def __init__(self,
layers,
in_channels=192,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1),
pretrained=False,
deep_base=False) -> None:
super(ResNetDCT_345, self).__init__()
if layers == 18:
resnet = resnet18(pretrained, deep_base, strides=strides, dilations=dilations)
elif layers == 34:
resnet = resnet34(pretrained, deep_base, strides=strides, dilations=dilations)
elif layers == 50:
resnet = resnet50(pretrained, deep_base, strides=strides, dilations=dilations)
elif layers == 101:
resnet = resnet101(pretrained, deep_base, strides=strides, dilations=dilations)
self.layer2, self.layer3, self.layer4, self.avgpool, self.fc = \
resnet.layer2, resnet.layer3, resnet.layer4, resnet.avgpool, resnet.fc
self.relu = nn.ReLU(inplace=True)
out_ch = self.layer2[0].conv1.out_channels
ks = self.layer2[0].conv1.kernel_size
stride = self.layer2[0].conv1.stride
padding = self.layer2[0].conv1.padding
self.layer2[0].conv1 = nn.Conv2d(in_channels, out_ch, kernel_size=ks, stride=stride, padding=padding, bias=False)
init_weight(self.layer2[0].conv1)
out_ch = self.layer2[0].downsample[0].out_channels
self.layer2[0].downsample[0] = nn.Conv2d(in_channels, out_ch, kernel_size=1, stride=2, bias=False)
init_weight(self.layer2[0].downsample[0])
def test():
##
config = getConfig()
# define transform image
transform_test = transforms.Compose([
transforms.Resize((config.image_size, config.image_size)),
# transforms.CenterCrop(config.input_size),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
net = resnet101(pretrained=True, use_bap=False)
in_features = net.fc_new.in_features
new_linear = torch.nn.Linear(in_features=in_features, out_features=25)
net.fc_new = new_linear
# load checkpoint
checkpoint_path = os.path.join(config.checkpoint_path,
'model_best.pth.tar')
load_state_dict = torch.load(checkpoint_path,
map_location='cpu')['state_dict']
new_state_dict = {}
for key, value in load_state_dict.items():
new_key = key.replace('module.', '')
new_state_dict[new_key] = value
net.load_state_dict(new_state_dict)
img_dir = config.image
image = Image.open(img_dir).convert('RGB')
image = transform_test(image)
preds, _, _ = net(image.unsqueeze(0))
print(torch.sigmoid(preds))
def get_model(model_name, pho_size=299, num_classes=110):
if model_name == "vgg16":
model = VGG(num_classes=num_classes, pho_size=299)
elif model_name == "resnet101":
model = resnet101(num_classes=num_classes)
elif model_name == "resnet152":
model = resnet152(num_classes=num_classes)
elif model_name == "densenet":
model = DenseNet(growth_rate=12,
block_config=[(100 - 4) // 6 for _ in range(3)],
num_classes=num_classes,
small_inputs=False,
efficient=True,
pho_size=pho_size)
elif model_name == "InceptionResNetV2":
model = InceptionResNetV2(num_classes=num_classes)
elif model_name == "InceptionV4":
model = InceptionV4(num_classes=num_classes)
elif model_name == "Inception3":
model = Inception3(num_classes=num_classes)
elif model_name == "denoise":
model = get_denoise()
elif model_name == "Mymodel":
model = Mymodel()
elif model_name == 'Comdefend':
model = ComDefend()
elif model_name == 'Rectifi':
model = Rectifi()
return model
def __init__(self, layers=18, bins=(1, 2, 3, 6), dropout=0.1, classes=2, zoom_factor=8, use_ppm=True,
criterion=nn.CrossEntropyLoss(ignore_index=255), BatchNorm=nn.BatchNorm2d, flow=False, sd=False,
pretrained=True):
super(PSPNet, self).__init__()
assert layers in [18, 50, 101, 152]
assert 512 % len(bins) == 0
assert classes > 1
assert zoom_factor in [1, 2, 4, 8]
self.zoom_factor = zoom_factor
self.use_ppm = use_ppm
self.flow = flow
self.sd = sd
self.criterion = criterion
models.BatchNorm = BatchNorm
if layers == 50:
resnet = models.resnet50(pretrained=pretrained)
elif layers == 18:
resnet = models.resnet18(deep_base=False, pretrained=pretrained)
elif layers == 101:
resnet = models.resnet101(pretrained=pretrained)
else:
resnet = models.resnet152(pretrained=pretrained)
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool)
self.layer1, self.layer2, self.layer3, self.layer4 = resnet.layer1, resnet.layer2, resnet.layer3, resnet.layer4
for n, m in self.layer3.named_modules():
if 'conv2' in n:
m.dilation, m.padding, m.stride = (2, 2), (2, 2), (1, 1)
elif 'conv1' in n:
m.stride = (1, 1)
elif 'downsample.0' in n:
m.stride = (1, 1)
for n, m in self.layer4.named_modules():
if 'conv2' in n:
m.dilation, m.padding, m.stride = (4, 4), (4, 4), (1, 1)
elif 'conv1' in n:
m.stride = (1, 1)
elif 'downsample.0' in n:
m.stride = (1, 1)
fea_dim = 512
if use_ppm:
self.ppm = PPM(fea_dim, int(fea_dim / len(bins)), bins, BatchNorm)
fea_dim *= 2
self.cls = nn.Sequential(
nn.Conv2d(fea_dim, 256, kernel_size=3, padding=1, bias=False),
BatchNorm(256),
nn.ReLU(inplace=True),
nn.Dropout2d(p=dropout),
nn.Conv2d(256, classes, kernel_size=1)
)
if self.training:
self.aux = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, padding=1, bias=False),
BatchNorm(256),
nn.ReLU(inplace=True),
nn.Dropout2d(p=dropout),
nn.Conv2d(256, classes, kernel_size=1)
)
def __init__(self,
model,
modality='rgb',
inp=3,
num_classes=150,
input_size=224,
input_segments=8,
dropout=0.5):
super(tsn_model, self).__init__()
if modality == 'flow':
inp = 10
self.num_classes = num_classes
self.inp = inp
self.input_segments = input_segments
self._enable_pbn = False
if model == 'resnet18':
self.model = resnet.resnet18(inp=inp, pretrained=True)
elif model == 'resnet34':
self.model = resnet.resnet34(inp=inp, pretrained=True)
elif model == 'resnet50':
self.model = resnet.resnet50(inp=inp, pretrained=True)
elif model == 'resnet101':
self.model = resnet.resnet101(inp=inp, pretrained=True)
elif model == 'bn_inception':
self.model = bn_inception.bninception(inp=inp)
self.modality = modality
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.dropout = nn.Dropout(p=dropout)
in_channels = self.model.fc.in_features
self.model.fc = None
self.fc = nn.Linear(in_channels, num_classes)
self.consensus = basic_ops.ConsensusModule('avg')
def __init__(self,
embedding_size,
num_classes,
backbone='resnet18',
mode='t'):
