-
Notifications
You must be signed in to change notification settings - Fork 12
/
learner.py
149 lines (128 loc) · 6.67 KB
/
learner.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
import gfootball.env as football_env
import time, pprint, importlib
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.distributions import Categorical
import torch.multiprocessing as mp
from tensorboardX import SummaryWriter
def write_summary(writer, arg_dict, summary_queue, n_game, loss_lst, pi_loss_lst, v_loss_lst, \
entropy_lst, move_entropy_lst, optimization_step, self_play_board, win_evaluation, score_evaluation):
win, score, tot_reward, game_len = [], [], [], []
loop_t, forward_t, wait_t = [], [], []
for i in range(arg_dict["summary_game_window"]):
game_data = summary_queue.get()
a,b,c,d,opp_num,t1,t2,t3 = game_data
if arg_dict["env"] == "11_vs_11_kaggle":
if opp_num in self_play_board:
self_play_board[opp_num].append(a)
else:
self_play_board[opp_num] = [a]
if 'env_evaluation' in arg_dict and opp_num==arg_dict['env_evaluation']:
win_evaluation.append(a)
score_evaluation.append(b)
else:
win.append(a)
score.append(b)
tot_reward.append(c)
game_len.append(d)
loop_t.append(t1)
forward_t.append(t2)
wait_t.append(t3)
writer.add_scalar('game/win_rate', float(np.mean(win)), n_game)
writer.add_scalar('game/score', float(np.mean(score)), n_game)
writer.add_scalar('game/reward', float(np.mean(tot_reward)), n_game)
writer.add_scalar('game/game_len', float(np.mean(game_len)), n_game)
writer.add_scalar('train/step', float(optimization_step), n_game)
writer.add_scalar('time/loop', float(np.mean(loop_t)), n_game)
writer.add_scalar('time/forward', float(np.mean(forward_t)), n_game)
writer.add_scalar('time/wait', float(np.mean(wait_t)), n_game)
writer.add_scalar('train/loss', np.mean(loss_lst), n_game)
writer.add_scalar('train/pi_loss', np.mean(pi_loss_lst), n_game)
writer.add_scalar('train/v_loss', np.mean(v_loss_lst), n_game)
writer.add_scalar('train/entropy', np.mean(entropy_lst), n_game)
writer.add_scalar('train/move_entropy', np.mean(move_entropy_lst), n_game)
mini_window = max(1, int(arg_dict['summary_game_window']/3))
if len(win_evaluation)>=mini_window:
writer.add_scalar('game/win_rate_evaluation', float(np.mean(win_evaluation)), n_game)
writer.add_scalar('game/score_evaluation', float(np.mean(score_evaluation)), n_game)
win_evaluation, score_evaluation = [], []
for opp_num in self_play_board:
if len(self_play_board[opp_num]) >= mini_window:
label = 'self_play/'+opp_num
writer.add_scalar(label, np.mean(self_play_board[opp_num][:mini_window]), n_game)
self_play_board[opp_num] = self_play_board[opp_num][mini_window:]
return win_evaluation, score_evaluation
def save_model(model, arg_dict, optimization_step, last_saved_step):
if optimization_step >= last_saved_step + arg_dict["model_save_interval"]:
model_dict = {
'optimization_step': optimization_step,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': model.optimizer.state_dict(),
}
path = arg_dict["log_dir"]+"/model_"+str(optimization_step)+".tar"
torch.save(model_dict, path)
print("Model saved :", path)
return optimization_step
else:
return last_saved_step
def get_data(queue, arg_dict, model):
data = []
for i in range(arg_dict["buffer_size"]):
mini_batch_np = []
for j in range(arg_dict["batch_size"]):
rollout = queue.get()
mini_batch_np.append(rollout)
mini_batch = model.make_batch(mini_batch_np)
data.append(mini_batch)
return data
def learner(center_model, queue, signal_queue, summary_queue, arg_dict):
print("Learner process started")
imported_model = importlib.import_module("models." + arg_dict["model"])
imported_algo = importlib.import_module("algos." + arg_dict["algorithm"])
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = imported_model.Model(arg_dict, device)
model.load_state_dict(center_model.state_dict())
model.optimizer.load_state_dict(center_model.optimizer.state_dict())
algo = imported_algo.Algo(arg_dict)
for state in model.optimizer.state.values():
for k, v in state.items():
if isinstance(v, torch.Tensor):
state[k] = v.cuda()
model.to(device)
writer = SummaryWriter(logdir=arg_dict["log_dir"])
optimization_step = 0
if "optimization_step" in arg_dict:
optimization_step = arg_dict["optimization_step"]
last_saved_step = optimization_step
n_game = 0
loss_lst, pi_loss_lst, v_loss_lst, entropy_lst, move_entropy_lst = [], [], [], [], []
self_play_board = {}
win_evaluation, score_evaluation = [], []
while True:
if queue.qsize() > arg_dict["batch_size"]*arg_dict["buffer_size"]:
last_saved_step = save_model(model, arg_dict, optimization_step, last_saved_step)
signal_queue.put(1)
data = get_data(queue, arg_dict, model)
loss, pi_loss, v_loss, entropy, move_entropy = algo.train(model, data)
optimization_step += arg_dict["batch_size"]*arg_dict["buffer_size"]*arg_dict["k_epoch"]
print("step :", optimization_step, "loss", loss, "data_q", queue.qsize(), "summary_q", summary_queue.qsize())
loss_lst.append(loss)
pi_loss_lst.append(pi_loss)
v_loss_lst.append(v_loss)
entropy_lst.append(entropy)
move_entropy_lst.append(move_entropy)
center_model.load_state_dict(model.state_dict())
if queue.qsize() > arg_dict["batch_size"]*arg_dict["buffer_size"]:
print("warning. data remaining. queue size : ", queue.qsize())
if summary_queue.qsize() > arg_dict["summary_game_window"]:
win_evaluation, score_evaluation = write_summary(writer, arg_dict, summary_queue, n_game, loss_lst, pi_loss_lst,
v_loss_lst, entropy_lst, move_entropy_lst, optimization_step,
self_play_board, win_evaluation, score_evaluation)
loss_lst, pi_loss_lst, v_loss_lst, entropy_lst, move_entropy_lst = [], [], [], [], []
n_game += arg_dict["summary_game_window"]
_ = signal_queue.get()
else:
time.sleep(0.1)