-
Notifications
You must be signed in to change notification settings - Fork 2
/
run_pretrain.py
executable file
·103 lines (82 loc) · 3.08 KB
/
run_pretrain.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
import argparse
import hydra
from hydra import compose, initialize
from src.dataloaders import *
from src.envs import *
from src.envs.vec_env import VecEnv
from src.models import *
from src.common.logger import WandbTrainerLogger, AgentLogger, VecAgentLogger
from src.common.train_utils import set_global_seeds
from src.trainers import build_trainer
from typing import List
from dotmap import DotMap
import torch
import wandb
import numpy as np
import re
def run(args):
args = DotMap(args)
config_dir = args.config_dir
config_name = args.config_name
overrides = args.overrides
# Hydra Compose
config_path = './configs/' + config_dir
#hydra.core.global_hydra.GlobalHydra.instance().clear()
initialize(version_base=None, config_path=config_path)
cfg = compose(config_name=config_name, overrides=overrides)
# device
device = torch.device(cfg.device)
# dataset
torch.set_num_threads(1) # when dataset on disk
cfg.dataloader.device = cfg.device
train_loader, eval_act_loader, eval_rew_loader = build_dataloader(cfg.dataloader)
# shape config
cfg.env.game = cfg.dataloader.game
env, _ = build_env(cfg.env)
obs_shape = [cfg.dataloader.train.frame] + list(env.observation_space.shape[1:])
action_size = env.action_space.n
# initiaize not pre-defined hyperparameters
param_dict = {'obs_shape': obs_shape,
'action_size': action_size,
't_step': cfg.dataloader.train.t_step,
'batch_size': cfg.dataloader.train.batch_size}
for key, value in param_dict.items():
if key in cfg.model.backbone:
cfg.model.backbone[key] = value
if key in cfg.model.head:
cfg.model.head[key] = value
if key in cfg.model.policy:
cfg.model.policy[key] = value
if key in cfg.trainer:
cfg.trainer[key] = value
# logger
logger= WandbTrainerLogger(cfg)
agent_logger = AgentLogger(average_len=100)
# model
model = build_model(cfg.model)
# load pretrained
p_cfg = cfg.pretrain
p_cfg.env = cfg.dataloader.game
trainer = build_trainer(cfg=cfg.trainer,
train_loader=train_loader,
eval_act_loader=eval_act_loader,
eval_rew_loader=eval_rew_loader,
env=env,
device=device,
logger=logger,
agent_logger=agent_logger,
model=model)
# train
if cfg.debug:
trainer.debug()
else:
trainer.train()
wandb.finish()
return logger
if __name__ == '__main__':
parser = argparse.ArgumentParser(allow_abbrev=False)
parser.add_argument('--config_dir', type=str, default='atari/pretrain')
parser.add_argument('--config_name', type=str, default='simtpr')
parser.add_argument('--overrides', action='append', default=[])
args = parser.parse_args()
run(vars(args))