super(background_resnet, self).__init__()
self.trainMode = mode
self.backbone = backbone
# copying modules from pretrained models
if backbone == 'resnet50':
self.pretrained = resnet.resnet50(pretrained=False)
elif backbone == 'resnet101':
self.pretrained = resnet.resnet101(pretrained=False)
elif backbone == 'resnet152':
self.pretrained = resnet.resnet152(pretrained=False)
elif backbone == 'resnet18':
self.pretrained = resnet.resnet18(pretrained=False)
elif backbone == 'resnet34':
self.pretrained = resnet.resnet34(pretrained=False)
else:
raise RuntimeError('unknown backbone: {}'.format(backbone))
self.fc0 = nn.Linear(128, embedding_size[0])
# task specific layers for task 1
self.fc1 = nn.Linear(128, embedding_size[1])
self.bn1 = nn.BatchNorm1d(embedding_size[1])
self.relu1 = nn.ReLU()
self.last1 = nn.Linear(embedding_size[1], num_classes)
# task speicific layers for task 2
self.fc2 = nn.Linear(128, embedding_size[2])
self.bn2 = nn.BatchNorm1d(embedding_size[2])
self.relu2 = nn.ReLU()
self.last2 = nn.Linear(embedding_size[2], num_classes)
def validate():
##
engine = Engine()
config = getConfig()
device = torch.device("cuda:" + str(config.device))
# define dataset
transform_test = transforms.Compose([
transforms.Resize((config.image_size, config.image_size)),
# transforms.CenterCrop(config.input_size),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
val_dataset = CustomDataset('data/movie_val.csv',
'data/movie/images',
transform=transform_test)
val_loader = DataLoader(val_dataset,
batch_size=config.batch_size,
shuffle=False,
num_workers=config.workers,
pin_memory=True)
net = resnet101(pretrained=True, use_bap=False)
in_features = net.fc_new.in_features
new_linear = torch.nn.Linear(in_features=in_features, out_features=25)
net.fc_new = new_linear
# load checkpoint
use_gpu = torch.cuda.is_available() and config.use_gpu
if use_gpu:
net = net.to(device)
gpu_ids = [int(r) for r in config.gpu_ids.split(',')]
if use_gpu and config.multi_gpu:
net = torch.nn.DataParallel(net, device_ids=gpu_ids)
checkpoint_path = os.path.join(config.checkpoint_path,
'model_best.pth.tar')
load_state_dict = torch.load(checkpoint_path,
map_location=device)['state_dict']
new_state_dict = {}
for key, value in load_state_dict.items():
new_key = key.replace('module.', '')
new_state_dict[new_key] = value
net.load_state_dict(new_state_dict)
# define loss
criterion = torch.nn.BCEWithLogitsLoss()
if use_gpu:
criterion = criterion.cuda()
state = {
'model': net,
'val_loader': val_loader,
'criterion': criterion,
'config': config,
'device': device,
'step': 0,
'lr': config.lr
}
prec1, fprec, val_loss = engine.validate(state)
print(prec1)
def __init__(self,
layers=50,
bins=(1, 2, 3, 6),
dropout=0.1,
classes=2,
zoom_factor=8,
use_ppm=True,
BatchNorm=nn.BatchNorm2d,
pretrained=False):
""" """
super(Backbone, self).__init__()
assert layers in [50, 101, 152]
assert 2048 % len(bins) == 0
assert classes > 1
assert zoom_factor in [1, 2, 4, 8]
self.zoom_factor = zoom_factor
self.use_ppm = use_ppm
models.BatchNorm = BatchNorm
if layers == 50:
resnet = models.resnet50(pretrained=pretrained)
elif layers == 101:
resnet = models.resnet101(pretrained=pretrained)
else:
resnet = models.resnet152(pretrained=pretrained)
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu,
resnet.conv2, resnet.bn2, resnet.relu,
resnet.conv3, resnet.bn3, resnet.relu,
resnet.maxpool)
self.layer1, self.layer2, self.layer3, self.layer4 = resnet.layer1, resnet.layer2, resnet.layer3, resnet.layer4
for n, m in self.layer3.named_modules():
if 'conv2' in n:
m.dilation, m.padding, m.stride = (2, 2), (2, 2), (1, 1)
elif 'downsample.0' in n:
m.stride = (1, 1)
for n, m in self.layer4.named_modules():
if 'conv2' in n:
m.dilation, m.padding, m.stride = (4, 4), (4, 4), (1, 1)
elif 'downsample.0' in n:
m.stride = (1, 1)
fea_dim = 2048
if use_ppm:
self.ppm = PPM(fea_dim, int(fea_dim / len(bins)), bins, BatchNorm)
fea_dim *= 2
self.cls = nn.Sequential(
nn.Conv2d(fea_dim, 512, kernel_size=3, padding=1, bias=False),
BatchNorm(512), nn.ReLU(inplace=True), nn.Dropout2d(p=dropout),
nn.Conv2d(512, classes, kernel_size=1))
if self.training:
self.aux = nn.Sequential(
nn.Conv2d(1024, 256, kernel_size=3, padding=1, bias=False),
BatchNorm(256), nn.ReLU(inplace=True), nn.Dropout2d(p=dropout),
nn.Conv2d(256, classes, kernel_size=1))
def prepare_optimizee(args, sgd_in_names, obs_shape, hidden_size, actor_critic,
current_optimizee_step, prev_optimizee_step):
prev_optimizee_step += current_optimizee_step
current_optimizee_step = 0
model = resnet101(pretrained=True)
num_ftrs = model.fc.in_features
fc_layers = nn.Sequential(
nn.Linear(num_ftrs, 512),
nn.ReLU(inplace=True),
nn.Linear(512, args.num_class),
)
model.fc_new = fc_layers
train_blocks = args.train_blocks.split('.')
# default turn-off fc, turn-on fc_new
for param in model.fc.parameters():
param.requires_grad = False
##### Freeze several bottom layers (Optional) #####
non_train_blocks = [
'conv1', 'bn1', 'layer1', 'layer2', 'layer3', 'layer4', 'fc'
]
for name in train_blocks:
try:
non_train_blocks.remove(name)
except Exception:
print(
"cannot find block name %s\nAvailable blocks are: conv1, bn1, layer1, layer2, layer3, layer4, fc"
% name)
for name in non_train_blocks:
for param in getattr(model, name).parameters():
param.requires_grad = False
# Setup optimizer
sgd_in = []
for name in train_blocks:
if name != 'fc':
sgd_in.append({'params': get_params(model, [name]), 'lr': args.lr})
else:
sgd_in.append({
'params': get_params(model, ["fc_new"]),
'lr': args.lr
})
base_lrs = [group['lr'] for group in sgd_in]
optimizer = SGD(sgd_in,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
model = model.cuda()
model.eval()
return model, optimizer, current_optimizee_step, prev_optimizee_step
def create_model(opt=None):
assert opt.model in ['alexnet', 'googlenet', 'lenet', 'mobilenetv2', 'resnet34', 'resnet101', 'vgg11', 'vgg13', 'vgg16', 'vgg19', 'mobilenetv3small', 'mobilenetv3wen']
if opt.model == 'mobilenetv2':
model = mobilenet.MobileNetV2(num_classes=opt.n_classes, )
print('net is mobilenetv2!')
