def _register_classes(module, superclass, prefix=None, sep='.'):
for name in dir(module):
attr = getattr(module, name)
if isinstance(attr, type) and issubclass(attr, superclass):
if attr is superclass:
continue
if prefix is not None:
name = prefix + sep + name
mlconfig.register(attr, name=name)
import mlconfig from torch import optim from .rmsprop import TFRMSprop mlconfig.register(optim.SGD) mlconfig.register(optim.Adam) mlconfig.register(optim.lr_scheduler.MultiStepLR) mlconfig.register(optim.lr_scheduler.StepLR) mlconfig.register(optim.lr_scheduler.ExponentialLR)
import torch
import torch.nn.functional as F
import numpy as np
import mlconfig
mlconfig.register(torch.nn.CrossEntropyLoss)
if torch.cuda.is_available():
torch.backends.cudnn.benchmark = True
if torch.cuda.device_count() > 1:
device = torch.device('cuda:0')
else:
device = torch.device('cuda')
else:
device = torch.device('cpu')
@mlconfig.register
class SCELoss(torch.nn.Module):
def __init__(self, alpha, beta, num_classes=10):
super(SCELoss, self).__init__()
self.device = device
self.alpha = alpha
self.beta = beta
self.num_classes = num_classes
self.cross_entropy = torch.nn.CrossEntropyLoss()
def forward(self, pred, labels):
# CCE
ce = self.cross_entropy(pred, labels)
# RCE
def registerClasses():
global registered
if not registered:
mlconfig.register(ResCNN)
mlconfig.register(CachedNonLinearSelfPlay)
mlconfig.register(CachedLinearSelfPlay)
mlconfig.register(RemoteEvaluationAccess)
mlconfig.register(EvaluationWorker)
mlconfig.register(LocalEvaluationAccess)
mlconfig.register(FakeEvaluationAccess)
mlconfig.register(TreeSelfPlayWorker)
mlconfig.register(LearntThinkDecider)
mlconfig.register(PointsGaussServerLeague)
mlconfig.register(EloGaussServerLeague)
mlconfig.register(LeaguePlayerAccess)
mlconfig.register(FixedPlayerAccess)
mlconfig.register(FixedThinkDecider)
mlconfig.register(LeagueSelfPlayerWorker)
mlconfig.register(NoopPolicyUpdater)
mlconfig.register(NoopGameReporter)
mlconfig.register(DatasetPolicyTester)
mlconfig.register(ShuffleBatchedPolicyPlayer)
mlconfig.register(SolverBatchedPolicyPlayer)
mlconfig.register(PolicyPlayer)
mlconfig.register(PolicyIteratorPlayer)
mlconfig.register(LinearSelfPlayWorker)
mlconfig.register(MctsPolicyIterator)
mlconfig.register(TemperatureMoveDecider)
mlconfig.register(MNKGameState)
mlconfig.register(SingleProcessReporter)
mlconfig.register(SingleProcessUpdater)
mlconfig.register(PytorchPolicy)
mlconfig.register(dict)
mlconfig.register(PlayVs)
mlconfig.register(HumanMNKInterface)
mlconfig.register(Connect4GameState)
mlconfig.register(HumanConnect4Interface)
mlconfig.register(RandomPlayPolicy)
mlconfig.register(PonsSolver)
mlconfig.register(TestDatabaseGenerator)
mlconfig.register(BestPlayPolicy)
mlconfig.register(DistributedNetworkUpdater)
mlconfig.register(DistributedReporter)
mlconfig.register(TrainingWorker)
mlconfig.register(ConstantTrainingWindowManager)
mlconfig.register(StreamTrainingWorker)
mlconfig.register(ConstantWindowSizeManager)
mlconfig.register(LrStepSchedule)
mlconfig.register(SemiPerfectPolicy)
mlconfig.register(TestDatabaseGenerator2)
mlconfig.register(DatasetPolicyTester2)
mlconfig.register(FilePolicyUpdater)
mlconfig.register(SupervisedNetworkTrainer)
mlconfig.register(DistributedNetworkUpdater2)
