-
Notifications
You must be signed in to change notification settings - Fork 4
/
02_train_a2c.py
176 lines (150 loc) · 7.04 KB
/
02_train_a2c.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
from lib import environ, models
from common import agent, experience, helper
import os
import time
import logging
import argparse
import collections
import gym
import numpy as np
import pandas as pd
import torch
import torch.optim as optim
import torch.nn.utils as nn_utils
from common.writer import SummaryWriter
BATCH_SIZE = 32
BARS_COUNT = 50
REWARD_GROUPS = 10
STATS_GROUPS = 10
GAMMA = 0.9
ENTROPY_BETA = 0.01
REWARD_STEPS = 2
CLIP_GRAD = 0.1
LEARNING_RATE = 0.0001
CHECKPOINT_EVERY_STEP = 1000000
GOOGLE_COLAB_MAX_STEP = 5000000
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-y', '--year', default=2018, type=int, help='year of data to train (default: 2018')
parser.add_argument('--resume', default='', type=str, metavar='PATH', help='path to latest checkpoint (default: none)')
parser.add_argument('--cuda', default=False, action='store_true', help='enable cuda')
parser.add_argument('--colab', default=False, action='store_true', help='enable colab hosted runtime')
args = parser.parse_args()
device = torch.device('cuda' if args.cuda else 'cpu')
try:
from lib import data
train_data = data.load_data(year=args.year)
except ModuleNotFoundError:
# workaround that Ta-lib cannot be installed on Colab
train_data = (pd.read_csv('data/000001_prices_%d.csv' % args.year, index_col=0),
pd.read_csv('data/000001_factors_%d.csv' % args.year, index_col=0))
env = environ.StockEnv(train_data, bars_count=BARS_COUNT, commission=0.0, reset_on_sell=False)
# env = gym.wrappers.TimeLimit(env, max_episode_steps=1000)
net = models.A2CConv1d(env.observation_space.shape, env.action_space.n).to(device)
agent = agent.ProbabilityAgent(lambda x: net(x)[0], apply_softmax=True, device=device)
exp_source = experience.ExperienceSource(env, agent, GAMMA, steps_count=REWARD_STEPS)
optimizer = optim.Adam(net.parameters(), lr=LEARNING_RATE, eps=1e-3)
total_reward = []
total_steps = []
reward_buf = []
steps_buf = []
batch = []
frame_idx = 0
frame_prev = 0
ts = time.time()
stats = collections.defaultdict(list)
best_mean_reward = None
file_name = os.path.splitext(os.path.basename(__file__))[0]
file_name = file_name.split('_')[-1]
save_path = os.path.join('saves', file_name)
os.makedirs(save_path, exist_ok=True)
if args.resume:
print('==> Loading %s' % args.resume)
checkpoint = torch.load(os.path.join(save_path, 'checkpoints', args.resume))
total_reward = checkpoint['total_reward']
total_steps = checkpoint['total_steps']
frame_idx = checkpoint['frame_idx']
best_mean_reward = float(checkpoint['best_mean_reward']),
net.load_state_dict(checkpoint['state_dict']),
optimizer.load_state_dict(checkpoint['optimizer'])
print('==> Loaded %s' % args.resume)
logging.basicConfig(level=logging.INFO, format='%(levelname)s:%(message)s',
handlers=[logging.FileHandler(os.path.join(save_path, 'console.log')),
logging.StreamHandler()])
writer = SummaryWriter(os.path.join('runs', file_name))
for exp in exp_source:
frame_idx += 1
batch.append(exp)
if len(batch) < BATCH_SIZE:
continue
optimizer.zero_grad()
loss_val_v, loss_policy_v, loss_entropy_v = helper.a2c_loss(batch, net, GAMMA**REWARD_STEPS, ENTROPY_BETA, device)
batch.clear()
loss_policy_v.backward(retain_graph=True)
grads = np.concatenate([p.grad.data.cpu().numpy().flatten() for p in net.parameters() if p.grad is not None])
loss_v = loss_entropy_v + loss_val_v
loss_v.backward()
nn_utils.clip_grad_norm_(net.parameters(), CLIP_GRAD)
optimizer.step()
loss_v += loss_policy_v
stats['loss_value'].append(loss_val_v)
stats['loss_policy'].append(loss_policy_v)
stats['loss_entropy'].append(loss_entropy_v)
stats['loss_total'].append(loss_v)
stats['grad_l2'].append(np.sqrt(np.mean(np.square(grads))))
stats['grad_max'].append(np.max(np.abs(grads)))
stats['grad_var'].append(np.var(grads))
for stat in stats:
if len(stat) >= STATS_GROUPS:
writer.add_scalar(stat, torch.mean(torch.stack(stats[stat])).item(), frame_idx)
stats[stat].clear()
ep_reward, ep_steps = exp_source.pop_episode_result()
if ep_reward:
logging.info('%d done, Episode reward: %.4f, Episode step: %d' % (frame_idx, ep_reward, ep_steps))
reward_buf.append(ep_reward)
steps_buf.append(ep_steps)
if len(reward_buf) == REWARD_GROUPS:
reward = np.mean(reward_buf)
steps = np.mean(steps_buf)
reward_buf.clear()
steps_buf.clear()
total_reward.append(reward)
total_steps.append(steps)
speed = (frame_idx - frame_prev) / (time.time() - ts)
frame_prev = frame_idx
ts = time.time()
mean_reward = np.mean(total_reward[-100:])
mean_step = np.mean(total_steps[-100:])
logging.info('%d done, mean reward %.3f, mean step %d, speed %.2f f/s' % (frame_idx, mean_reward, mean_step, speed))
writer.add_scalar('speed', speed, frame_idx)
writer.add_scalar('reward', reward, frame_idx)
writer.add_scalar('reward_100', mean_reward, frame_idx)
writer.add_scalar('steps', steps, frame_idx)
writer.add_scalar('steps_100', mean_step, frame_idx)
if best_mean_reward is None or best_mean_reward < mean_reward:
torch.save(net.state_dict(), os.path.join(save_path, 'best_mean_reward.pth'))
try:
if best_mean_reward is not None:
logging.info('Best mean reward updated %.3f -> %.3f'
% (best_mean_reward, mean_reward))
except Exception as e:
print(e)
pass
finally:
best_mean_reward = mean_reward
if frame_idx % CHECKPOINT_EVERY_STEP == 0:
checkpoint = {'frame_idx': frame_idx,
'state_dict': net.state_dict(),
'optimizer': optimizer.state_dict(),
'total_reward': total_reward,
'total_steps': total_steps,
'best_mean_reward': best_mean_reward,
}
os.makedirs(os.path.join(save_path, 'checkpoints'), exist_ok=True)
torch.save(checkpoint, os.path.join(save_path, 'checkpoints', 'checkpoint-%d.pth' % frame_idx))
print('==> checkpoint saved at frame %d' % frame_idx)
# workaround Colab's time limit
if args.colab:
if frame_idx % GOOGLE_COLAB_MAX_STEP == 0:
break
writer.close()