opt.model_save_dir='./weights/mobilenetv2'
elif opt.model == 'alexnet':
model = alexnet.AlexNet(num_classes=opt.n_classes, init_weights=True)
print('net is alexnet!')
opt.model_save_dir = './weights/alexnet'
elif opt.model == 'googlenet':
model = googlenet.GoogLeNet(num_classes=opt.n_classes, init_weights=True)
print('net is googlenet!')
opt.model_save_dir = './weights/googlenet'
elif opt.model == 'lenet':
model = lenet.LeNet(num_classes=opt.n_classes)
print('net is lenet!')
opt.model_save_dir = './weights/lenet'
elif opt.model == 'resnet34':
model = resnet.resnet34(num_classes=opt.n_classes)
print('net is resnet34!')
opt.model_save_dir = './weights/resnet34'
elif opt.model == 'resnet101':
model = resnet.resnet101(num_classes=opt.n_classes)
print('net is resnet101!')
opt.model_save_dir = './weights/resnet101'
elif opt.model == 'vgg11':
model = vgg.vgg(model_name="vgg11", num_classes=opt.n_classes, init_weights=True)
print('net is vgg11!')
opt.model_save_dir = './weights/vgg11'
elif opt.model == 'vgg13':
model = vgg.vgg(model_name="vgg13", num_classes=opt.n_classes, init_weights=True)
print('net is vgg13!')
opt.model_save_dir = './weights/vgg13'
elif opt.model == 'vgg16':
model = vgg.vgg(model_name="vgg16", num_classes=opt.n_classes, init_weights=True)
print('net is vgg16!')
opt.model_save_dir = './weights/vgg16'
elif opt.model == 'vgg19':
model = vgg.vgg(model_name="vgg19", num_classes=opt.n_classes, init_weights=True)
print('net is vgg19!')
opt.model_save_dir = './weights/vgg19'
elif opt.model == 'mobilenetv3small':
model = mobilenetv3_small.MobileNetV3_small(num_classes=opt.n_classes)
print('net is mobilenetv3small!')
opt.model_save_dir = './weights/mobilenetv3small'
elif opt.model == 'mobilenetv3wen':
model = mobilenetv3_wen.MobileNetV3_small(num_classes=opt.n_classes)
print('net is mobilenetv3wen!')
opt.model_save_dir = './weights/mobilenetv3wen'
return opt, model
def __init__(self, layers=50, dropout=0.1, classes=1, zoom_factor=8, criterion=nn.CrossEntropyLoss(ignore_index=255), BatchNorm=nn.BatchNorm2d, pretrained=True, args=None):
super(MGLNet, self).__init__()
assert layers in [50, 101, 152]
assert classes == 1
assert zoom_factor in [1, 2, 4, 8]
self.zoom_factor = zoom_factor
self.criterion = criterion
self.args = args
models.BatchNorm = BatchNorm
self.gamma = 1.0
if layers == 50:
resnet = models.resnet50(pretrained=pretrained)
elif layers == 101:
resnet = models.resnet101(pretrained=pretrained)
else:
resnet = models.resnet152(pretrained=pretrained)
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu, resnet.conv2, resnet.bn2, resnet.relu, resnet.conv3, resnet.bn3, resnet.relu, resnet.maxpool)
self.layer1, self.layer2, self.layer3, self.layer4 = resnet.layer1, resnet.layer2, resnet.layer3, resnet.layer4
for n, m in self.layer3.named_modules():
if 'conv2' in n:
m.dilation, m.padding, m.stride = (2, 2), (2, 2), (1, 1)
elif 'downsample.0' in n:
m.stride = (1, 1)
for n, m in self.layer4.named_modules():
if 'conv2' in n:
m.dilation, m.padding, m.stride = (4, 4), (4, 4), (1, 1)
elif 'downsample.0' in n:
m.stride = (1, 1)
self.dim = 512
self.pred = nn.Sequential(
nn.Conv2d(2048, self.dim, kernel_size=3, padding=1, bias=False),
BatchNorm(self.dim),
nn.ReLU(inplace=True),
nn.Dropout2d(p=dropout),
nn.Conv2d(self.dim, classes, kernel_size=1)
)
self.region_conv = self.pred[0:4] # 2048 -> 512
self.edge_cat = ConcatNet(BatchNorm) # concat low-level feature map to predict edge
# cascade mutual net
self.mutualnet0 = MutualNet(BatchNorm, dim=self.dim, num_clusters=args.num_clusters, dropout=dropout)
if args.stage == 1:
self.mutualnets = nn.ModuleList([self.mutualnet0])
elif args.stage == 2:
self.mutualnet1 = MutualNet(BatchNorm, dim=self.dim, num_clusters=args.num_clusters, dropout=dropout)
self.mutualnets = nn.ModuleList([self.mutualnet0, self.mutualnet1])
def __init__(self, input_size=(512, 1024)):
super(FPN, self).__init__()
self.input_size = input_size
self.input_row = input_size[0]
self.input_col = input_size[1]
# self.base_net = InceptionResNetV2()
self.base_net = resnet101()
# !!!! Can be replaced with light-head/RFB/ASPP ect. to improve the receptive field !!!!
self.c6 = nn.Conv2d(2048, 256, kernel_size=3, stride=2, padding=1)
self.c5_down = nn.Conv2d(2048, 256, kernel_size=1, stride=1)
self.c4_down = nn.Conv2d(1024, 256, kernel_size=1, stride=1)
self.c3_down = nn.Conv2d(512, 256, kernel_size=1, stride=1)
# !!!! I add another two dilated convolution to it to improve the receptive field !!!!
# !!!! Of course they can be replaced by light-head/ASPP/RBF etc. !!!!
# !!!! Of course they can also be replaced by Inception Block !!!!
self.c2_up = nn.Conv2d(256, 256, kernel_size=1, stride=1)
