-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathexplain_policy.py
208 lines (176 loc) · 9.07 KB
/
explain_policy.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
import shap
import pybullet as p
import pybullet_data
from mojograsp.simcore.environment import EnvironmentDefault
from demos.rl_demo import rl_env
from demos.rl_demo import manipulation_phase_rl
# import rl_env
from demos.rl_demo.rl_state import StateRL, GoalHolder, RandomGoalHolder
from demos.rl_demo import rl_action
from demos.rl_demo import rl_reward
from demos.rl_demo import rl_gym_wrapper
import pandas as pd
from mojograsp.simcore.record_data import RecordDataJSON, RecordDataPKL, RecordDataRLPKL
from mojograsp.simobjects.two_finger_gripper import TwoFingerGripper
from mojograsp.simobjects.ik_gripper import IKGripper
from mojograsp.simobjects.object_with_velocity import ObjectWithVelocity
from mojograsp.simcore.priority_replay_buffer import ReplayBufferPriority
from mojograsp.simcore.data_combination import data_processor
import pickle as pkl
import json
from stable_baselines3 import A2C, PPO
from stable_baselines3.common.evaluation import evaluate_policy
from stable_baselines3.common.monitor import Monitor
import wandb
import numpy as np
import os
import time
from typing import Callable
def build_state_from_pkl(data, args):
"""
Method takes in data loaded from pickle file object and config
Extracts state information from state_container and returns it as a list based on
current used states contained in self.state_list
"""
angle_keys = ["finger0_segment0_joint","finger0_segment1_joint","finger1_segment0_joint","finger1_segment1_joint"]
list_of_states = []
data = data['timestep_list']
for i in range(args['tsteps']):
# print('doing the thing')
state = []
if args['pv'] > 0:
for j in range(1,args['pv']+1):
for key in args['state_list']:
if key == 'op':
state.extend(data[max(i-j,0)]['state']['obj_2']['pose'][0][0:2])
elif key == 'oo':
state.extend(data[max(i-j,0)]['state']['obj_2']['pose'][1])
elif key == 'ftp':
state.extend(data[max(i-j,0)]['state']['f1_pos'][0:2])
state.extend(data[max(i-j,0)]['state']['f2_pos'][0:2])
elif key == 'fbp':
state.extend(data[max(i-j,0)]['state']['f1_base'][0:2])
state.extend(data[max(i-j,0)]['state']['f2_base'][0:2])
elif key == 'fcp':
state.extend(data[max(i-j,0)]['state']['f1_contact_pos'][0:2])
state.extend(data[max(i-j,0)]['state']['f2_contact_pos'][0:2])
elif key == 'ja':
state.extend([data[max(i-j,0)]['state']['two_finger_gripper']['joint_angles'][item] for item in angle_keys])
elif key == 'fta':
state.extend([data[max(i-j,0)]['state']['f1_ang'],data[max(i-j,0)]['state']['f2_ang']])
elif key == 'eva':
state.extend(data[max(i-j,0)]['state']['two_finger_gripper']['eigenvalues'])
elif key == 'evc':
state.extend(data[max(i-j,0)]['state']['two_finger_gripper']['eigenvectors'])
elif key == 'evv':
evecs = data[max(i-j,0)]['state']['two_finger_gripper']['eigenvectors']
evals = data[max(i-j,0)]['state']['two_finger_gripper']['eigenvalues']
scaled = [evals[0]*evecs[0],evals[0]*evecs[2],evals[1]*evecs[1],evals[1]*evecs[3],
evals[2]*evecs[4],evals[2]*evecs[6],evals[3]*evecs[5],evals[3]*evecs[7]]
state.extend(scaled)
elif key == 'gp':
state.extend(data[max(i-j,0)]['state']['goal_pose']['goal_pose'])
else:
raise Exception('key does not match list of known keys')
for key in args['state_list']:
if key == 'op':
state.extend(data[i]['state']['obj_2']['pose'][0][0:2])
elif key == 'oo':
state.extend(data[i]['state']['obj_2']['pose'][1])
elif key == 'ftp':
state.extend(data[i]['state']['f1_pos'][0:2])
state.extend(data[i]['state']['f2_pos'][0:2])
elif key == 'fbp':
state.extend(data[i]['state']['f1_base'][0:2])
state.extend(data[i]['state']['f2_base'][0:2])
elif key == 'fcp':
state.extend(data[i]['state']['f1_contact_pos'][0:2])
state.extend(data[i]['state']['f2_contact_pos'][0:2])
elif key == 'ja':
