-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathuser_solution.py
331 lines (295 loc) · 18.3 KB
/
user_solution.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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
import pathlib
from collections import deque
import gym
import numpy as np
from mani_skill_learn.env import get_env_info
from mani_skill_learn.env.observation_process import process_mani_skill_base
from mani_skill_learn.methods.builder import build_brl
from mani_skill_learn.utils.data import to_np, unsqueeze
from mani_skill_learn.utils.meta import Config
from mani_skill_learn.utils.torch import load_checkpoint
import os
class ObsProcess:
# modified from Saself.agentenRLWrapper
def __init__(self, env, obs_mode, stack_frame=1):
"""
Stack k last frames for point clouds or rgbd
"""
self.env = env
self.obs_mode = obs_mode
self.stack_frame = stack_frame
self.buffered_data = {}
def _update_buffer(self, obs):
for key in obs:
if key not in self.buffered_data:
self.buffered_data[key] = deque([obs[key]] * self.stack_frame, maxlen=self.stack_frame)
else:
self.buffered_data[key].append(obs[key])
def _get_buffer_content(self):
axis = 0 if self.obs_mode == 'pointcloud' else -1
return {key: np.concatenate(self.buffered_data[key], axis=axis) for key in self.buffered_data}
def process_observation(self, observation):
if self.obs_mode == "state":
return observation
observation = process_mani_skill_base(observation, self.env)
visual_data = observation[self.obs_mode]
self._update_buffer(visual_data)
visual_data = self._get_buffer_content()
state = observation['agent']
# Convert dict of array to list of array with sorted key
ret = {}
ret[self.obs_mode] = visual_data
ret['state'] = state
return ret
class BasePolicy(object):
def __init__(self, opts=None):
self.obs_mode = 'pointcloud'
def act(self, observation):
raise NotImplementedError()
def reset(self): # if you use an RNN-based policy, you need to implement this function
pass
class UserPolicy(BasePolicy):
def __init__(self, env_name):
super().__init__()
self.env = gym.make(env_name)
self.obs_mode = 'pointcloud' # remember to set this!
self.env.set_env_mode(obs_mode=self.obs_mode)
self.stack_frame = 1
# cfg_path = str(pathlib.Path('./configs/bc/mani_skill_point_cloud_transformer.py').resolve())
# cfg_path = os.path.join(os.path.dirname(__file__), './configs/cql/mani_skill_point_cloud_transformer.py')
# cfg_path = os.path.join(os.path.dirname(__file__), './configs/bc/mani_skill_point_cloud_transformer.py')
# cfg_path = os.path.join(os.path.dirname(__file__), './configs/bc/mani_skill_point_cloud_transformer2.py')
# cfg_path = str(pathlib.Path(cfg_path).resolve())
# cfg = Config.fromfile(cfg_path)
# cfg.env_cfg['env_name'] = env_name
# obs_shape, action_shape, action_space = get_env_info(cfg.env_cfg)
# cfg.agent['obs_shape'] = obs_shape
# cfg.agent['action_shape'] = action_shape
# cfg.agent['action_space'] = action_space
self.agents = []
if env_name.find('Door') >= 0:
matrix_index = [i for i in range(20)]
model_paths = [
'./work_dirs/bc_pointnet_transformer_door10_0/BC/models/model_110000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_1/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_2/BC/models/model_100000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_3/BC/models/model_110000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_4/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_5/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_6/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_7/BC/models/model_100000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_8/BC/models/model_100000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_9/BC/models/model_110000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_10/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_11/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_12/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_13/BC/models/model_130000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_14/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_15/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_16/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_17/BC/models/model_110000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_18/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_door10_19/BC/models/model_120000.ckpt',
]
config7 = False #True
model7_path = './work_dirs/bc_pointnet_transformer_door7/BC/models/model_150000.ckpt'
elif env_name.find('Drawer') >= 0:
matrix_index = [i for i in range(20)]
model_paths = [