# !!!! Can be replaced with light-head/RFB/ASPP etc. to improve the receptive field !!!!
# self.conv_fuse = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
self.p4_atrous = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=2,
dilation=2), SwitchNorm2d(256,
using_moving_average=True),
nn.ReLU(inplace=True),
nn.Conv2d(256, 256, kernel_size=1, stride=1, padding=0),
SwitchNorm2d(256, using_moving_average=True),
nn.ReLU(inplace=True))
self.conv_fuse = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=2,
dilation=2))
# nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=4, dilation=4),
# nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=8, dilation=8))
# downsample Conv
self.conv_down = nn.Conv2d(256,
256,
kernel_size=3,
stride=2,
padding=1)
self.conv1x1_cat = nn.Conv2d(512,
256,
kernel_size=1,
stride=1,
padding=0)
self.sn = SwitchNorm2d(256, using_moving_average=True)
self.bn = nn.BatchNorm2d(256, eps=0.001)
def get_model(model_name):
if model_name=="vgg16":
model=VGG(num_classes=110)
elif model_name=="resnet101":
model=resnet101(num_classes=1000)
elif model_name=="densenet":
model=DenseNet(
growth_rate=12,
block_config=[(100 - 4) // 6 for _ in range(3)],
num_classes=110,
small_inputs=False,
efficient=True,
)
else:
model=None
return model
def __init__(self, num_class=19):
super(DFN, self).__init__()
self.num_class = num_class
self.resnet_features = resnet101(pretrained=False)
self.layer0 = nn.Sequential(self.resnet_features.conv1, self.resnet_features.bn1,
self.resnet_features.relu #self.resnet_features.conv3,
#self.resnet_features.bn3, self.resnet_features.relu3
)
self.layer1 = nn.Sequential(self.resnet_features.maxpool, self.resnet_features.layer1)
self.layer2 = self.resnet_features.layer2
self.layer3 = self.resnet_features.layer3
self.layer4 = self.resnet_features.layer4
# this is for smooth network
self.out_conv = nn.Conv2d(2048,self.num_class,kernel_size=1,stride=1)
self.global_pool = nn.AdaptiveAvgPool2d(1)
self.cab1 = CAB(self.num_class*2,self.num_class)
self.cab2 = CAB(self.num_class*2,self.num_class)
self.cab3 = CAB(self.num_class*2,self.num_class)
self.cab4 = CAB(self.num_class*2,self.num_class)
self.rrb_d_1 = RRB(256, self.num_class)
self.rrb_d_2 = RRB(512, self.num_class)
self.rrb_d_3 = RRB(1024, self.num_class)
self.rrb_d_4 = RRB(2048, self.num_class)
self.upsample = nn.Upsample(scale_factor=2,mode="bilinear")
self.upsample_4 = nn.Upsample(scale_factor=4, mode="bilinear")
self.upsample_8 = nn.Upsample(scale_factor=8, mode="bilinear")
self.rrb_u_1 = RRB(self.num_class,self.num_class)
self.rrb_u_2 = RRB(self.num_class,self.num_class)
self.rrb_u_3 = RRB(self.num_class,self.num_class)
self.rrb_u_4 = RRB(self.num_class,self.num_class)
## this is for boarder net work
self.rrb_db_1 = RRB(256, self.num_class)
self.rrb_db_2 = RRB(512, self.num_class)
self.rrb_db_3 = RRB(1024, self.num_class)
self.rrb_db_4 = RRB(2048, self.num_class)
self.rrb_trans_1 = RRB(self.num_class,self.num_class)
self.rrb_trans_2 = RRB(self.num_class,self.num_class)
self.rrb_trans_3 = RRB(self.num_class,self.num_class)
def __init__(self, opt):
super(resnet_mil, self).__init__()
import model.resnet as resnet
resnet = resnet.resnet101()
resnet.load_state_dict(
torch.load('/media/jxgu/d2tb/model/resnet/resnet101.pth'))
self.conv = torch.nn.Sequential()
self.conv.add_module("conv1", resnet.conv1)
self.conv.add_module("bn1", resnet.bn1)
self.conv.add_module("relu", resnet.relu)
self.conv.add_module("maxpool", resnet.maxpool)
self.conv.add_module("layer1", resnet.layer1)
self.conv.add_module("layer2", resnet.layer2)
self.conv.add_module("layer3", resnet.layer3)
self.conv.add_module("layer4", resnet.layer4)
self.l1 = nn.Sequential(nn.Linear(2048, 1000), nn.ReLU(True),
nn.Dropout(0.5))
self.att_size = 7
self.pool_mil = nn.MaxPool2d(kernel_size=self.att_size, stride=0)
def __init__(self, opt):
super(resnet_mil, self).__init__()
import model.resnet as resnet
resnet = resnet.resnet101()
resnet.load_state_dict(torch.load('/media/jxgu/d2tb/model/resnet/resnet101.pth'))
self.conv = torch.nn.Sequential()
self.conv.add_module("conv1", resnet.conv1)
self.conv.add_module("bn1", resnet.bn1)
self.conv.add_module("relu", resnet.relu)
self.conv.add_module("maxpool", resnet.maxpool)
self.conv.add_module("layer1", resnet.layer1)
self.conv.add_module("layer2", resnet.layer2)
self.conv.add_module("layer3", resnet.layer3)
self.conv.add_module("layer4", resnet.layer4)
self.l1 = nn.Sequential(nn.Linear(2048, 1000),
nn.ReLU(True),
nn.Dropout(0.5))
self.att_size = 7
self.pool_mil = nn.MaxPool2d(kernel_size=self.att_size, stride=0)
def __init__(self, embedding_size, num_classes, backbone='resnet50'):
super(background_resnet, self).__init__()
self.backbone = backbone
# copying modules from pretrained models
if backbone == 'resnet50':
self.pretrained = resnet.resnet50(pretrained=False)
elif backbone == 'resnet101':
self.pretrained = resnet.resnet101(pretrained=False)
elif backbone == 'resnet152':
self.pretrained = resnet.resnet152(pretrained=False)
elif backbone == 'resnet18':
self.pretrained = resnet.resnet18(pretrained=False)
elif backbone == 'resnet34':
self.pretrained = resnet.resnet34(pretrained=False)
else:
raise RuntimeError('unknown backbone: {}'.format(backbone))
self.fc0 = nn.Linear(512, embedding_size)
self.bn0 = nn.BatchNorm1d(embedding_size)
self.relu = nn.ReLU()
self.last = nn.Linear(embedding_size, num_classes)
def __init__(self,
layers,
in_channels=192,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1),
pretrained=False,
deep_base=False) -> None:
super(ResNetDCT_2345, self).__init__()
self.layers = layers
if layers == 18:
resnet = resnet18(pretrained, deep_base, strides=strides, dilations=dilations)
elif layers == 34:
resnet = resnet34(pretrained, deep_base, strides=strides, dilations=dilations)
elif layers == 50:
resnet = resnet50(pretrained, deep_base, strides=strides, dilations=dilations)
elif layers == 101:
resnet = resnet101(pretrained, deep_base, strides=strides, dilations=dilations)
self.layer1, self.layer2, self.layer3, self.layer4, self.avgpool, self.fc = \
resnet.layer1, resnet.layer2, resnet.layer3, resnet.layer4, resnet.avgpool, resnet.fc
self.relu = nn.ReLU(inplace=True)
if layers in [18, 34]:
in_ch = self.layer1[0].conv1.in_channels
self.down_layer = nn.Sequential(
nn.Conv2d(in_channels, in_ch, kernel_size=1, stride=1, bias=False),
nn.BatchNorm2d(in_ch),
nn.ReLU(inplace=True)
)
# initialize the weight for only one layer
for m in self.down_layer.modules():
init_weight(m)
else:
out_ch = self.layer1[0].conv1.out_channels
self.layer1[0].conv1 = nn.Conv2d(in_channels, out_ch, kernel_size=1, stride=1, bias=False)
