File size: 7,536 Bytes
0c51b93 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 |
from torch.utils.tensorboard import SummaryWriter
from collections import defaultdict
import json
import os
import csv
import shutil
import torch
import numpy as np
from termcolor import colored
COMMON_TRAIN_FORMAT = [
('episode', 'E', 'int'),
('step', 'S', 'int'),
('episode_reward', 'R', 'float'),
('true_episode_reward', 'TR', 'float'),
('total_feedback', 'TF', 'int'),
('labeled_feedback', 'LR', 'int'),
('noisy_feedback', 'NR', 'int'),
('duration', 'D', 'time'),
('total_duration', 'TD', 'time'),
]
COMMON_EVAL_FORMAT = [
('episode', 'E', 'int'),
('step', 'S', 'int'),
('episode_reward', 'R', 'float'),
('true_episode_reward', 'TR', 'float'),
('true_episode_success', 'TS', 'float'),
]
AGENT_TRAIN_FORMAT = {
'sac': [
('batch_reward', 'BR', 'float'),
('actor_loss', 'ALOSS', 'float'),
('critic_loss', 'CLOSS', 'float'),
('alpha_loss', 'TLOSS', 'float'),
('alpha_value', 'TVAL', 'float'),
('actor_entropy', 'AENT', 'float'),
('bc_loss', 'BCLOSS', 'float'),
],
'ppo': [
('batch_reward', 'BR', 'float'),
],
}
class AverageMeter(object):
def __init__(self):
self._sum = 0
self._count = 0
def update(self, value, n=1):
self._sum += value
self._count += n
def value(self):
return self._sum / max(1, self._count)
class MetersGroup(object):
def __init__(self, file_name, formating):
self._csv_file_name = self._prepare_file(file_name, 'csv')
self._formating = formating
self._meters = defaultdict(AverageMeter)
self._csv_file = open(self._csv_file_name, 'w')
self._csv_writer = None
def _prepare_file(self, prefix, suffix):
file_name = f'{prefix}.{suffix}'
if os.path.exists(file_name):
os.remove(file_name)
return file_name
def log(self, key, value, n=1):
self._meters[key].update(value, n)
def _prime_meters(self):
data = dict()
for key, meter in self._meters.items():
if key.startswith('train'):
key = key[len('train') + 1:]
else:
key = key[len('eval') + 1:]
key = key.replace('/', '_')
data[key] = meter.value()
return data
def _dump_to_csv(self, data):
if self._csv_writer is None:
self._csv_writer = csv.DictWriter(self._csv_file,
fieldnames=sorted(data.keys()),
restval=0.0)
self._csv_writer.writeheader()
self._csv_writer.writerow(data)
self._csv_file.flush()
def _format(self, key, value, ty):
if ty == 'int':
value = int(value)
return f'{key}: {value}'
elif ty == 'float':
return f'{key}: {value:.04f}'
elif ty == 'time':
return f'{key}: {value:04.1f} s'
else:
raise f'invalid format type: {ty}'
def _dump_to_console(self, data, prefix):
prefix = colored(prefix, 'yellow' if prefix == 'train' else 'green')
pieces = [f'| {prefix: <14}']
for key, disp_key, ty in self._formating:
value = data.get(key, 0)
pieces.append(self._format(disp_key, value, ty))
print(' | '.join(pieces))
def dump(self, step, prefix, save=True):
if len(self._meters) == 0:
return
if save:
data = self._prime_meters()
data['step'] = step
self._dump_to_csv(data)
self._dump_to_console(data, prefix)
self._meters.clear()
class Logger(object):
def __init__(self,
log_dir,
save_tb=False,
log_frequency=10000,
agent='sac'):
self._log_dir = log_dir
self._log_frequency = log_frequency
if save_tb:
tb_dir = os.path.join(log_dir, 'tb')
if os.path.exists(tb_dir):
try:
shutil.rmtree(tb_dir)
except:
print("logger.py warning: Unable to remove tb directory")
pass
self._sw = SummaryWriter(tb_dir)
else:
self._sw = None
# each agent has specific output format for training
assert agent in AGENT_TRAIN_FORMAT
train_format = COMMON_TRAIN_FORMAT + AGENT_TRAIN_FORMAT[agent]
self._train_mg = MetersGroup(os.path.join(log_dir, 'train'),
formating=train_format)
self._eval_mg = MetersGroup(os.path.join(log_dir, 'eval'),
formating=COMMON_EVAL_FORMAT)
def _should_log(self, step, log_frequency):
cur_step = step
log_frequency = log_frequency or self._log_frequency
# import pdb; pdb.set_trace()
return step % log_frequency == 0
def _try_sw_log(self, key, value, step):
if self._sw is not None:
self._sw.add_scalar(key, value, step)
def _try_sw_log_video(self, key, frames, step):
if self._sw is not None:
frames = torch.from_numpy(np.array(frames))
frames = frames.unsqueeze(0)
self._sw.add_video(key, frames, step, fps=30)
def _try_sw_log_histogram(self, key, histogram, step):
if self._sw is not None:
self._sw.add_histogram(key, histogram, step)
def log(self, key, value, step, n=1, log_frequency=1):
cur_step = step
# import pdb; pdb.set_trace()
if not self._should_log(step, log_frequency):
return
assert key.startswith('train') or key.startswith('eval')
if type(value) == torch.Tensor:
value = value.item()
self._try_sw_log(key, value / n, step)
mg = self._train_mg if key.startswith('train') else self._eval_mg
mg.log(key, value, n)
def log_param(self, key, param, step, log_frequency=None):
if not self._should_log(step, log_frequency):
return
self.log_histogram(key + '_w', param.weight.data, step)
if hasattr(param.weight, 'grad') and param.weight.grad is not None:
self.log_histogram(key + '_w_g', param.weight.grad.data, step)
if hasattr(param, 'bias') and hasattr(param.bias, 'data'):
self.log_histogram(key + '_b', param.bias.data, step)
if hasattr(param.bias, 'grad') and param.bias.grad is not None:
self.log_histogram(key + '_b_g', param.bias.grad.data, step)
def log_video(self, key, frames, step, log_frequency=None):
if not self._should_log(step, log_frequency):
return
assert key.startswith('train') or key.startswith('eval')
self._try_sw_log_video(key, frames, step)
def log_histogram(self, key, histogram, step, log_frequency=None):
if not self._should_log(step, log_frequency):
return
assert key.startswith('train') or key.startswith('eval')
self._try_sw_log_histogram(key, histogram, step)
def dump(self, step, save=True, ty=None):
if ty is None:
self._train_mg.dump(step, 'train', save)
self._eval_mg.dump(step, 'eval', save)
elif ty == 'eval':
self._eval_mg.dump(step, 'eval', save)
elif ty == 'train':
self._train_mg.dump(step, 'train', save)
else:
raise f'invalid log type: {ty}'
|