std = [0.229]
device = torch.device('cuda')
model_path = 'checkpoints/019.pt'
test_set = TinyMind(mode='test',
root='data/',
fpath="test2.txt",
trans=T.Compose([Norm(mean, std),
to_tensor()]))
test_loader = DataLoader(dataset=test_set,
batch_size=256,
shuffle=False,
num_workers=0)
net = resnet50().to(device)
state = torch.load(model_path)
net.load_state_dict(state['state_dict'])
classes = [
'且', '世', '东', '九', '亭', '今', '从', '令', '作', '使', '侯', '元', '光', '利', '印',
'去', '受', '右', '司', '合', '名', '周', '命', '和', '唯', '堂', '士', '多', '夜', '奉',
'女', '好', '始', '字', '孝', '守', '宗', '官', '定', '宜', '室', '家', '寒', '左', '常',
'建', '徐', '御', '必', '思', '意', '我', '敬', '新', '易', '春', '更', '朝', '李', '来',
'林', '正', '武', '氏', '永', '流', '海', '深', '清', '游', '父', '物', '玉', '用', '申',
'白', '皇', '益', '福', '秋', '立', '章', '老', '臣', '良', '莫', '虎', '衣', '西', '起',
'足', '身', '通', '遂', '重', '陵', '雨', '高', '黄', '鼎'
]
results_csv = csv.writer(
open('./data/result.csv', 'w', encoding='utf-8', newline=''))
results_csv.writerow(['filename', 'label'])
def network(args):
if args.net == 'alexnet':
from models.AlexNet import alexnet
net = alexnet(num_class=args.num_class)
elif args.net == 'fractalnet34':
from models.FratalNet import fractalnet34
net = fractalnet34(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'googlenet':
from models.GoogLeNet import googlenet
net = googlenet(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'inception_v3':
from models.InceptionV3 import inception_v3
net = inception_v3(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'mobilenet_v1':
from models.MobileNetV1 import mobilenet_v1
net = mobilenet_v1(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'preactresnet18':
from models.PreActResNet import peractresnet18
net = peractresnet18(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'preactresnet34':
from models.PreActResNet import peractresnet34
net = peractresnet34(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'preactresnet50':
from models.PreActResNet import peractresnet50
net = peractresnet50(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'preactresnet101':
from models.PreActResNet import peractresnet101
net = peractresnet101(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'preactresnet152':
from models.PreActResNet import peractresnet152
net = peractresnet152(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'resnet18':
from models.ResNet import resnet18
net = resnet18(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'resnet34':
from models.ResNet import resnet34
net = resnet34(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'resnet50':
from models.ResNet import resnet50
net = resnet50(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'resnet101':
from models.ResNet import resnet101
net = resnet101(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'resnet152':
from models.ResNet import resnet152
net = resnet152(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'se_resnet50':
from models.SEResNet import se_resnet50
net = se_resnet50(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'se_resnet101':
from models.SEResNet import se_resnet101
net = se_resnet101(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'se_resnet152':
from models.SEResNet import se_resnet152
net = se_resnet152(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'shufflenet_05x_g1':
from models.ShuffleNet import shufflenet_05x_g1
net = shufflenet_05x_g1(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'shufflenet_05x_g2':
from models.ShuffleNet import shufflenet_05x_g2
net = shufflenet_05x_g2(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'shufflenet_05x_g3':
from models.ShuffleNet import shufflenet_05x_g3
net = shufflenet_05x_g3(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'shufflenet_05x_g4':
from models.ShuffleNet import shufflenet_05x_g4
net = shufflenet_05x_g4(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'shufflenet_025x_g1':
from models.ShuffleNet import shufflenet_025x_g1
net = shufflenet_025x_g1(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'shufflenet_025x_g2':
from models.ShuffleNet import shufflenet_025x_g2
net = shufflenet_025x_g2(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'shufflenet_025x_g3':
from models.ShuffleNet import shufflenet_025x_g3
net = shufflenet_025x_g3(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'shufflenet_025x_g4':
from models.ShuffleNet import shufflenet_025x_g4
net = shufflenet_025x_g4(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'shufflenet_1x_g1':
from models.ShuffleNet import shufflenet_1x_g1