mlconfig.register(StreamTrainingWorker2)
mlconfig.register(OneCycleSchedule)
mlconfig.register(SlowWindowSizeManager)
registered = True
import argparse
import datetime
import os
import shutil
import time
import numpy as np
import dataset
import mlconfig
import torch
import util
import madrys
import models
from evaluator import Evaluator
from trainer import Trainer
mlconfig.register(madrys.MadrysLoss)
# General Options
parser = argparse.ArgumentParser(description='ClasswiseNoise')
parser.add_argument('--seed', type=int, default=0, help='seed')
parser.add_argument('--version', type=str, default="resnet18")
parser.add_argument('--exp_name', type=str, default="test_exp")
parser.add_argument('--config_path', type=str, default='configs/cifar10')
parser.add_argument('--load_model', action='store_true', default=False)
parser.add_argument('--data_parallel', action='store_true', default=False)
parser.add_argument('--train', action='store_true', default=False)
parser.add_argument('--save_frequency', default=-1, type=int)
# Datasets Options
parser.add_argument('--train_face', action='store_true', default=False)
parser.add_argument('--train_portion', default=1.0, type=float)
parser.add_argument('--train_batch_size',
default=128,
import mlconfig from torch import optim mlconfig.register(optim.Adam) mlconfig.register(optim.lr_scheduler.StepLR)
import torch
import torch.nn as nn
import torch.nn.functional as F
import mlconfig
import torchvision
mlconfig.register(torchvision.models.resnet50)
mlconfig.register(torch.optim.SGD)
mlconfig.register(torch.optim.Adam)
mlconfig.register(torch.optim.lr_scheduler.MultiStepLR)
mlconfig.register(torch.optim.lr_scheduler.CosineAnnealingLR)
mlconfig.register(torch.optim.lr_scheduler.StepLR)
mlconfig.register(torch.optim.lr_scheduler.ExponentialLR)
class ConvBrunch(nn.Module):
def __init__(self, in_planes, out_planes, kernel_size=3):
super(ConvBrunch, self).__init__()
padding = (kernel_size - 1) // 2
self.out_conv = nn.Sequential(
nn.Conv2d(in_planes,
out_planes,
kernel_size=kernel_size,
padding=padding), nn.BatchNorm2d(out_planes), nn.ReLU())
def forward(self, x):
return self.out_conv(x)
@mlconfig.register
class ToyModel(nn.Module):
def __init__(self, type='CIFAR10'):
import mlconfig import torch import torch.nn as nn import torchvision from . import DenseNet, ResNet, ToyModel, inception_resnet_v1 mlconfig.register(torch.optim.SGD) mlconfig.register(torch.optim.Adam) mlconfig.register(torch.optim.lr_scheduler.MultiStepLR) mlconfig.register(torch.optim.lr_scheduler.CosineAnnealingLR) mlconfig.register(torch.optim.lr_scheduler.StepLR) mlconfig.register(torch.optim.lr_scheduler.ExponentialLR) mlconfig.register(torch.nn.CrossEntropyLoss) # Models mlconfig.register(ResNet.ResNet) mlconfig.register(ResNet.ResNet18) mlconfig.register(ResNet.ResNet34) mlconfig.register(ResNet.ResNet50) mlconfig.register(ResNet.ResNet101) mlconfig.register(ResNet.ResNet152) mlconfig.register(ToyModel.ToyModel) mlconfig.register(DenseNet.DenseNet121) mlconfig.register(inception_resnet_v1.InceptionResnetV1) # torchvision models mlconfig.register(torchvision.models.resnet18) mlconfig.register(torchvision.models.resnet50) mlconfig.register(torchvision.models.densenet121) # CUDA Options