'./work_dirs/bc_pointnet_transformer_drawer10_0/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_1/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_2/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_3/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_4/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_5/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_6/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_7/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_8/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_9/BC/models/model_100000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_10/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_11/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_12/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_13/BC/models/model_130000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_14/BC/models/model_110000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_15/BC/models/model_130000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_16/BC/models/model_110000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_17/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_18/BC/models/model_130000.ckpt',
'./work_dirs/bc_pointnet_transformer_drawer10_19/BC/models/model_120000.ckpt',
]
config7 = False #True
model7_path = './work_dirs/bc_pointnet_transformer_drawer7/BC/models/model_120000.ckpt'
elif env_name.find('Bucket') >= 0:
matrix_index = [i for i in range(20)]
model_paths = [
'./work_dirs/bc_pointnet_transformer_bucket10_0/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_1/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_2/BC/models/model_110000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_3/BC/models/model_100000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_4/BC/models/model_110000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_5/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_6/BC/models/model_110000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_7/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_8/BC/models/model_100000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_9/BC/models/model_100000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_10/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_11/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_12/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_13/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_14/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_15/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_16/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_17/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_18/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_bucket10_19/BC/models/model_140000.ckpt',
]
config7 = False #True #False
model7_path = './work_dirs/bc_pointnet_transformer_bucket7/BC/models/model_30000.ckpt'
elif env_name.find('Chair') >= 0:
matrix_index = [i for i in range(20)]
model_paths = [
'./work_dirs/bc_pointnet_transformer_chair10_0/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_1/BC/models/model_110000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_2/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_3/BC/models/model_130000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_4/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_5/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_6/BC/models/model_100000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_7/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_8/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_9/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_10/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_11/BC/models/model_110000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_12/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_13/BC/models/model_150000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_14/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_15/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_16/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_17/BC/models/model_140000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_18/BC/models/model_120000.ckpt',
'./work_dirs/bc_pointnet_transformer_chair10_19/BC/models/model_100000.ckpt',
]
config7 = False #True #False
model7_path = './work_dirs/bc_pointnet_transformer_chair7/BC/models/model_150000.ckpt'
for idx, i in enumerate(matrix_index):
cfg_path = os.path.join(os.path.dirname(__file__), './configs/bc/mani_skill_point_cloud_transformer10.py')
cfg_path = str(pathlib.Path(cfg_path).resolve())
cfg = Config.fromfile(cfg_path)
cfg.env_cfg['env_name'] = env_name
obs_shape, action_shape, action_space = get_env_info(cfg.env_cfg)
cfg.agent['obs_shape'] = obs_shape
cfg.agent['action_shape'] = action_shape
cfg.agent['action_space'] = action_space
cfg.agent['policy_cfg']['nn_cfg']['matrix_index'] = i
cur_agent = build_brl(cfg.agent)
ckpt_path = os.path.join(os.path.dirname(__file__), model_paths[idx])
