-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_vae.py
179 lines (148 loc) · 6.35 KB
/
train_vae.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
import argparse
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import gym
from tqdm import tqdm
import os
from vae import VAE
import time
from coolname import generate_slug
import utils
import json
from log import Logger
import d4rl
from utils import get_lr
parser = argparse.ArgumentParser()
parser.add_argument('--seed', type=int, default=0)
# dataset
parser.add_argument('--env', type=str, default='hopper')
parser.add_argument('--dataset', type=str, default='medium') # medium, medium-replay, medium-expert, expert
parser.add_argument('--version', type=str, default='v2')
# model
parser.add_argument('--model', default='VAE', type=str)
parser.add_argument('--hidden_dim', type=int, default=750)
parser.add_argument('--beta', type=float, default=0.5)
# train
parser.add_argument('--num_iters', type=int, default=int(1e5))
parser.add_argument('--batch_size', type=int, default=256)
parser.add_argument('--lr', type=float, default=1e-3)
parser.add_argument('--weight_decay', default=0, type=float)
parser.add_argument('--scheduler', default=False, action='store_true')
parser.add_argument('--gamma', default=0.95, type=float)
parser.add_argument('--no_max_action', default=False, action='store_true')
parser.add_argument('--clip_to_eps', default=False, action='store_true')
parser.add_argument('--eps', default=1e-4, type=float)
parser.add_argument('--latent_dim', default=None, type=int, help="default: action_dim * 2")
parser.add_argument('--no_normalize', default=False, action='store_true', help="do not normalize states")
parser.add_argument('--eval_data', default=0.0, type=float, help="proportion of data used for validation, e.g. 0.05")
# work dir
parser.add_argument('--work_dir', type=str, default='train_vae')
parser.add_argument('--notes', default=None, type=str)
args = parser.parse_args()
# make directory
base_dir = 'runs'
utils.make_dir(base_dir)
base_dir = os.path.join(base_dir, args.work_dir)
utils.make_dir(base_dir)
args.work_dir = os.path.join(base_dir, args.env + '_' + args.dataset)
utils.make_dir(args.work_dir)
ts = time.gmtime()
ts = time.strftime("%m-%d-%H:%M", ts)
exp_name = str(args.env) + '-' + str(args.dataset) + '-' + ts + '-bs' \
+ str(args.batch_size) + '-s' + str(args.seed) + '-b' + str(args.beta) + \
'-h' + str(args.hidden_dim) + '-lr' + str(args.lr) + '-wd' + str(args.weight_decay)
exp_name += '-' + generate_slug(2)
if args.notes is not None:
exp_name = args.notes + '_' + exp_name
args.work_dir = args.work_dir + '/' + exp_name
utils.make_dir(args.work_dir)
args.model_dir = os.path.join(args.work_dir, 'model')
utils.make_dir(args.model_dir)
with open(os.path.join(args.work_dir, 'args.json'), 'w') as f:
json.dump(vars(args), f, sort_keys=True, indent=4)
utils.snapshot_src('.', os.path.join(args.work_dir, 'src'), '.gitignore')
logger = Logger(args.work_dir, use_tb=True)
utils.set_seed_everywhere(args.seed)
device = 'cuda'
# load data
env_name = f"{args.env}-{args.dataset}-{args.version}"
env = gym.make(env_name)
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
max_action = float(env.action_space.high[0])
if args.no_max_action:
max_action = None
print(state_dim, action_dim, max_action)
latent_dim = action_dim * 2
if args.latent_dim is not None:
latent_dim = args.latent_dim
replay_buffer = utils.ReplayBuffer(state_dim, action_dim)
replay_buffer.convert_D4RL(d4rl.qlearning_dataset(env))
if not args.no_normalize:
mean, std = replay_buffer.normalize_states()
else:
print("No normalize")
if args.clip_to_eps:
replay_buffer.clip_to_eps(args.eps)
states = replay_buffer.state
actions = replay_buffer.action
if args.eval_data:
eval_size = int(states.shape[0] * args.eval_data)
eval_idx = np.random.choice(states.shape[0], eval_size, replace=False)
train_idx = np.setdiff1d(np.arange(states.shape[0]), eval_idx)
eval_states = states[eval_idx]
eval_actions = actions[eval_idx]
states = states[train_idx]
actions = actions[train_idx]
else:
eval_states = None
eval_actions = None
# train
if args.model == 'VAE':
vae = VAE(state_dim, action_dim, latent_dim, max_action, hidden_dim=args.hidden_dim).to(device)
else:
raise NotImplementedError
optimizer = torch.optim.Adam(vae.parameters(), lr=args.lr, weight_decay=args.weight_decay)
if args.scheduler:
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer=optimizer, gamma=args.gamma)
total_size = states.shape[0]
batch_size = args.batch_size
for step in tqdm(range(args.num_iters + 1), desc='train'):
idx = np.random.choice(total_size, batch_size)
train_states = torch.from_numpy(states[idx]).to(device)
train_actions = torch.from_numpy(actions[idx]).to(device)
# Variational Auto-Encoder Training
recon, mean, std = vae(train_states, train_actions)
recon_loss = F.mse_loss(recon, train_actions)
KL_loss = -0.5 * (1 + torch.log(std.pow(2)) - mean.pow(2) - std.pow(2)).mean()
vae_loss = recon_loss + args.beta * KL_loss
logger.log('train/recon_loss', recon_loss, step=step)
logger.log('train/KL_loss', KL_loss, step=step)
logger.log('train/vae_loss', vae_loss, step=step)
optimizer.zero_grad()
vae_loss.backward()
optimizer.step()
if step % 5000 == 0:
logger.dump(step)
torch.save(vae.state_dict(), '%s/vae_model_%s_%s_b%s_%s.pt' %
(args.model_dir, args.env, args.dataset, str(args.beta), step))
if eval_states is not None and eval_actions is not None:
vae.eval()
with torch.no_grad():
eval_states_tensor = torch.from_numpy(eval_states).to(device)
eval_actions_tensor = torch.from_numpy(eval_actions).to(device)
# Variational Auto-Encoder Evaluation
recon, mean, std = vae(eval_states_tensor, eval_actions_tensor)
recon_loss = F.mse_loss(recon, eval_actions_tensor)
KL_loss = -0.5 * (1 + torch.log(std.pow(2)) - mean.pow(2) - std.pow(2)).mean()
vae_loss = recon_loss + args.beta * KL_loss
logger.log('eval/recon_loss', recon_loss, step=step)
logger.log('eval/KL_loss', KL_loss, step=step)
logger.log('eval/vae_loss', vae_loss, step=step)
vae.train()
if args.scheduler and (step + 1) % 10000 == 0:
logger.log('train/lr', get_lr(optimizer), step=step)
scheduler.step()
logger._sw.close()