Spaces:
Sleeping
Sleeping
File size: 8,601 Bytes
2d9a728 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 |
import csv
import datetime
from collections import defaultdict
import numpy as np
import torch
import torchvision
import wandb
from termcolor import colored
from torch.utils.tensorboard import SummaryWriter
COMMON_TRAIN_FORMAT = [('frame', 'F', 'int'), ('step', 'S', 'int'),
('episode', 'E', 'int'), ('episode_length', 'L', 'int'),
('episode_reward', 'R', 'float'),
('fps', 'FPS', 'float'), ('total_time', 'T', 'time')]
COMMON_EVAL_FORMAT = [('frame', 'F', 'int'), ('step', 'S', 'int'),
('episode', 'E', 'int'), ('episode_length', 'L', 'int'),
('episode_reward', 'R', 'float'),
('total_time', 'T', 'time')]
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, csv_file_name, formating, use_wandb):
self._csv_file_name = csv_file_name
self._formating = formating
self._meters = defaultdict(AverageMeter)
self._csv_file = None
self._csv_writer = None
self.use_wandb = use_wandb
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 _remove_old_entries(self, data):
rows = []
with self._csv_file_name.open('r') as f:
reader = csv.DictReader(f)
for row in reader:
if 'episode' in row:
# BUGFIX: covers weird cases where CSV are badly written
if row['episode'] == '':
rows.append(row)
continue
if type(row['episode']) == type(None):
continue
if float(row['episode']) >= data['episode']:
break
rows.append(row)
with self._csv_file_name.open('w') as f:
# To handle CSV that have more keys than new data
keys = set(data.keys())
if len(rows) > 0: keys = keys | set(row.keys())
keys = sorted(list(keys))
#
writer = csv.DictWriter(f,
fieldnames=keys,
restval=0.0)
writer.writeheader()
for row in rows:
writer.writerow(row)
def _dump_to_csv(self, data):
if self._csv_writer is None:
should_write_header = True
if self._csv_file_name.exists():
self._remove_old_entries(data)
should_write_header = False
self._csv_file = self._csv_file_name.open('a')
self._csv_writer = csv.DictWriter(self._csv_file,
fieldnames=sorted(data.keys()),
restval=0.0)
if should_write_header:
self._csv_writer.writeheader()
# To handle components that start training later
# (restval covers only when data has less keys than the CSV)
if self._csv_writer.fieldnames != sorted(data.keys()) and \
len(self._csv_writer.fieldnames) < len(data.keys()):
self._csv_file.close()
self._csv_file = self._csv_file_name.open('r')
dict_reader = csv.DictReader(self._csv_file)
rows = [row for row in dict_reader]
self._csv_file.close()
self._csv_file = self._csv_file_name.open('w')
self._csv_writer = csv.DictWriter(self._csv_file,
fieldnames=sorted(data.keys()),
restval=0.0)
self._csv_writer.writeheader()
for row in rows:
self._csv_writer.writerow(row)
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':
value = str(datetime.timedelta(seconds=int(value)))
return f'{key}: {value}'
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_to_wandb(self, data):
wandb.log(data)
def dump(self, step, prefix):
if len(self._meters) == 0:
return
data = self._prime_meters()
data['frame'] = step
if self.use_wandb:
wandb_data = {prefix + '/' + key: val for key, val in data.items()}
self._dump_to_wandb(data=wandb_data)
# self._dump_to_csv(data)
self._dump_to_console(data, prefix)
self._meters.clear()
class Logger(object):
def __init__(self, log_dir, use_tb, use_wandb):
self._log_dir = log_dir
self._train_mg = MetersGroup(log_dir / 'train.csv',
formating=COMMON_TRAIN_FORMAT,
use_wandb=use_wandb)
self._eval_mg = MetersGroup(log_dir / 'eval.csv',
formating=COMMON_EVAL_FORMAT,
use_wandb=use_wandb)
if use_tb:
self._sw = SummaryWriter(str(log_dir / 'tb'))
else:
self._sw = None
self.use_wandb = use_wandb
def _try_sw_log(self, key, value, step):
if self._sw is not None:
self._sw.add_scalar(key, value, step)
def log(self, key, value, step):
assert key.startswith('train') or key.startswith('eval')
if type(value) == torch.Tensor:
value = value.item()
self._try_sw_log(key, value, step)
mg = self._train_mg if key.startswith('train') else self._eval_mg
mg.log(key, value)
def log_metrics(self, metrics, step, ty):
for key, value in metrics.items():
self.log(f'{ty}/{key}', value, step)
def dump(self, step, ty=None):
if ty is None or ty == 'eval':
self._eval_mg.dump(step, 'eval')
if ty is None or ty == 'train':
self._train_mg.dump(step, 'train')
def log_and_dump_ctx(self, step, ty):
return LogAndDumpCtx(self, step, ty)
def log_visual(self, data, step):
if self._sw is not None:
for k, v in data.items():
if len(v.shape) == 3:
self._sw.add_image(k, v)
else:
if len(v.shape) == 4:
v = np.expand_dims(v, axis=0)
self._sw.add_video(k, v, global_step=step, fps=15)
if self.use_wandb:
for k, v in data.items():
if type(v) is not np.ndarray:
v = v.cpu()
if v.dtype not in [np.uint8]:
v = v*255
v = np.uint8(v)
if len(v.shape) == 3:
if v.shape[0] == 3:
v = v.transpose(1,2,0)
# Note: defaulting to save only one image/video to save storage on wandb
wandb.log({k: wandb.Image(v)},)
else:
# Note: defaulting to save only one image/video to save storage on wandb
wandb.log({k: wandb.Video(v, fps=15, format="gif")},)
class LogAndDumpCtx:
def __init__(self, logger, step, ty):
self._logger = logger
self._step = step
self._ty = ty
def __enter__(self):
return self
def __call__(self, key, value):
self._logger.log(f'{self._ty}/{key}', value, self._step)
def __exit__(self, *args):
self._logger.dump(self._step, self._ty)
|