-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_utils.py
186 lines (166 loc) · 6.94 KB
/
train_utils.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
180
181
182
183
184
185
186
import os.path
import pathlib
from collections import defaultdict
import jax.numpy as jnp
import numpy as np
from tqdm import tqdm
from dreamer.utils import evaluate_model, get_mixed_precision_policy
def do_episode(agent, training, environment, config, pbar, render):
episode_summary = defaultdict(list)
steps = 0
done = False
observation = environment.reset()
while not done:
action = agent(observation, training)
next_observation, reward, done, info = environment.step(action)
terminal = done and not info.get('TimeLimit.truncated', False)
if training:
agent.observe(dict(observation=observation,
next_observation=next_observation,
action=action.astype(np.float32),
reward=np.array(reward, np.float32),
terminal=np.array(terminal, np.float32),
info=info))
episode_summary['observation'].append(observation)
episode_summary['next_observation'].append(next_observation)
episode_summary['action'].append(action)
episode_summary['reward'].append(reward)
episode_summary['terminal'].append(terminal)
episode_summary['info'].append(info)
observation = next_observation
if render:
episode_summary['image'].append(
environment.render(mode='rgb_array'))
pbar.update(config.action_repeat)
steps += config.action_repeat
episode_summary['steps'] = [steps]
return steps, episode_summary
def interact(agent, environment, steps, config, training=True,
on_episode_end=None):
pbar = tqdm(total=steps)
steps_count = 0
episodes = []
while steps_count < steps:
episode_steps, episode_summary = do_episode(agent, training,
environment, config,
pbar,
len(episodes) <
config.render_episodes and
not training)
steps_count += episode_steps
episodes.append(episode_summary)
if on_episode_end is not None:
on_episode_end(episode_summary, steps_count)
pbar.close()
return steps, episodes
def make_summary(summaries, prefix):
epoch_summary = {prefix + '/average_return': np.asarray([
sum(episode['reward']) for episode in summaries]).mean(),
prefix + '/average_episode_length': np.asarray([
episode['steps'][0]
for episode in summaries]).mean()}
return epoch_summary
def evaluate(agent, train_env, logger, config, steps):
evaluation_steps, evaluation_episodes_summaries = interact(
agent, train_env, config.evaluation_steps_per_epoch, config,
training=False)
if config.render_episodes:
videos = list(map(lambda episode: episode.get('image'),
evaluation_episodes_summaries[
:config.render_episodes]))
logger.log_video(np.array(videos, copy=False).transpose([0, 1, 4, 2, 3])
, steps, name='videos/overview')
if config.evaluate_model:
more_vidoes = evaluate_model(
jnp.asarray(evaluation_episodes_summaries[0]['observation']),
jnp.asarray(evaluation_episodes_summaries[0]['action']),
next(agent.rng_seq),
agent.model, agent.model.params,
get_mixed_precision_policy(config.precision)
)
for vid, name in zip(more_vidoes, ('gt', 'infered', 'generated')):
logger.log_video(
np.array(vid, copy=False).transpose([0, 1, 4, 2, 3]), steps,
name='videos/' + name)
return make_summary(evaluation_episodes_summaries, 'evaluation')
def on_episode_end(episode_summary, logger, global_step, steps_count):
episode_return = sum(episode_summary['reward'])
steps = global_step + steps_count
print("\nFinished episode with return: {}".format(episode_return))
summary = {'training/episode_return': episode_return}
logger.log_evaluation_summary(summary, steps)
def train(config, agent, environment, logger):
if not config.jit:
from jax.config import config as jax_config
jax_config.update('jax_disable_jit', True)
np.random.seed(config.seed)
steps = 0
if pathlib.Path(config.log_dir, 'agent_data').exists():
agent.load(os.path.join(config.log_dir, 'agent_data'))
steps = agent.training_step
print("Loaded {} steps. Continuing training from {}".format(
steps,
config.log_dir))
while steps < config.steps:
print("Performing a training epoch.")
training_steps, training_episodes_summaries = interact(
agent, environment, config.training_steps_per_epoch, config,
training=True,
on_episode_end=lambda episode_summary, steps_count: on_episode_end(
episode_summary, logger=logger, global_step=steps,
steps_count=steps_count))
steps += training_steps
training_summary = make_summary(training_episodes_summaries, 'training')
if config.evaluation_steps_per_epoch:
print("Evaluating.")
evaluation_summaries = evaluate(agent, environment, logger, config,
steps)
training_summary.update(evaluation_summaries)
logger.log_evaluation_summary(training_summary, steps)
# agent.write(os.path.join(config.log_dir, 'agent_data'))
environment.close()
return agent
def load_config():
import argparse
import ruamel.yaml as yaml
def args_type(default):
def parse_string(x):
if default is None:
return x
if isinstance(default, bool):
return bool(['False', 'True'].index(x))
if isinstance(default, int):
return float(x) if ('e' in x or '.' in x) else int(x)
if isinstance(default, (list, tuple)):
return tuple(args_type(default[0])(y) for y in x.split(','))
return type(default)(x)
def parse_object(x):
if isinstance(default, (list, tuple)):
return tuple(x)
return x
return lambda x: parse_string(x) if isinstance(x, str) else parse_object(x)
parser = argparse.ArgumentParser()
parser.add_argument('--configs', nargs='+', required=True)
args, remaining = parser.parse_known_args()
with open('dreamer/config.yaml') as file:
configs = yaml.safe_load(file)
defaults = {}
for name in args.configs:
defaults.update(configs[name])
updated_remaining = []
for idx in range(0, len(remaining), 2):
stripped = remaining[idx].strip('-')
if '.' in stripped:
params_group, key = stripped.split('.')
orig_value = defaults[params_group][key]
arg_type = args_type(orig_value)
defaults[params_group][key] = arg_type(remaining[idx + 1])
else:
updated_remaining.append(remaining[idx])
updated_remaining.append(remaining[idx + 1])
remaining = updated_remaining
parser = argparse.ArgumentParser()
for key, value in sorted(defaults.items(), key=lambda x: x[0]):
arg_type = args_type(value)
parser.add_argument(f'--{key}', type=arg_type, default=arg_type(value))
return parser.parse_args(remaining)