state.extend([data[i]['state']['two_finger_gripper']['joint_angles'][item] for item in angle_keys])
elif key == 'fta':
state.extend([data[i]['state']['f1_ang'],data[i]['state']['f2_ang']])
elif key == 'eva':
state.extend(data[i]['state']['two_finger_gripper']['eigenvalues'])
elif key == 'evc':
state.extend(data[i]['state']['two_finger_gripper']['eigenvectors'])
elif key == 'evv':
evecs = data[i]['state']['two_finger_gripper']['eigenvectors']
evals = data[i]['state']['two_finger_gripper']['eigenvalues']
scaled = [evals[0]*evecs[0],evals[0]*evecs[2],evals[1]*evecs[1],evals[1]*evecs[3],
evals[2]*evecs[4],evals[2]*evecs[6],evals[3]*evecs[5],evals[3]*evecs[7]]
state.extend(scaled)
elif key == 'gp':
state.extend(data[i]['state']['goal_pose']['goal_pose'])
else:
raise Exception('key does not match list of known keys')
list_of_states.append(state)
return list_of_states
def explain_policy(filepath):
'''
Uses shapely values to explain which state values are most important. expects a filepath to the folder containing experiment_config.json
with a trained model present and a full set of evaluations in filepath + Eval/
'''
with open(filepath+'experiment_config.json', 'r') as argfile:
args = json.load(argfile)
eval_files = os.listdir(filepath + 'Eval')
state_info = []
print('Eval file length',len(eval_files))
for filename in eval_files:
# print(filename)
with open(filepath + 'Eval/'+filename, 'rb') as file:
data = pkl.load(file)
state_list = build_state_from_pkl(data, args)
state_info.extend(state_list)
physics_client = p.connect(p.DIRECT)
hand_id = p.loadURDF(args['hand_path'], useFixedBase=True,
basePosition=[0.0, 0.0, 0.05], flags=p.URDF_ENABLE_CACHED_GRAPHICS_SHAPES)
obj_id = p.loadURDF(args['object_path'], basePosition=[0.0, 0.10, .05], flags=p.URDF_ENABLE_CACHED_GRAPHICS_SHAPES)
x = [0]
y = [0]
pose_list = [x[0],y[0]]
eval_pose_list = pose_list.copy()
hand = IKGripper(hand_id, path=args['hand_path'])
obj = ObjectWithVelocity(obj_id, path=args['object_path'],name='obj_2')
goal_poses = GoalHolder(pose_list)
eval_goal_poses = GoalHolder(eval_pose_list)
state = StateRL(objects=[hand, obj, goal_poses], prev_len=args['pv'],eval_goals = eval_goal_poses)
if args['freq'] ==240:
action = rl_action.ExpertAction()
else:
action = rl_action.InterpAction(args['freq'])
reward = rl_reward.ExpertReward()
arg_dict = args.copy()
if args['action'] == 'Joint Velocity':
arg_dict['ik_flag'] = False
else:
arg_dict['ik_flag'] = True
# replay buffer
replay_buffer = ReplayBufferPriority(buffer_size=4080000)
# environment and recording
# env = rl_env.ExpertEnv(hand=hand, obj=obj, hand_type=arg_dict['hand'], rand_start=args['rstart'])
env = EnvironmentDefault()
# env = rl_env.ExpertEnv(hand=hand, obj=cylinder)
# Create phase
manipulation = manipulation_phase_rl.ManipulationRL(
hand, obj, x, y, state, action, reward, env, replay_buffer=replay_buffer, args=arg_dict)
# data recording
record_data = RecordDataRLPKL(
data_path=args['save_path'], state=state, action=action, reward=reward, save_all=False, controller=manipulation.controller)
gym_env = rl_gym_wrapper.GymWrapper(env, manipulation, record_data, args)
model = PPO("MlpPolicy", gym_env, tensorboard_log=args['tname'], policy_kwargs={'log_std_init':-2.3}).load(args['save_path']+'best_model')
def F(state):
out = np.zeros([4,len(state)])
for i,v in enumerate(state):
temp, _ = model.predict(v)
out[:,i] = temp
# print(out.shape)
return out[0,:].flatten()
state_info = np.array(state_info) # this should be a 30000x50 or so array
print(state_info.shape)
explainer = shap.KernelExplainer(F, state_info)
# test = explainer(state_info[0])
shap_values = explainer.shap_values(state_info[0:500,:], nsamples=50)
shap.summary_plot(shap_values, state_info[0:500,:])
return explainer, state_info
explainer, state_info = explain_policy('/home/mothra/mojo-grasp/demos/rl_demo/data/wedge_double/wedge_l-r/')