load_checkpoint(cur_agent,
str(pathlib.Path(ckpt_path).resolve()),
map_location='cpu'
)
cur_agent.to('cuda') # dataparallel not done here
cur_agent.eval()
self.agents.append(cur_agent)
if config7:
cfg_path = os.path.join(os.path.dirname(__file__), './configs/bc/mani_skill_point_cloud_transformer7.py')
cfg_path = str(pathlib.Path(cfg_path).resolve())
cfg = Config.fromfile(cfg_path)
cfg.env_cfg['env_name'] = env_name
obs_shape, action_shape, action_space = get_env_info(cfg.env_cfg)
cfg.agent['obs_shape'] = obs_shape
cfg.agent['action_shape'] = action_shape
cfg.agent['action_space'] = action_space
cfg.agent['policy_cfg']['nn_cfg']['matrix_index'] = -1
cur_agent = build_brl(cfg.agent)
ckpt_path = os.path.join(os.path.dirname(__file__), model7_path)
load_checkpoint(cur_agent,
str(pathlib.Path(ckpt_path).resolve()),
map_location='cpu'
)
cur_agent.to('cuda') # dataparallel not done here
cur_agent.eval()
self.agents.append(cur_agent)
# self.agent = build_brl(cfg.agent)
# if env_name.find('Bucket') >= 0:
# ckpt_path = os.path.join(os.path.dirname(__file__), './work_dirs/cql_transformer_bucket/CQL/models/model_115000.ckpt')
# if env_name.find('Chair') >= 0:
# ckpt_path = os.path.join(os.path.dirname(__file__), './work_dirs/cql_transformer_chair/CQL/models/model_115000.ckpt')
# if env_name.find('Door') >= 0:
# ckpt_path = os.path.join(os.path.dirname(__file__), './work_dirs/base_bc_point_transformer_door/BC/models/model_140000.ckpt')
# ckpt_path = os.path.join(os.path.dirname(__file__), './work_dirs/bc_pointnet_transformer_door3/BC/models/model_5000.ckpt')
# ckpt_path = os.path.join(os.path.dirname(__file__), './work_dirs/bc_pointnet_transformer_door3/BC/models/model_25000.ckpt')
# if env_name.find('Drawer') >= 0:
# ckpt_path = os.path.join(os.path.dirname(__file__), './work_dirs/cql_transformer_drawer/CQL/models/model_90000.ckpt')
# load_checkpoint(self.agent,
# str(pathlib.Path('./example_mani_skill_data/OpenCabinetDrawer_1045_link_0-v0_PN_Transformer.ckpt').resolve()),
# map_location='cpu'
# )
# load_checkpoint(self.agent,
# str(pathlib.Path(ckpt_path).resolve()),
# map_location='cpu'
# )
# self.agent.to('cuda') # dataparallel not done here
# self.agent.eval()
self.lstm_obs = []
self.obsprocess = ObsProcess(self.env, self.obs_mode, self.stack_frame)
# def act(self, observation):
# ##### Replace with your code
# observation = self.obsprocess.process_observation(observation)
# return to_np(self.agent(unsqueeze(observation, axis=0), mode='eval'))[0]
def reset(self):
self.lstm_obs = []
return super().reset()
def act(self, observation):
##### Replace with your code
# observation = self.obsprocess.process_observation(observation)
# return to_np(self.agent(unsqueeze(observation, axis=0), mode='eval'))[0]
obs = self.obsprocess.process_observation(observation)
action_list = []
if len(self.lstm_obs) < self.agents[0].lstm_len:
self.lstm_obs.append(obs)
else:
for i in range(self.agents[0].lstm_len - 1):
self.lstm_obs[i] = self.lstm_obs[i + 1]
self.lstm_obs[-1] = obs
for cur_agent in self.agents:
if cur_agent.lstm_len == 1 or len(self.lstm_obs) < cur_agent.lstm_len:
action = to_np(cur_agent(unsqueeze(obs, axis=0), mode='eval'))[0]
action_list.append(action)
else:
# for k in merge_obs:
# print("k: %s; v: %s" % (k, merge_obs[k]))
state_list = []
xyz_list = []
rgb_list = []
seg_list = []
for i in range(cur_agent.lstm_len):
state_list.append(self.lstm_obs[i]['state'])
xyz_list.append(self.lstm_obs[i]['pointcloud']['xyz'])
rgb_list.append(self.lstm_obs[i]['pointcloud']['rgb'])
seg_list.append(self.lstm_obs[i]['pointcloud']['seg'])
# k = 'state'
# merge_obs[k] = [merge_obs[k], lstm_obs[i+1].get(k)]
# k = 'pointcloud'
# merge_obs[k] = {sub_k: [merge_obs[k][sub_k], lstm_obs[i+1][k][sub_k]] for sub_k in merge_obs[k]}
# merge_obs = {k: [merge_obs[k], lstm_obs[i+1].get(k)] for k in merge_obs}
merge_obs = self.lstm_obs[0]
merge_obs['state'] = np.stack(state_list)
# print("state: %s" % (str(merge_obs['state'].shape)))
# k = 'state'
# # merge_obs = {k: np.stack(merge_obs[k]) for k in merge_obs}
# merge_obs[k] = np.stack(merge_obs[k])
# print("k: %s; v: %s" % (k, merge_obs[k].shape))
k = 'pointcloud'
# merge_obs[k] = {sub_k: np.stack(merge_obs[k][sub_k]) for sub_k in merge_obs[k]}
merge_obs[k]['xyz'] = np.stack(xyz_list)
merge_obs[k]['rgb'] = np.stack(rgb_list)
merge_obs[k]['seg'] = np.stack(seg_list)
# for sub_k in merge_obs[k]:
# print("sub_k: %s; v: %s" % (str(sub_k), str(merge_obs[k][sub_k].shape)))
# action = to_np(self.agent(unsqueeze(merge_obs, axis=0), mode=self.sample_mode))[0]
action = to_np(cur_agent(merge_obs, mode='eval'))[0]
# print("cur action: %s" % (str(action.shape)))
action_list.append(action)
action_list = np.stack(action_list)
action = np.mean(action_list, axis=0)
return action