init_weight(self.layer1[0].conv1)
out_ch = self.layer1[0].downsample[0].out_channels
self.layer1[0].downsample[0] = nn.Conv2d(in_channels, out_ch, kernel_size=1, stride=1, bias=False)
init_weight(self.layer1[0].downsample[0])
def get_model(model_type='resnet50', num_classes=1000):
# TODO: Add more backbones
if model_type == 'resnet34':
model = resnet.resnet34(pretrained=True)
elif model_type == 'resnet50':
model = resnet.resnet50(pretrained=True)
elif model_type == 'resnet101':
model = resnet.resnet101(pretrained=True)
elif model_type == 'resnet152':
model = resnet.resnet152(pretrained=True)
elif model_type == 'resnext50_32x4d':
model = resnet.resnext50_32x4d(pretrained=True)
elif model_type == 'resnext101_32x8d':
model = resnet.resnext101_32x8d(pretrained=True)
elif model_type == 'res2net_v1b_50':
model = res2net50_v1b_26w_4s(pretrained=True)
elif model_type == 'res2net_v1b_101':
model = res2net101_v1b_26w_4s(pretrained=True)
elif model_type == 'res2net50_26w_4s':
model = res2net50_26w_4s(pretrained=True)
elif model_type == 'res2net101_26w_4s':
model = res2net101_26w_4s(pretrained=True)
elif model_type == 'res2next50':
model = res2next50(pretrained=True)
elif model_type == 'senet154':
model = senet.senet154(num_classes=num_classes, pretrained='imagenet')
elif model_type == 'resnest50':
model = resnest50(pretrained=True)
elif model_type == 'resnest101':
model = resnest101(pretrained=True)
elif model_type == 'resnest200':
model = resnest200(pretrained=True)
elif model_type == 'resnest269':
model = resnest269(pretrained=True)
else:
model = resnet.resnet50(pretrained=True)
return model
def test():
##
config = getConfig()
device = torch.device("cuda:" + str(config.device))
# define transform image
transform_test = transforms.Compose([
transforms.Resize((config.image_size, config.image_size)),
# transforms.CenterCrop(config.input_size),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
net = resnet101(pretrained=True, use_bap=False)
in_features = net.fc_new.in_features
new_linear = torch.nn.Linear(in_features=in_features, out_features=25)
net.fc_new = new_linear
# load checkpoint
use_gpu = torch.cuda.is_available() and config.use_gpu
if use_gpu:
net = net.to(device)
gpu_ids = [int(r) for r in config.gpu_ids.split(',')]
if use_gpu and config.multi_gpu:
net = torch.nn.DataParallel(net, device_ids=gpu_ids)
checkpoint_path = os.path.join(config.checkpoint_path,
'model_best.pth.tar')
load_state_dict = torch.load(checkpoint_path,
map_location=device)['state_dict']
new_state_dict = {}
for key, value in load_state_dict.items():
new_key = key.replace('module.', '')
new_state_dict[new_key] = value
net.load_state_dict(new_state_dict)
img_dir = config.image
image = Image.open(img_dir).convert('RGB')
image = transform_test(image)
preds, _, _ = net(image.unsqueeze(0).to(device))
print(torch.sigmoid(preds))
elif args.layer == '34':
net = resnet34(pretrained=False,
progress=True,
activate=activate,
hidden=hidden,
num_classes=10)
elif args.layer == '50':
net = resnet50(pretrained=False,
progress=True,
activate=activate,
hidden=hidden,
num_classes=10)
elif args.layer == '101':
net = resnet101(pretrained=False,
progress=True,
activate=activate,
hidden=hidden,
num_classes=10)
elif args.layer == '152':
net = resnet152(pretrained=False,
progress=True,
activate=activate,
hidden=hidden,
num_classes=10)
else:
raise ValueError('layer should be 18 / 34 / 50 / 101 / 152')
net.to(device)
net.initialize_weights(zero_init_residual=True)
# ============================ step 3/6 损失函数 ============================
parser.add_argument("--gamma_name", type=str)
parser.add_argument("--gamma_name_path", type=str)
parser.add_argument("--record_path", type=str)
args = parser.parse_args()
######## Choose whether to use Cuda
device = torch.device("cuda" if (
args.use_cuda and torch.cuda.is_available()) else "cpu")
torch.cuda.empty_cache()
####### Pick according model
if args.model_name == 'res50':
model = resnet.resnet50().to(device)
elif args.model_name == 'res101':
model = resnet.resnet101().to(device)
elif args.model_name == 'res152':
model = resnet.resnet152().to(device)
elif args.model_name == 'res34':
model = resnet.resnet34().to(device)
elif args.model_name == 'res18':
model = resnet.resnet18().to(device)
elif args.model_name == 'alexnet':
model = alexnet.alexnet().to(device)
else:
print('Wrong Model Name')
########### Whether to parallel the model
if args.use_cuda:
if args.parallel:
model = nn.DataParallel(model)
def train():
# input params
config = getConfig()
torch.manual_seed(GLOBAL_SEED)
torch.cuda.manual_seed(GLOBAL_SEED)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = True
torch.cuda.set_device(config.device)
best_prec1 = 0.
rate = 0.875
device = torch.device("cuda:" + str(config.device))
# define train_dataset and loader
transform_train = transforms.Compose([
transforms.Resize(
(int(config.input_size // rate), int(config.input_size // rate))),
transforms.RandomCrop((config.input_size, config.input_size)),
transforms.RandomVerticalFlip(),
transforms.ColorJitter(brightness=32. / 255., saturation=0.5),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
train_dataset = CustomDataset('data/movie_train.csv',
'data/movie/images',
transform=transform_train)
train_loader = DataLoader(train_dataset,
batch_size=config.batch_size,
shuffle=True,
num_workers=config.workers,
pin_memory=True,
worker_init_fn=_init_fn)
transform_test = transforms.Compose([
transforms.Resize((config.image_size, config.image_size)),
transforms.CenterCrop(config.input_size),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
val_dataset = CustomDataset('data/movie_val.csv',
'data/movie/images',
transform=transform_test)
val_loader = DataLoader(val_dataset,
batch_size=config.batch_size * 2,
shuffle=False,
num_workers=config.workers,
pin_memory=True,
worker_init_fn=_init_fn)
# logging dataset info
print('Train:[{train_num}], Val:[{val_num}]'.format(
train_num=len(train_dataset), val_num=len(val_dataset)))
print('Batch Size:[{0}], Total:::Train Batches:[{1}],Val Batches:[{2}]'.
format(config.batch_size, len(train_loader), len(val_loader)))
net = resnet101(pretrained=True, use_bap=False)
in_features = net.fc_new.in_features
new_linear = torch.nn.Linear(in_features=in_features,
out_features=train_dataset.num_classes)
net.fc_new = new_linear
# gpu config
use_gpu = torch.cuda.is_available() and config.use_gpu
if use_gpu:
net = net.to(device)
gpu_ids = [int(r) for r in config.gpu_ids.split(',')]
if use_gpu and config.multi_gpu:
net = torch.nn.DataParallel(net, device_ids=gpu_ids)
# define optimizer
assert config.optim in ['sgd', 'adam'], 'optim name not found!'
