configer.update(["max_batch_size"], old_max_bs - 1)
else:
raise RuntimeError(e)
from segmentor.trainer import Trainer
Log.info("self.configer.get('network','interp') {}".format(
configer.get("network", "interp")))
model_id = configer.get("network", "model_name")
bestTrial = getBestTrial("cityscapes", model_id + "_")
configer.add(["trial_nb"], bestTrial)
bestWeights = "checkpoints/cityscapes/{}__trial{}_max_performance.pth".format(
model_id, bestTrial - 1)
dic = torch.load(bestWeights)
configer.add(["teacher_interp"], 0)
model = Trainer(configer)
model.configer.resume(dic['config_dict'])
model.seg_net.load_state_dict(dic["state_dict"])
Log.info("loading {}".format(bestWeights))
model.configer.add(["teacher_interp"], 0)
model.configer.add(["network", "interp"], False)
model.val(model.data_loader.get_valloader(dataset='val'),
retSimMap=True,
printallIoU=True,
endEval=True,
gradcam=configer.get("test", "gradcam"))
def run(configer, trial):
if configer.get('phase') == 'train' and configer.get("network", "optuna"):
configer.update(["train", "batch_size"],
trial.suggest_int("batch_size",
2 * torch.cuda.device_count(),
configer.get("max_batch_size"),
step=1))
configer.update(["lr", "base_lr"],
trial.suggest_float("base_lr",
0.0001,
0.0019,
step=0.0003))
configer.update(["lr", "lr_policy"],
trial.suggest_categorical("lr_policy",
["step", "lambda_poly"]))
configer.get("lr", "step")["gamma"] = trial.suggest_float("gamma",
0.3,
0.9,
step=0.2)
configer.get("lr",
"step")["step_size"] = trial.suggest_int("step_size",
50,
400,
step=50)
configer.update(["optim", "optim_method"],
trial.suggest_categorical("optim_method",
["sgd", "adam"]))
configer.update(["lr", "nbb_mult"],
trial.suggest_float("nbb_mult", 0.25, 4, step=0.25))
if configer.get("optim", "optim_method") == "sgd":
configer.get("optim", "sgd")["weight_decay"] = trial.suggest_float(
"weight_decay", 0.00001, 0.0002, step=0.00003)
configer.get("optim",
"sgd")["nesterov"] = trial.suggest_categorical(
"nesterov", [False, True])
else:
configer.get("optim",
"adam")["weight_decay"] = trial.suggest_float(
"weight_decay", 0.0001, 0.0010, step=0.0002)
configer.get("network",
"loss_weights")["aux_loss"] = trial.suggest_float(
"aux_loss", 0.2, 0.8, step=0.2)
configer.get("network",
"loss_weights")["seg_loss"] = trial.suggest_float(
"seg_loss", .5, 2.0, step=0.5)
if configer.get("network", "use_teach"):
configer.update(["teacher_temp"],
trial.suggest_float("teacher_temp", 1, 21, step=5))
configer.update(["teacher_interp"],
trial.suggest_float("teacher_interp",
0.1,
1,
step=0.1))
else:
configer.update(["teacher_interp"], 0)
lossType = trial.suggest_categorical(
"loss_type", ["fs_auxce_loss", "fs_auxohemce_loss "])
lossType = lossType.replace(" ", "")
configer.update(["loss", "loss_type"], lossType)
if not configer.exists("trial_nb"):
configer.add(["trial_nb"], trial.number)
else:
configer.update(["trial_nb"], trial.number)
model = None
if configer.get('method') == 'fcn_segmentor':
if configer.get('phase') == 'train':
from segmentor.trainer import Trainer
model = Trainer(configer)
elif configer.get('phase') == 'test':
from segmentor.tester import Tester
model = Tester(configer)
elif configer.get('phase') == 'test_offset':
from segmentor.tester_offset import Tester
model = Tester(configer)
else:
Log.error('Method: {} is not valid.'.format(configer.get('task')))
exit(1)
if configer.get('phase') == 'train':
miou = model.train(trial)
return miou
elif configer.get('phase').startswith('test') and configer.get(
'network', 'resume') is not None:
model.test()
else:
Log.error('Phase: {} is not valid.'.format(configer.get('phase')))
exit(1)
log_file, time.strftime("%Y-%m-%d_%X", time.localtime()))
configer.update(['logging', 'log_file'], new_log_file)
else:
configer.update(['logging', 'logfile_level'], None)
Log.init(logfile_level=configer.get('logging', 'logfile_level'),
stdout_level=configer.get('logging', 'stdout_level'),
log_file=configer.get('logging', 'log_file'),
log_format=configer.get('logging', 'log_format'),
rewrite=configer.get('logging', 'rewrite'))
model = None
if configer.get('method') == 'fcn_segmentor':
if configer.get('phase') == 'train':
from segmentor.trainer import Trainer
model = Trainer(configer)
elif configer.get('phase') == 'test':
from segmentor.tester import Tester
model = Tester(configer)
elif configer.get('phase') == 'test_offset':
from segmentor.tester_offset import Tester
model = Tester(configer)
else:
Log.error('Method: {} is not valid.'.format(configer.get('task')))
exit(1)
if configer.get('phase') == 'train':
model.train()
elif configer.get('phase').startswith('test') and configer.get(
'network', 'resume') is not None:
model.test()