net = shufflenet_1x_g1(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'shufflenet_1x_g2':
from models.ShuffleNet import shufflenet_1x_g2
net = shufflenet_1x_g2(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'shufflenet_1x_g3':
from models.ShuffleNet import shufflenet_1x_g3
net = shufflenet_1x_g3(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'shufflenet_1x_g4':
from models.ShuffleNet import shufflenet_1x_g4
net = shufflenet_1x_g4(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'squeezenet':
from models.SqueezeNet import squeezenet
net = squeezenet(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'stochastic_depth18':
from models.StochasticDepth import stochastic_depth18
net = stochastic_depth18(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'stochastic_depth34':
from models.StochasticDepth import stochastic_depth34
net = stochastic_depth34(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'stochastic_depth50':
from models.StochasticDepth import stochastic_depth50
net = stochastic_depth50(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'stochastic_depth101':
from models.StochasticDepth import stochastic_depth101
net = stochastic_depth101(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'stochastic_depth152':
from models.StochasticDepth import stochastic_depth152
net = stochastic_depth152(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'vgg_11':
from models.VGG import vgg_11
net = vgg_11(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'vgg_11_lrn':
from models.VGG import vgg_11_lrn
net = vgg_11_lrn(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'vgg_13':
from models.VGG import vgg_13
net = vgg_13(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'vgg_16_c':
from models.VGG import vgg_16_c
net = vgg_16_c(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'vgg_16_d':
from models.VGG import vgg_16_d
net = vgg_16_d(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'vgg_19':
from models.VGG import vgg_19
net = vgg_19(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'wide_resnet40_4':
from models.Wide_ResNet import wide_resnet40_4
net = wide_resnet40_4(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'wide_resnet16_8':
from models.Wide_ResNet import wide_resnet16_8
net = wide_resnet16_8(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'wide_resnet28_10':
from models.Wide_ResNet import wide_resnet28_10
net = wide_resnet28_10(num_class=args.num_class,
initialize=args.initialize)
elif args.net == 'resnet110':
from models.Wide_ResNet import resnet110
net = resnet110(num_class=args.num_class, initialize=args.initialize)
elif args.net == 'xception':
from models.Xception import xception
net = xception(num_class=args.num_class, initialize=args.initialize)
return net
# train teacher model from pretrained resnet
teacher_model = torch_models.resnet34(pretrained=True)
teacher_model.fc = nn.Linear(512, 10)
# freeze earlier layers (first 7 out of 10)
ct = 0
for child in teacher_model.children():
ct += 1
if ct < 7:
for param in child.parameters():
param.requires_grad = False
teacher_model = teacher_model.to(device)
teacher_model = nn.DataParallel(teacher_model, device_ids=[0, 1])
# teacher_model.load_state_dict(checkpoint['teacher_model'])
student_model = resnet50(in_channels=3, n_classes=10).to(device)
student_model = nn.DataParallel(student_model, device_ids=[0, 1])
# student_model.load_state_dict(checkpoint['student_model'])
with wandb.init(project="MPL-Cifar10"):
train_mpl(teacher_model,
student_model,
sup_dl,
unsup_dl,
batch_size=BATCH_SIZE,
dataset=DATASET,
num_epochs=N_EPOCHS,
learning_rate=LR,
weight_u=WEIGHT_U,
uda_threshold=UDA_THRESHOLD,
n_student_steps=N_STUDENT_STEPS,
download=True,
transform=transform)
val_loader = torch.utils.data.DataLoader(valset,
batch_size=10,
shuffle=True,
num_workers=2)
alexnet = AlexNet().cuda()
alexnetexpander = alexnetexpander(sparsity=10).cuda()
alexnetskip = alexnetskip(sparsity=10).cuda()
resnet34 = resnet34().cuda()
resnet34expander = resnet34expander(sparsity=5).cuda()
resnet34skip = resnet34skip(sparsity=5).cuda()
resnet50 = resnet50().cuda()
resnet50expander = resnet50expander(sparsity=10).cuda()
resnet50skip = resnet50skip(sparsity=10).cuda()
vgg16 = vgg16().cuda()
vggexpander16 = vggexpander16(sparsity=10).cuda()
vggskip16 = vggskip16(sparsity=2).cuda()
#model_10 = Netexpander(sparsity=45).cuda()
model = Net().cuda()
modelx = Netexpander(sparsity=10).cuda()
#Wprune = adj_matrix(modelx,10)
#W_binary = adj_matrix(model,10)
#Wprune,eprune,vprune = EVD(modelx)
#raw = low_rank(Wprune,eprune,vprune)