if config.optim == 'sgd':
optimizer = torch.optim.SGD(net.parameters(),
lr=config.lr,
momentum=config.momentum,
weight_decay=config.weight_decay)
elif config.optim == 'adam':
optimizer = torch.optim.Adam(net.parameters(),
lr=config.lr,
weight_decay=config.weight_decay)
# define learning scheduler
assert config.scheduler in ['plateau', 'step', 'cosine_annealing'
], 'scheduler not supported!!!'
if config.scheduler == 'plateau':
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
patience=3,
factor=0.1)
elif config.scheduler == 'step':
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=2,
gamma=0.9)
# define loss
criterion = torch.nn.BCEWithLogitsLoss()
# train val parameters dict
state = {
'model': net,
'train_loader': train_loader,
'val_loader': val_loader,
'criterion': criterion,
'config': config,
'optimizer': optimizer,
'device': device,
'step': 0,
'lr': config.lr
}
## train and val
engine = Engine()
print(config)
best_prec = 0
for e in range(config.epochs):
if config.scheduler == 'step':
scheduler.step()
lr_val = get_lr(optimizer)
print("Start epoch %d ==========,lr=%f" % (e, lr_val))
train_prec, train_loss = engine.train(state, e)
prec, val_loss = engine.validate(state)
is_best = prec > best_prec
best_prec = max(prec, best_prec)
print('Epoch: {}, Train-Loss: {:.4f}, Train-accuracy: {:.4f},'\
'Test-accuracy: {:.4f}'.format(e + 1, train_loss, train_prec, prec))
print('Best accuracy: {:.4f}'.format(best_prec))
save_checkpoint(
{
'epoch': e + 1,
'state_dict': net.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, config.output)
if config.scheduler == 'plateau':
scheduler.step(val_loss)
'./data/celebA/list_attr_celeba.txt', '0',
'./data/celebA/img_align_celeba/', transform_train)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=opt.batchSize,
shuffle=True,
num_workers=opt.workers)
valset = celeba('./data/celebA/list_eval_partition.txt',
'./data/celebA/list_attr_celeba.txt', '1',
'./data/celebA/img_align_celeba/', transform_val)
valloader = torch.utils.data.DataLoader(valset,
batch_size=opt.batchSize,
shuffle=True,
num_workers=opt.workers)
resnet = resnet101(num_classes=40)
resnet.cuda()
criterion = nn.MSELoss(reduce=False)
optimizer = optim.SGD(resnet.parameters(),
lr=opt.lr,
momentum=0.9,
weight_decay=5e-4)
scheduler = StepLR(optimizer, step_size=3)
kk = 100
count = 0
weight = torch.FloatTensor(40).fill_(1.0).cuda()
print(type(weight))
val_loss = []
def validation():
record_dir = 'record/%s/%s' % (args.dataset, args.method)
if not os.path.exists(record_dir):
os.makedirs(record_dir)
record_file = os.path.join(
record_dir, '%s_net_%s_%s_to_%s_num_%s' %
(args.method, args.net, args.source, args.target, args.num))
torch.cuda.manual_seed(args.seed)
if args.net == 'resnet34':
G = resnet34()
inc = 512
elif args.net == 'resnet50':
G = resnet50()
inc = 2048
elif args.net == 'resnet101':
G = resnet101()
inc = 2048
elif args.net == "alexnet":
G = AlexNetBase()
inc = 4096
elif args.net == "vgg":
G = VGGBase()
inc = 4096
else:
raise ValueError('Model cannot be recognized.')
params = []
for key, value in dict(G.named_parameters()).items():
if value.requires_grad:
if 'classifier' not in key:
params += [{
def __init__(self,
layers=50,
bins=(1, 2, 3, 6),
dropout=0.1,
classes=2,
zoom_factor=8,
use_ppm=True,
criterion=nn.CrossEntropyLoss(ignore_index=255),
BatchNorm=nn.BatchNorm2d,
pretrained=True):
super(PSPNet, self).__init__()
assert layers in [18, 50, 101, 152]
assert 2048 % len(bins) == 0
assert classes > 1
assert zoom_factor in [1, 2, 4, 8]
self.zoom_factor = zoom_factor
self.use_ppm = use_ppm
self.criterion = criterion
self.criterion_reg = nn.MSELoss(reduce=False)
models.BatchNorm = BatchNorm
if layers == 18:
resnet = models_origin.resnet18(pretrained=True)
elif layers == 50:
resnet = models.resnet50(pretrained=pretrained)
elif layers == 101:
resnet = models.resnet101(pretrained=pretrained)
else:
resnet = models.resnet152(pretrained=pretrained)
'''self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool)
self.layer1, self.layer2, self.layer3, self.layer4 = resnet.layer1, resnet.layer2, resnet.layer3, resnet.layer4
for n, m in self.layer3.named_modules():
if 'conv2' in n:
m.dilation, m.padding, m.stride = (2, 2), (2, 2), (1, 1)
elif 'downsample.0' in n:
m.stride = (1, 1)
for n, m in self.layer4.named_modules():
if 'conv2' in n:
m.dilation, m.padding, m.stride = (4, 4), (4, 4), (1, 1)
elif 'downsample.0' in n:
m.stride = (1, 1)
fea_dim = 512
if use_ppm:
self.ppm = PPM(fea_dim, int(fea_dim/len(bins)), bins, BatchNorm)
fea_dim *= 2'''
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu,
resnet.conv2, resnet.bn2, resnet.relu,
resnet.conv3, resnet.bn3, resnet.relu,
resnet.maxpool)
self.layer1, self.layer2, self.layer3, self.layer4 = resnet.layer1, resnet.layer2, resnet.layer3, resnet.layer4
for n, m in self.layer3.named_modules():
if 'conv2' in n:
m.dilation, m.padding, m.stride = (2, 2), (2, 2), (1, 1)
elif 'downsample.0' in n:
m.stride = (1, 1)
for n, m in self.layer4.named_modules():
if 'conv2' in n:
m.dilation, m.padding, m.stride = (4, 4), (4, 4), (1, 1)
elif 'downsample.0' in n:
m.stride = (1, 1)
fea_dim = 2048
if use_ppm:
self.ppm = PPM(fea_dim, int(fea_dim / len(bins)), bins, BatchNorm)
fea_dim *= 2
'''self.cls = nn.Sequential(
nn.Conv2d(fea_dim, 512, kernel_size=3, padding=1, bias=False),
BatchNorm(512),
nn.ReLU(inplace=True),
nn.Dropout2d(p=dropout),
nn.Conv2d(512, classes, kernel_size=1)
)'''
self.cls3 = nn.Sequential(
nn.Conv2d(fea_dim, 512, kernel_size=3, padding=1, bias=False),
BatchNorm(512), nn.ReLU(inplace=True),
nn.Conv2d(512, 96 * 2, kernel_size=1), BatchNorm(96 * 2),
nn.ReLU(inplace=True))
self.cls2 = nn.Sequential(
nn.Conv2d(96 * 2 + 96, 96, kernel_size=3, padding=1, bias=False),
BatchNorm(96), nn.ReLU(inplace=True), nn.Dropout2d(p=dropout),
nn.Conv2d(96, classes, kernel_size=1))
self.reg = nn.Sequential(
nn.Conv2d(fea_dim, 512, kernel_size=3, padding=1, bias=False),
BatchNorm(512), nn.ReLU(inplace=True),
nn.Conv2d(512, 96, kernel_size=1))
self.reg2 = nn.Sequential(
