def __init__(self, metrics=None):
super(MetricsHandler, self).__init__()
self.metrics = AlphaConfig()
self._items_ = []
for m in metrics:
if m not in METRICS.keys():
logger.raise_error(
NotImplementedError,
"Handler of metric {} is not in implemented metrics set {}"
.format(m, _METRICS_))
self._items_.append(m)
def test_deepcopy(test_dict):
config = AlphaConfig(test_dict)
copied_config = copy.deepcopy(config)
print("expect: False, got: {}".format(config is copied_config))
copied_config.attr_1 = 10
print(copied_config)
print(config)
def test_build(test_dict):
# config = AlphaConfig(path2json="test.json")
configs = AlphaConfig(test_dict, attr_3="value_3")
print(configs.cvt2dict())
with open("test.yaml", 'r') as fp:
src_dict = yaml.safe_load(fp)
configs.yaml = src_dict
configs.attr_1 = "value_1"
configs.attr_2 = {"attr_2_1": 1, "attr_2_2": [1, 2, 4]}
print(configs)
def update(self, dataset, phase, epoch, metric, value):
epoch = str(epoch)
if dataset not in self.metrics.keys():
self.metrics[dataset] = {}
if phase not in self.metrics[dataset].keys():
self.metrics[dataset][phase] = {}
if epoch not in self.metrics[dataset][phase].keys():
self.metrics[dataset][phase][epoch] = {}
if metric not in self.metrics[dataset][phase][epoch].keys():
self.metrics[dataset][phase][epoch][metric] = AlphaConfig(value=0,
counts=0)
self.metrics[dataset][phase][epoch][metric].counts += 1
counts = self.metrics[dataset][phase][epoch][metric].counts
pre_value = self.metrics[dataset][phase][epoch][metric].value
new_value = value / counts + pre_value * (counts - 1) / counts
self.metrics[dataset][phase][epoch][metric].value = new_value
return new_value
def test_read_only(test_dict):
config = AlphaConfig(test_dict)
config.cvt_state(read_only=True)
print("is read only? {}".format(config.is_read_only()))
try:
config.test_attr_1
except AttributeError:
print("Successfully raise AttributeError #1")
try:
config.test_attr_2 = "test_value"
except AttributeError:
print("Successfully raise AttributeError #2")
print(config)
from alphaconfig import AlphaConfig loss_fn = AlphaConfig() # ======== MSELoss ======== loss_fn.MSELoss.WEIGHT1 = 1.0 loss_fn.MSELoss.WEIGHT2 = 1.0 # ======== MAELoss ======== loss_fn.MAELoss.WEIGHT1 = 1.0 loss_fn.MAELoss.WEIGHT2 = 1.0 # ======== MSESSIMLoss ======== loss_fn.MSESSIMLoss.mse = 1.0 loss_fn.MSESSIMLoss.ssim = 0.5
r"""
Info:
Configurations for datasets.
Author:
Yiqun Chen
"""
from alphaconfig import AlphaConfig
data = AlphaConfig()
# ======== Dataset1 ========
data.Dataset1.dir = "/home/chenyiqun/data/SOTS/outdoor"
data.Dataset1.train = True
data.Dataset1.valid = True
data.Dataset1.test = False
data.Dataset1.infer = ["valid"]
data.Dataset1.train_reso = (256, 256)
data.Dataset1.extra_reso = (128, 128)
data.Dataset1.data_ratio = 1.0
# ======== Dataset2 ========
data.Dataset2.dir = "/home/chenyiqun/data/SOTS/indoor"
data.Dataset2.train = False
data.Dataset2.valid = True
data.Dataset2.test = False
data.Dataset2.infer = ["valid"]
data.Dataset2.train_reso = (256, 256)
data.Dataset2.extra_reso = (128, 128)
data.Dataset2.data_ratio = 1.0
r"""
Info:
Configurations for optimizer
Author:
Yiqun Chen
"""
from alphaconfig import AlphaConfig
from .args import args
optim = AlphaConfig()
# ======== Adam ========
optim.Adam.lr = args.lr
optim.Adam.finetune = 1.0
optim.Adam.weight_decay = 0.0
# ======== AdamW ========
optim.AdamW.lr = args.lr
optim.AdamW.finetune = 1.0
optim.AdamW.weight_decay = 0.1
def test_iter(test_dict):
config = AlphaConfig(test_dict)
for it in config:
print(it)
def test_get_key(test_dict):
config = AlphaConfig(test_dict)
