-
Notifications
You must be signed in to change notification settings - Fork 1
/
collect_demos.py
190 lines (154 loc) · 6.74 KB
/
collect_demos.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
from huge.algo import huge
from numpy import VisibleDeprecationWarning
import doodad as dd
import huge.doodad_utils as dd_utils
import argparse
import wandb
import numpy as np
import yaml
import os
from huge.envs.room_env import PointmassGoalEnv
import skvideo.io
def run(model_name, run_path,
reward_layers="600,600",
fourier=False,
fourier_goal_selector=False,
buffer_size=20000,
maze_type=0,
random_goal=False,
num_blocks=1,
seed=0,
network_layers="128,128",
normalize=False,
task_config="slide_cabinet,microwave",
continuous_action_space=False,
goal_threshold=-1,
env_name='pointmass_empty',
num_demos=0,
max_path_length=100,
goal_selector_buffer_size=50000,
gpu=0,
noise=0,
goal_selector_name='', **extra_params):
import gym
import numpy as np
import rlkit.torch.pytorch_util as ptu
ptu.set_gpu_mode(True, 0)
import rlutil.torch as torch
import rlutil.torch.pytorch_util as ptu
# Envs
from huge import envs
from huge.envs.env_utils import DiscretizedActionEnv
# Algo
from huge.algo import buffer, variants, networks
ptu.set_gpu(gpu)
if not gpu:
print('Not using GPU. Will be slow.')
torch.manual_seed(seed)
np.random.seed(seed)
env = envs.create_env(env_name, task_config, num_blocks, random_goal, maze_type, continuous_action_space, goal_threshold)
env_params = envs.get_env_params(env_name)
env_params['max_trajectory_length']=max_path_length
env_params['network_layers']=network_layers
env_params['reward_layers'] = reward_layers
env_params['buffer_size'] = buffer_size
env_params['use_horizon'] = False
env_params['fourier'] = fourier
env_params['fourier_goal_selector'] = fourier_goal_selector
env_params['normalize']=normalize
env_params['env_name'] = env_name
env_params['goal_selector_buffer_size'] = goal_selector_buffer_size
print(env_params)
env_params['goal_selector_name']=goal_selector_name
env_params['continuous_action_space'] = continuous_action_space
env, policy, goal_selector, replay_buffer, goal_selector_buffer, gcsl_kwargs = variants.get_params(env, env_params)
expert_policy = wandb.restore(f"checkpoint/{model_name}.h5", run_path=run_path)
policy.load_state_dict(torch.load(expert_policy.name, map_location=f"cuda:{gpu}"))
policy = policy.to(f"cuda:{gpu}")
gcsl_kwargs['max_path_length']=max_path_length
os.makedirs(f"demos/{env_name}", exist_ok=True)
collect_demos(env, policy, num_demos, env_name, max_path_length, noise)
def env_distance(env, state, goal):
obs = env.observation(state)
if isinstance(env.wrapped_env, PointmassGoalEnv):
return env.base_env.room.get_shaped_distance(obs, goal)
else:
return env.get_shaped_distance(obs, goal)
def create_video(images, video_filename):
images = np.array(images).astype(np.uint8)
images = images.transpose(0,3,1,2)
wandb.log({"demos_video_trajectories":wandb.Video(images, fps=10)})
def collect_demos(env, policy, num_demos, env_name, max_path_length, noise):
policy.eval()
i = 0
while i < num_demos:
actions = []
states = []
if env_name == "kitchenSeq":
goal = np.array([0.3360342, 0.23376076, -0.26612756, 1., 1., 1., -0.583948, 0.7008386, 0.27896893, 0.3000104, -0.5212525, 0.5438033, 2.5910416])
else:
goal = env.extract_goal(env.sample_goal())
state = env.reset()
video = []
for t in range(max_path_length):
video.append(env.render_image())
observation = env.observation(state)
horizon = np.arange(max_path_length) >= (max_path_length - 1 - t) # Temperature encoding of horizon
action = policy.act_vectorized(observation[None], goal[None], horizon=horizon[None], greedy=False, noise=noise)[0]
if "ravens" in env_name:
action = action + np.random.normal(0, noise)
elif np.random.random() < noise:
action = np.random.randint(env.action_space.n)
actions.append(action)
states.append(state)
state, _, done , info = env.step(action)
if done and not("ravens" in env_name):
break
if "ravens" in env_name:
success = env.compute_success(env.observation(states[-1]), goal)
print("pre success", success)
success = success == 4
else:
success = env.compute_success(env.observation(states[-1]), goal)
if success:
final_dist_commanded = env_distance(env, states[-1], goal)
create_video(video, f"{env_name}_{final_dist_commanded}")
print("Final distance 1", final_dist_commanded)
# put actions states into npy file
actions = np.array(actions)
states = np.array(states)
env.plot_trajectories([states], [goal])
np.save(f"demos/{env_name}/demo_{i}_actions.npy", actions)
np.save(f"demos/{env_name}/demo_{i}_states.npy", states)
i += 1
else:
print("Failed trajectory")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--gpu",type=int, default=0)
parser.add_argument("--seed",type=int, default=0)
parser.add_argument("--env_name", type=str, default='pointmass_empty')
parser.add_argument("--epsilon_greedy_rollout",type=float, default=None)
parser.add_argument("--task_config", type=str, default=None)
parser.add_argument("--num_demos", type=int, default=10)
parser.add_argument("--run_path", type=str, default=None)
parser.add_argument("--max_path_length", type=int, default=None)
parser.add_argument("--model_name", type=str, default='best_model_02_04_2023_09:36:41')
parser.add_argument("--noise", type=float, default=0)
parser.add_argument("--num_blocks", type=int, default=None)
args = parser.parse_args()
with open("config.yaml") as file:
config = yaml.safe_load(file)
params = config["common"]
params.update(config[args.env_name])
for key in args.__dict__:
value = args.__dict__[key]
if value is not None:
params[key] = value
data_folder_name = f"{args.env_name}_"
data_folder_name = data_folder_name+"_use_oracle_"
data_folder_name = data_folder_name + str(args.seed)
params["data_folder"] = data_folder_name
wandb.init(project=args.env_name+"demos", name=f"{args.env_name}_demos", config=params, dir="/data/pulkitag/data/marcel/wandb")
run(**params)
# dd_utils.launch(run, params, mode='local', instance_type='c4.xlarge')