nn.Conv2d(96, 96, kernel_size=3, padding=1, bias=False),
BatchNorm(96), nn.ReLU(inplace=True),
nn.Conv2d(96, 96, kernel_size=1), nn.ReLU(inplace=True))
self.reg3 = nn.Sequential(
nn.Conv2d(96, 96, kernel_size=3, padding=1, bias=False),
BatchNorm(96), nn.ReLU(inplace=True),
nn.Conv2d(96, 96, kernel_size=1), nn.ReLU(inplace=True))
self.bn = BatchNorm(96)
self.bn2 = BatchNorm(96)
self.mean = torch.tensor(np.load('../data/meanvar/mean.npy'),
requires_grad=False)
self.var = torch.tensor(np.sqrt(np.load('../data/meanvar/var.npy')),
requires_grad=False)
self.mean_2d = torch.tensor(np.load('../data/meanvar/mean_2d.npy'),
requires_grad=False) #.cuda().float()
self.var_2d = torch.tensor(np.sqrt(
np.load('../data/meanvar/var_2d.npy')),
requires_grad=False) #.cuda().float()
self.var = self.var[0, :]
self.mean = torch.unsqueeze(self.mean, 0)
self.mean = torch.unsqueeze(self.mean, 2)
self.mean = torch.unsqueeze(self.mean, 2)
self.mean = self.mean.repeat(1, 1, 60, 60)
self.var = torch.unsqueeze(self.var, 0)
self.var = torch.unsqueeze(self.var, 2)
self.var = torch.unsqueeze(self.var, 2)
self.var = self.var.repeat(1, 1, 60, 60)
writer = SummaryWriter(tf_record)
else:
writer = None
# batch size
train_batchSize = [args.label_batch_size, args.unlabel_batch_size]
# backbone architecture
if args.arch == 'resnet18':
backbone = resnet.resnet18(feature_len=args.feat_len)
elif args.arch == 'resnet34':
backbone = resnet.resnet34(feature_len=args.feat_len)
elif args.arch == 'resnet50':
backbone = resnet.resnet50(feature_len=args.feat_len)
elif args.arch == 'resnet101':
backbone = resnet.resnet101(feature_len=args.feat_len)
elif args.arch == 'resnet152':
backbone = resnet.resnet152(feature_len=args.feat_len)
elif args.arch == 'usr':
backbone = model_usr
else:
raise NameError(
'Arch %s is not support. Please enter from [resnet18, resnet34, resnet50, resnet101, resnet152, usr]'
% args.arch)
# head
model_head = arc.ArcMarginProduct_virface(in_features=args.feat_len,
out_features=args.num_ids,
s=32,
m=0.5,
device='cuda')
experiment_name = str(N_CLASSES) + '_classes' + '_1000_train_16'
model_path = os.path.join('model', 'resnet101-5d3b4d8f.pth')
#model_path = os.path.join(save_dir, experiment_name, MODEL_NAME)
if resume and os.path.isfile(model_path):
model = torch.load(model_path)
model.eval()
with open(os.path.join(save_dir, experiment_name, MODEL_LOG_NAME),
'r') as fp:
words = fp.readline().split()
starting_iter = int(words[2])
starting_epoch = int(words[5])
print("Resuming model at epoch %d and iter %d, with %d classes" %
(starting_epoch, starting_iter, N_CLASSES))
else:
model = resnet101(pretrained=pretrained, num_classes=N_CLASSES)
model = nn.DataParallel(model)
starting_iter = 0
starting_epoch = 0
print(
"Training ResNet-101 from scratch with %d classes (initialized with pretrained weights)"
% N_CLASSES)
model.cuda()
# Cross entropy loss takes the logits directly, so we don't need to apply softmax in our CNN
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(),
lr=0.1,
weight_decay=0.0001,
momentum=0.9)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.1,
def main():
args = get_args()
PID = os.getpid()
print("<< ============== JOB (PID = %d) %s ============== >>" %
(PID, args.save_dir))
prepare_seed(args.rand_seed)
if args.timestamp == 'none':
args.timestamp = "{:}".format(
time.strftime('%h-%d-%C_%H-%M-%s', time.gmtime(time.time())))
torch.set_num_threads(1)
# Log outputs
args.save_dir = args.save_dir + \
"/Visda17-L2O.train.Res101-%s-train.%s-LR%.2E-epoch%d-batch%d-seed%d"%(
"LWF" if args.lwf > 0 else "XE", args.train_blocks, args.lr, args.epochs, args.batch_size, args.rand_seed) + \
"%s/%s"%('/'+args.resume if args.resume != 'none' else '', args.timestamp)
logger = prepare_logger(args)
best_prec1 = 0
#### preparation ###########################################
data_transforms = {
'train':
transforms.Compose([
transforms.Resize(224),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
'val':
transforms.Compose([
transforms.Resize(224),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
}
kwargs = {'num_workers': 20, 'pin_memory': True}
trainset = VisDA17(txt_file=os.path.join(args.data,
"train/image_list.txt"),
root_dir=os.path.join(args.data, "train"),
transform=data_transforms['train'])
valset = VisDA17(txt_file=os.path.join(args.data,
"validation/image_list.txt"),
root_dir=os.path.join(args.data, "validation"),
transform=data_transforms['val'],
label_one_hot=True)
train_loader = DataLoader(trainset,
batch_size=args.batch_size,
shuffle=True,
**kwargs)
val_loader = DataLoader(valset,
batch_size=args.batch_size,
shuffle=False,
**kwargs)
train_loader_iter = iter(train_loader)
current_optimizee_step, prev_optimizee_step = 0, 0
model_old = None
if args.lwf > 0:
# create a fixed model copy for Life-long learning
model_old = resnet101(pretrained=True)
for param in model_old.parameters():
param.requires_grad = False
model_old.eval()
model_old.cuda()
############################################################
### Agent Settings ########################################
RANDOM = False # False | True | 'init'
action_space = np.arange(0, 1.1, 0.1)
obs_avg = True
_window_size = 1
window_size = 1 if obs_avg else _window_size
window_shrink_size = 20 # larger: controller will be updated more frequently
sgd_in_names = [
"conv1", "bn1", "layer1", "layer2", "layer3", "layer4", "fc_new"
]
coord_size = len(sgd_in_names)
ob_name_lstm = ["loss", "loss_kl", "step", "fc_mean", "fc_std"]
ob_name_scalar = []
obs_shape = (len(ob_name_lstm) * window_size + len(ob_name_scalar) +
coord_size, )
_hidden_size = 20
hidden_size = _hidden_size * len(ob_name_lstm)
actor_critic = Policy(coord_size,
input_size=(len(ob_name_lstm), len(ob_name_scalar)),
action_space=len(action_space),
hidden_size=_hidden_size,
window_size=window_size)
actor_critic.cuda()
actor_critic.eval()
partial = torch.load(args.agent_load_dir,
map_location=lambda storage, loc: storage)
state = actor_critic.state_dict()
pretrained_dict = {k: v for k, v in partial.items()}