print(config.items())
def test_hasattr(test_dict):
print("Testing hasattr")
config = AlphaConfig(test_dict)
config.cvt_state(read_only=True)
print("expect: False, got: {}".format(hasattr(config, "attr")))
r"""
Info:
Configurations for learning rate scheduler
Author:
Yiqun Chen
"""
from alphaconfig import AlphaConfig
from .args import args
scheduler = AlphaConfig()
# ======== LinearLRScheduler ========
scheduler.LinearLRScheduler.min_lr = 2.5E-6
scheduler.LinearLRScheduler.warmup = 10
# ======== StepLRScheduler ========
scheduler.StepLRScheduler.min_lr = 2.5E-6
scheduler.StepLRScheduler.warmup = 10
scheduler.StepLRScheduler.update_epoch = range(int(args.max_epoch * 0.1),
args.max_epoch,
int(args.max_epoch * 0.1))
scheduler.StepLRScheduler.update_coeff = 0.5
class MetricsHandler:
def __init__(self, metrics=None):
super(MetricsHandler, self).__init__()
self.metrics = AlphaConfig()
self._items_ = []
for m in metrics:
if m not in METRICS.keys():
logger.raise_error(
NotImplementedError,
"Handler of metric {} is not in implemented metrics set {}"
.format(m, _METRICS_))
self._items_.append(m)
def register(self, metric):
if m not in METRICS.keys():
logger.raise_error(
NotImplementedError,
"Handler of metric {} is not in implemented metrics set {}".
format(m, _METRICS_))
self._items_.append(m)
def update(self, dataset, phase, epoch, metric, value):
epoch = str(epoch)
if dataset not in self.metrics.keys():
self.metrics[dataset] = {}
if phase not in self.metrics[dataset].keys():
self.metrics[dataset][phase] = {}
if epoch not in self.metrics[dataset][phase].keys():
self.metrics[dataset][phase][epoch] = {}
if metric not in self.metrics[dataset][phase][epoch].keys():
self.metrics[dataset][phase][epoch][metric] = AlphaConfig(value=0,
counts=0)
self.metrics[dataset][phase][epoch][metric].counts += 1
counts = self.metrics[dataset][phase][epoch][metric].counts
pre_value = self.metrics[dataset][phase][epoch][metric].value
new_value = value / counts + pre_value * (counts - 1) / counts
self.metrics[dataset][phase][epoch][metric].value = new_value
return new_value
def calc_metrics(self, dataset, phase, epoch, out, trg, data_range, *args,
**kwargs):
kwargs["device"] = out.device
for metric in self._items_:
value = METRICS[metric](out,
trg,
data_range=data_range,
*args,
**kwargs)
self.update(dataset, phase, epoch, metric, value)
return self.metrics[dataset][phase][epoch]
def summarize(self, dataset, phase, epoch, logger=None, *args, **kwargs):
fmt = "{:<20}" * (len(self._items_) + 1)
values = [dataset]
for idx in range(len(self._items_)):
values.append(
str(
round(
self.metrics[dataset][phase][epoch][
self._items_[idx]].value.item(), 5)))
info = fmt.format("dataset", *self._items_)
data = fmt.format(*values)
msg = "\n========Metrics========\n" + info + "\n" + data
utils.notify(msg, logger=logger)
Author:
Yiqun Chen
Docs:
Configurations, should not call other custom modules.
"""
import os, sys, copy
from alphaconfig import AlphaConfig
from .args import args
from .optim import optim
from .dataset import data
from .loss_fn import loss_fn
from .scheduler import scheduler
configs = AlphaConfig()
cfg = configs
# ================================
# GENERAL
# ================================
cfg.gnrl.root = os.path.join(os.getcwd(), ".")
cfg.gnrl.id = "{}".format(args.id)
cfg.gnrl.batch = args.batch_size
cfg.gnrl.resume = True if args.resume == "true" else False
cfg.gnrl.cuda = eval(args.cuda)
cfg.gnrl.ckphs = range(int(args.max_epoch * 0.6), args.max_epoch,
int(args.max_epoch * 0.1))
cfg.gnrl.infer = "0"
# ================================