if args.planner == 'trpo':
cg_args = {'iters': args.cg_iter, 'damping': args.cg_damping}
planner_agent = NPG(mb_env, policy, baseline, normalized_step_size=args.kl_dist, \
hvp_sample_frac=args.hvp_sample_frac, seed=args.seed, FIM_invert_args=cg_args, save_logs=True)
elif args.planner == 'ppo':
planner_agent = PPO(mb_env, policy, baseline, clip_coef=args.clip_coef, epochs=args.ppo_epochs, \
mb_size=args.ppo_batch_size, learn_rate=args.ppo_lr, save_logs=True)
else:
raise NotImplementedError('Chosen Planner not yet supported')
# ==============================================
# ============== MAIN LOOP START ===============
# ==============================================
n_iter = 0
best_policy_score = -float('inf')
greedy_scores, sample_scores, greedy_mmds, sample_mmds = [], [], [], []
while n_iter<args.n_iter:
logger.info(f"{'='*10} Main Episode {n_iter+1} {'='*10}")
# ============= Evaluate, Save, Plot ===============
scores, mmds = evaluate(n_iter, logger, writer, args, inf_env, \
planner_agent.policy, cost_function, num_traj=20)
save_and_plot(n_iter, args, dirs, scores, mmds)
if scores['greedy'] > best_policy_score:
best_policy_score = scores['greedy']
save_checkpoint(dirs, planner_agent, cost_function, 'best', agent_type=args.planner)
if (n_iter+1) % args.save_iter == 0:
save_checkpoint(dirs, planner_agent, cost_function, n_iter+1, agent_type=args.planner)
# =============== FIT DYNAMICS ===============
logger.info('====== Updating Dynamics Model =======')
if n_iter == 0:
online_samples = sample_data_batch(args.n_pretrain_samples, inf_env, planner_agent.policy, \
base_seed=args.seed, num_cpu=args.num_cpu, eval_mode=args.greedy_pretrain)
else:
online_samples = sample_data_batch(args.n_dynamics_samples, inf_env, planner_agent.policy, \
base_seed=args.seed, num_cpu=args.num_cpu, eval_mode=args.greedy_pretrain)
states = torch.from_numpy(np.concatenate([traj['observations'] for traj in online_samples], axis=0)).float()
actions = torch.from_numpy(np.concatenate([traj['actions'] for traj in online_samples], axis=0)).float()
next_states = torch.from_numpy(np.concatenate([traj['next_observations'] for traj in online_samples], axis=0)).float()
model_ensemble.add_data(states, actions, next_states)
training_epochs = args.n_pretrain_epochs if n_iter == 0 else args.n_epochs
model_train_info = model_ensemble.train(training_epochs, logger=logger, grad_clip=args.grad_clip)
for n, info in enumerate(model_train_info):
writer.add_scalars(f'data/dynamics_model_{n+1}', {'start_loss': info[1],
'min_loss': info[0]}, n_iter)
for n_grad, grad in enumerate(info[2]):
writer.add_scalar(f'data/grad_norms_{n+1}', grad, n_iter*args.n_epochs + n_grad)
model_ensemble.compute_threshold()
logger.info(f">>>>> Computed Maximum Discrepancy for Ensemble: {model_ensemble.threshold}")
mb_env.env.update_model(model_ensemble)
# BW update
#if args.update_bw:
# old_bw = cost_function.bw
# cost_function.update_bandwidth(model_ensemble.dataset)
# logger.info(f">>>>> Updated Bandwidth from {old_bw} to {cost_function.bw}")
# =============== START MINMAX ==============
for j in range(args.n_minmax):
logger.info(f'====== TRPO Step {j} =======')
best_baseline_optim, best_baseline = None, None
best_policy, best_prev_policy = None, None
curr_max_reward, curr_min_vloss = -float('inf'), float('inf')
for i in range(args.pg_iter):
reward_kwargs = dict(reward_func=cost_function, ensemble=model_ensemble, device=device, \
i=i, use_ground_truth=args.use_ground_truth, update_bw=args.update_bw, \
use_next_state=args.use_next_state, logger=logger)
planner_args = dict(N=args.samples_per_step, env=mb_env, sample_mode='model_based', \
gamma=args.gamma, gae_lambda=args.gae_lambda, num_cpu=args.num_cpu, \
reward_kwargs=reward_kwargs)
prev_policy = planner_agent.policy.get_param_values()
r_mean, r_std, r_min, r_max, _, infos = planner_agent.train_step(**planner_args)
# Policy Heuristic
if r_mean > curr_max_reward:
curr_max_reward = r_mean
best_policy = planner_agent.policy.get_param_values()
best_prev_policy = prev_policy
#best_baseline = planner_agent.baseline.model.state_dict()
#best_baseline_optim = planner_agent.baseline.optimizer.state_dict()
# Baseline Heuristic
if infos['vf_loss_end'] < curr_min_vloss:
curr_min_vloss = infos['vf_loss_end']
best_baseline = planner_agent.baseline.model.state_dict()
best_baseline_optim = planner_agent.baseline.optimizer.state_dict()
# Stderr Logging
reward_mean = np.array(infos['reward']).mean()
int_mean = np.array(infos['int']).mean()
ext_mean = np.array(infos['ext']).mean()
len_mean = np.array(infos['ep_len']).mean()
ground_truth_mean = np.array(infos['ground_truth_reward']).mean()
if infos['mb_mmd'] is not None:
logger.info(f'Model MMD: {infos["mb_mmd"]}')
logger.info(f'Bonus MMD: {infos["bonus_mmd"]}')
logger.info(f'Model Ground Truth Reward: {ground_truth_mean}')
logger.info('PG Iteration {} reward | int | ext | ep_len ---- {:.2f} | {:.2f} | {:.2f} | {:.2f}' \
.format(i+1, reward_mean, int_mean, ext_mean, len_mean))
# Tensorboard Logging
step_count = n_iter*args.n_minmax + j*args.pg_iter + i
if len(infos['int']) != 0:
writer.add_scalar('data/reward_mean', reward_mean, step_count)
writer.add_scalar('data/ext_reward_mean', ext_mean, step_count)
writer.add_scalar('data/int_reward_mean', int_mean, step_count)
writer.add_scalar('data/ep_len_mean', len_mean, step_count)
writer.add_scalar('data/true_reward_mean', ground_truth_mean, step_count)
writer.add_scalar('data/value_loss', infos['vf_loss_end'], step_count)
if infos['mb_mmd'] is not None:
writer.add_scalar('data/mb_mmd', infos['mb_mmd'], n_iter*args.n_minmax + j)
if infos['bonus_mmd'] is not None:
writer.add_scalar('data/bonus_mmd', infos['bonus_mmd'], step_count)
# repopulate planner_agent.policy and baseline, optimizer model weights
planner_agent.policy.set_param_values(best_prev_policy, set_new=False, set_old=True) # Set old
planner_agent.policy.set_param_values(best_policy, set_new=True, set_old=False) # Set new
planner_agent.baseline.model.load_state_dict(best_baseline)
planner_agent.baseline.optimizer.load_state_dict(best_baseline_optim)
n_iter += 1
if __name__ == '__main__':
main()