state.update(pretrained_dict)
actor_critic.load_state_dict(state)
################################################################
_min_iter = 10
# reset optmizee
model, optimizer, current_optimizee_step, prev_optimizee_step = prepare_optimizee(
args, sgd_in_names, obs_shape, hidden_size, actor_critic,
current_optimizee_step, prev_optimizee_step)
epoch_size = len(train_loader)
total_steps = epoch_size * args.epochs
bar_format = '{desc}[{elapsed}<{remaining},{rate_fmt}]'
pbar = tqdm(range(int(epoch_size * args.epochs)),
file=sys.stdout,
bar_format=bar_format,
ncols=100)
_window_size = max(
_min_iter,
current_optimizee_step + prev_optimizee_step // window_shrink_size)
train_loader_iter, obs, loss, loss_kl, fc_mean, fc_std = train_step(
args,
_window_size,
train_loader_iter,
train_loader,
model,
optimizer,
obs_avg,
args.lr,
pbar,
current_optimizee_step + prev_optimizee_step,
total_steps,
model_old=model_old)
logger.writer.add_scalar("loss/ce", loss,
current_optimizee_step + prev_optimizee_step)
logger.writer.add_scalar("loss/kl", loss_kl,
current_optimizee_step + prev_optimizee_step)
logger.writer.add_scalar("loss/total", loss + loss_kl,
current_optimizee_step + prev_optimizee_step)
logger.writer.add_scalar("fc/mean", fc_mean,
current_optimizee_step + prev_optimizee_step)
logger.writer.add_scalar("fc/std", fc_std,
current_optimizee_step + prev_optimizee_step)
current_optimizee_step += _window_size
pbar.update(_window_size)
prev_obs = obs.unsqueeze(0)
prev_hidden = torch.zeros(actor_critic.net.num_recurrent_layers, 1,
hidden_size).cuda()
for epoch in range(args.epochs):
print("\n===== Epoch %d / %d =====" % (epoch + 1, args.epochs))
print("<< ============== JOB (PID = %d) %s ============== >>" %
(PID, args.save_dir))
while current_optimizee_step < epoch_size:
# Sample actions
with torch.no_grad():
if not RANDOM:
value, action, action_log_prob, recurrent_hidden_states, distribution = actor_critic.act(
prev_obs, prev_hidden, deterministic=False)
action = action.squeeze()
action_log_prob = action_log_prob.squeeze()
value = value.squeeze()
for idx in range(len(action)):
logger.writer.add_scalar(
"action/%s" % sgd_in_names[idx], action[idx],
current_optimizee_step + prev_optimizee_step)
logger.writer.add_scalar(
"entropy/%s" % sgd_in_names[idx],
distribution.distributions[idx].entropy(),
current_optimizee_step + prev_optimizee_step)
optimizer.param_groups[idx]['lr'] = float(
action_space[action[idx]]) * args.lr
logger.writer.add_scalar(
"LR/%s" % sgd_in_names[idx],
optimizer.param_groups[idx]['lr'],
current_optimizee_step + prev_optimizee_step)
else:
if RANDOM is True or RANDOM == 'init':
for idx in range(coord_size):
optimizer.param_groups[idx]['lr'] = float(
choice(action_space)) * args.lr
if RANDOM == 'init':
RANDOM = 'done'
for idx in range(coord_size):
logger.writer.add_scalar(
"LR/%s" % sgd_in_names[idx],
optimizer.param_groups[idx]['lr'],
current_optimizee_step + prev_optimizee_step)
# Obser reward and next obs
_window_size = max(
_min_iter, current_optimizee_step +
prev_optimizee_step // window_shrink_size)
_window_size = min(_window_size,
epoch_size - current_optimizee_step)
train_loader_iter, obs, loss, loss_kl, fc_mean, fc_std = train_step(
args,
_window_size,
train_loader_iter,
train_loader,
model,
optimizer,
obs_avg,
args.lr,
pbar,
current_optimizee_step + prev_optimizee_step,
total_steps,
model_old=model_old)
logger.writer.add_scalar(
"loss/ce", loss, current_optimizee_step + prev_optimizee_step)
logger.writer.add_scalar(
"loss/kl", loss_kl,
current_optimizee_step + prev_optimizee_step)
logger.writer.add_scalar(
"loss/total", loss + loss_kl,
current_optimizee_step + prev_optimizee_step)
logger.writer.add_scalar(
"fc/mean", fc_mean,
current_optimizee_step + prev_optimizee_step)
logger.writer.add_scalar(
"fc/std", fc_std, current_optimizee_step + prev_optimizee_step)
current_optimizee_step += _window_size
pbar.update(_window_size)
prev_obs = obs.unsqueeze(0)
if not RANDOM: prev_hidden = recurrent_hidden_states
prev_optimizee_step += current_optimizee_step
current_optimizee_step = 0
# evaluate on validation set
prec1 = validate(val_loader, model, args)
logger.writer.add_scalar("prec", prec1, epoch)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
args.save_dir, {
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
}, is_best)
logging.info('Best accuracy: {prec1:.3f}'.format(prec1=best_prec1))
transform_test = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
testset = CustomData('./data/test1',
transform=transform_test,
train=False,
test=True)
testloader = torch.utils.data.DataLoader(testset,
batch_size=opt.batchSize,
shuffle=False,
num_workers=opt.num_workers)
model = resnet101(pretrained=True)
model.fc = nn.Linear(2048, 2)
model.load_state_dict(torch.load('ckp/model.pth'))
model.cuda()
model.eval()
results = []
with torch.no_grad():
for image, label in testloader:
image = image.cuda()
out = model(image)
label = label.numpy().tolist()
_, predicted = torch.max(out.data, 1)
predicted = predicted.data.cpu().numpy().tolist()
results.extend([[i, ";".join(str(j))]
for (i, j) in zip(label, predicted)])
'./data/celebA/list_attr_celeba.txt', '0',
'./data/celebA/img_align_celeba/', transform_train)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=opt.batchSize,
shuffle=True,
num_workers=opt.workers)
valset = celeba('./data/celebA/list_eval_partition.txt',
'./data/celebA/list_attr_celeba.txt', '1',
'./data/celebA/img_align_celeba/', transform_val)
valloader = torch.utils.data.DataLoader(valset,
batch_size=opt.batchSize,
shuffle=True,
num_workers=opt.workers)
resnet = resnet101(pretrained=True, num_classes=40)
resnet.cuda()
criterion = nn.MSELoss(reduce=True)
optimizer = optim.SGD(resnet.parameters(),
lr=opt.lr,
momentum=0.9,
weight_decay=5e-4)
scheduler = StepLR(optimizer, step_size=3)
def train(epoch):
print('\nEpoch: %d' % epoch)
scheduler.step()
resnet.train()
for batch_idx, (images, attrs) in enumerate(trainloader):
images = Variable(images.cuda())