-
Notifications
You must be signed in to change notification settings - Fork 0
/
prepocess.py
161 lines (139 loc) · 5.95 KB
/
prepocess.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
import argparse
import json
import os
import numpy as np
import imageio
import minerl
from sklearn.cluster import KMeans
import tqdm
import torch.nn.functional as F
import torch
import pickle
from collections import deque
parser = argparse.ArgumentParser()
def launch_params():
parser.add_argument('--ROOT',
help='root',
default = '/home/huijie/EECS545/EECS_545_Final_Project')
parser.add_argument('--DATASET_LOC',
help='location of the dataset',
default = '/home/huijie/EECS545/EECS_545_Final_Project/data/rawdata/MineRLTreechopVectorObf-v0')
parser.add_argument('--env',
help='the environment for minerl',
default = 'MineRLTreechopVectorObf-v0')
##### actionspace
parser.add_argument('--actionNum', type=int,
help='the number of discrete action combination',
default=32)
parser.add_argument('--ACTIONSPACE_TYPE',choices=['manually', 'k_means'],
help='way to define the actionsapce',
default='k_means')
##### prepare dataset
parser.add_argument('--PREPARE_DATASET',
help='if True, would automatically prepare dataset',
default=True)
parser.add_argument('--n', type = int,
help='n -step',
default = 25)
parser.add_argument('--gamma', type = float,
help='gamma',
default = 0.99)
# parser.add_argument('--DATA_TOTAL',
# help='total data from demonstration',
# default=400000)
# parser.add_argument('--DATA_PER_FILE',
# help='total data from demonstration',
# default=5000)
def create_actionspace(args):
actionspace = {}
actionspace_path = os.path.join(args.ROOT, "actionspace")
if args.ACTIONSPACE_TYPE == 'k_means':
actionspaceFile = os.path.join(actionspace_path, args.env + "_" + args.ACTIONSPACE_TYPE + "_" + str(args.actionNum) + ".pickle")
if os.path.exists(actionspaceFile):
with open(actionspaceFile, 'rb') as f:
kmeans = pickle.load(f)
else:
dat = minerl.data.make(args.env)
act_vectors = []
L = 100000
for _, act, _, _,_ in tqdm.tqdm(dat.batch_iter(1, 1, 1, preload_buffer_size=20)):
act_vectors.append(act['vector'])
if len(act_vectors) > L:
break
# Reshape these the action batches
acts = np.concatenate(act_vectors).reshape(-1, 64)
kmeans_acts = acts[:L]
# Use sklearn to cluster the demonstrated actions
kmeans = KMeans(n_clusters=args.actionNum, random_state=0).fit(kmeans_acts)
print(kmeans)
with open(actionspaceFile, "wb") as f:
pickle.dump(kmeans, f)
return kmeans
# def vectorize(act):
# vec1 = np.concatenate([act["camera"][0][0],act["attack"][0],act["back"][0]])
# vec2 = np.concatenate([act["forward"][0],act["left"][0],act["right"][0]])
# vec = np.concatenate([vec1,vec2,act["jump"][0]])
# return vec
# def vectorize_v2(act):
# vec = np.array([ act["camera"][0],
# act["camera"][1],
# act["attack"],
# act["back"],
# act["forward"],
# act["left"],
# act["right"],
# act["jump"],
# act['sprint'],
# act['sneak'] ])
# return vec
def prepare_dataset(args, actionspace):
## matrixlize the actionspace
## act_vec is actionNum X actionDim
root = os.path.join(args.ROOT, "data", "processdata" ,args.env + "_preprocess")
if not os.path.exists(root):
os.mkdir(root)
videoindex = 0
frame_index = 0
video_root = os.path.join(root, "{:03d}".format(videoindex))
if not os.path.exists(video_root):
os.mkdir(video_root)
data = minerl.data.make(args.env)
r_memory = deque()
for current_state, action, reward, _, done in data.batch_iter(
batch_size=1, num_epochs=1, seq_len=1):
s = current_state['pov'][0, 0, :, :, :].astype(np.float32)/255.0
s = np.moveaxis(s, -1, 0)
r = np.array([reward[0, 0]])
action_index = actionspace.predict(action['vector'][0, :])
action_one_hot = F.one_hot(torch.tensor([int(action_index)]), args.actionNum).squeeze()
t = np.array([not done[0, 0]])
## n-step reward
r_memory.append((s.reshape((-1, 64, 64)), action_one_hot, r, t))
if t == False:
## the frame is terminal
r_memory_list = list(r_memory)
for data in r_memory_list:
s, a, r, t = r_memory.popleft()
for i in range(len(r_memory)):
r += r_memory[i][2] * (args.gamma**i)
np.savez(os.path.join(video_root, "{:04d}.npz".format(frame_index - args.n - i + len(r_memory) )), s, a, r, t)
else:
## after accumulate n-step
if len(r_memory) >= args.n:
s, a, r, t = r_memory.popleft()
for i in range(len(r_memory)):
r += r_memory[i][2] * (args.gamma**(i + 1))
np.savez(os.path.join(video_root, "{:04d}.npz".format(frame_index - args.n + 1)), s, a, r, t)
frame_index += 1
if done:
videoindex += 1
frame_index = 0
video_root = os.path.join(root, "{:03d}".format(videoindex))
if not os.path.exists(video_root):
os.mkdir(video_root)
if __name__ == "__main__":
launch_params()
args = parser.parse_args()
actionspace = create_actionspace(args)
if args.PREPARE_DATASET:
prepare_dataset(args, actionspace)