-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
173 lines (128 loc) · 6.18 KB
/
main.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
import argparse
import os
import shutil
import pickle as pickle
import ray
import numpy as np
import random
from tqdm import tqdm
from tensorboardX import SummaryWriter
from worker import Worker
from environments import Env_config
ENV_CONFIG = Env_config(
name='rough',
ground_roughness=0,
pit_gap=[1,2],
stump_width=None,
stump_height=None,
stump_float=None,
stair_height=None,
stair_width=None,
stair_steps=None,
)
def initialize_model_simple(args):
np.random.seed(10)
h1_size = 100
if args.saved_model is not None:
model = pickle.load(open(args.saved_model, 'rb'))
else:
model = {}
model['W0'] = np.random.randn(24, h1_size) / np.sqrt(24)
model['W1'] = np.random.randn(h1_size, 4) / np.sqrt(h1_size)
model['morph'] = args.initial_scalar*(1.0 + (np.random.rand(8)*2-1.0)*0.5)
return model
def get_parallelized_reward_array(workers, perturbed_models, args):
R = np.zeros(len(perturbed_models))
num_workers = len(workers)
# Fully saturate al the workers to start
assert(len(workers) <= len(perturbed_models))
batches_to_assign_to_workers = group_jobs_for_workers(jobs=perturbed_models, num_workers=num_workers)
ongoing_ids = [workers[j].evaluate_model.remote(batches_to_assign_to_workers[j], num_rollouts=3) for j in range(len(workers))]
returns = ray.get(ongoing_ids)
for r in returns:
for model_data in r:
model_idx, score = model_data
R[model_idx] = score
return R
def group_jobs_for_workers(jobs, num_workers):
import math as m
batch_size = m.floor(len(jobs)/num_workers)
tail = len(jobs) - batch_size*num_workers
batches = []
idx = 0
for j in range(tail):
batch = []
for i in range(batch_size+1):
batch.append([idx, jobs.pop(0)])
idx += 1
batches.append(batch)
while len(jobs) > 0:
batch = []
for i in range(batch_size):
batch.append([idx, jobs.pop(0)])
idx += 1
batches.append(batch)
return batches
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='Train BipedWalker')
parser.add_argument('--log_dir', type=str, help='log directory')
parser.add_argument('--num_cores', type=int, help='num cpu cores')
parser.add_argument('--npop', type=int, help='num cpu cores')
parser.add_argument('--num_workers', type=int, help='number of data-collecting workers')
parser.add_argument('--saved_model', type=str, default=None, help='saved model path')
parser.add_argument('--sigma', type=float, default=0.1)
parser.add_argument('--alpha', type=float, default=0.03)
parser.add_argument('--scale_limit_lower', type=float, default=1)
parser.add_argument('--scale_limit_upper', type=float, default=1)
parser.add_argument('--initial_scalar', type=float, default=1)
parser.add_argument('--debug', type=bool, default=True)
parser.add_argument('--save_interval', type=int, default=10)
args = parser.parse_args()
# Logging init
log_dir = args.log_dir
if os.path.exists(os.path.join(log_dir, 'tmp-ray-logs')): shutil.rmtree(os.path.join(log_dir, 'tmp-ray-logs'))
if not os.path.exists(log_dir): os.makedirs(log_dir)
if not os.path.exists(os.path.join(log_dir, 'tmp-ray-logs')): os.makedirs(os.path.join(log_dir, 'tmp-ray-logs'))
if not os.path.exists(os.path.join(log_dir, 'models')): os.makedirs(os.path.join(log_dir, 'models'))
# Start
ray.init(object_store_memory=int(10e8), temp_dir=os.path.join(log_dir, f'tmp-ray-logs'), configure_logging=False, num_cpus=args.num_cores)
workers = [Worker.remote(ENV_CONFIG) for i in range(args.num_workers)]
global_evaluator_worker = Worker.remote(ENV_CONFIG)
model = initialize_model_simple(args)
npop = args.npop
aver_reward = None
writer = SummaryWriter(log_dir=os.path.join(args.log_dir,f'writer'))
for i in tqdm(range(10000)):
# npop different perturbations to each weight matrix W1,W2,W2
N = {}
for k, v in model.items():
if k == 'morph': continue
N[k] = np.random.randn(npop, v.shape[0], v.shape[1])
N['morph'] = np.vstack([random.uniform(args.scale_limit_lower, args.scale_limit_upper)*(1.0 + (np.random.rand(8)*2-1.0)*0.5) for i in range(npop)])
# npop different scores
R = np.zeros(npop)
# npop different perturbed models
perturbed_models = []
for j in range(npop):
model_try = {}
for k, v in model.items():
model_try[k] = v + args.sigma*N[k][j]
perturbed_models.append(model_try)
# Only for tracking where the overall previous model has reached
cur_reward_worker_id = global_evaluator_worker.evaluate_model.remote([[-1, model]], num_rollouts=3, debug=args.debug)
# Launch num_workers workers and keep pushing the npop different perturbed models to them
# until they are all done
R = get_parallelized_reward_array(workers, perturbed_models, args)
# More confident that it is done here
cur_reward = ray.get(cur_reward_worker_id)[0][1]
aver_reward = aver_reward * 0.9 + cur_reward * 0.1 if aver_reward is not None else cur_reward
print(f'iter {i}, cur_reward {cur_reward}, aver_reward {aver_reward} morphology {model["morph"]}')
writer.add_scalar('cur_reward', cur_reward, i)
writer.add_scalar('aver_reward', aver_reward, i)
if i % args.save_interval == 0: pickle.dump(model, open(os.path.join(args.log_dir, f'''models/model-pedal-{cur_reward}-{model['morph'][0]}-{model['morph'][1]}-{model['morph'][2]}-{model['morph'][3]}-{model['morph'][4]}-{model['morph'][5]}-{model['morph'][6]}-{model['morph'][7]}.p'''), 'wb'))
# New model is a weighted combination (based on resulting rewards) of perturbed models + old model
A = (R - np.mean(R)) / np.std(R)
for k in model:
if k == 'morph': continue
model[k] = model[k] + args.alpha/(npop*args.sigma) * np.dot(N[k].transpose(1, 2, 0), A)
model['morph'] = model['morph'] + args.alpha /(npop*args.sigma) * np.dot(N['morph'].transpose(1, 0), A)