def __init__(self, layers=18, classes=2, with_sp=True):
super(BiseNet, self).__init__()
self.with_sp = with_sp
if layers == 18:
resnet = models.resnet18(pretrained=True,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
elif layers == 34:
resnet = models.resnet34(pretrained=True,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
if self.with_sp:
self.sp = SpatialPath(in_channels=3, out_channels=128)
self.cp = ContextPath(in_channels=3, out_channels=128, backbone=resnet)
self.ffm = FeatureFusionModule(in_channels=256,
out_channels=256) # concat: 128+128
self.conv_out = BiseNetHead(in_channels=256,
mid_channels=256,
classes=classes)
if self.training:
self.conv_out16 = BiseNetHead(in_channels=128,
mid_channels=64,
classes=classes)
self.conv_out32 = BiseNetHead(in_channels=128,
mid_channels=64,
classes=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_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 __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 __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, layers=18, dropout=0.1, classes=2):
super(FFTNet23, self).__init__()
if layers == 18:
resnet = models.resnet18(pretrained=True,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
elif layers == 34:
resnet = models.resnet34(pretrained=True,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
elif layers == 50:
resnet = models.resnet50_semseg(pretrained=True,
deep_base=True,
strides=(1, 2, 1, 1),
dilations=(1, 1, 2, 4))
if layers == 18 or layers == 34:
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu,
resnet.maxpool)
else:
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
if layers == 18 or layers == 34:
fea_dim = 512
aux_dim = 256
else:
fea_dim = 2048
aux_dim = 1024
self.freq = nn.ModuleList()
for i in range(6, 10): # the number of in_channels is 2^i
self.freq.append(
FeatureFrequencySeparationModule(
in_channels=2**i,
up_channels=2**i if i == 6 else 2**(i - 1),
smf_channels=128,
high_ratio=1 - 0.2 * (i - 5),
# high_ratio=0.5,
low_ratio=0.2,
up_flag=False if i == 6 else True,
smf_flag=True if i % 2 == 0 else False,
))
self.fa_cls_seg = nn.Sequential(nn.Dropout2d(p=dropout),
nn.Conv2d(256, classes, kernel_size=1))
if self.training:
self.aux = nn.Sequential(
nn.Conv2d(aux_dim,
aux_dim // 4,
kernel_size=3,
padding=1,
bias=False), nn.BatchNorm2d(aux_dim // 4),
nn.ReLU(inplace=True), nn.Dropout2d(p=dropout),
nn.Conv2d(aux_dim // 4, classes, kernel_size=1))
def __init__(self,
out_planes=1,
ccm=True,
norm_layer=nn.BatchNorm2d,
is_training=True,
expansion=2,
base_channel=32):
super(CPFNet, self).__init__()
self.backbone = resnet34(pretrained=True)
self.expansion = expansion
self.base_channel = base_channel
if self.expansion == 4 and self.base_channel == 64:
expan = [512, 1024, 2048]
spatial_ch = [128, 256]
elif self.expansion == 4 and self.base_channel == 32:
expan = [256, 512, 1024]
spatial_ch = [32, 128]
conv_channel_up = [256, 384, 512]
elif self.expansion == 2 and self.base_channel == 32:
expan = [128, 256, 512]
spatial_ch = [64, 64]
conv_channel_up = [128, 256, 512]
conv_channel = expan[0]
self.is_training = is_training
self.sap = SAPblock(expan[-1])
self.decoder5 = DecoderBlock(expan[-1],
expan[-2],
relu=False,
last=True) #256
self.decoder4 = DecoderBlock(expan[-2], expan[-3], relu=False) #128
self.decoder3 = DecoderBlock(expan[-3], spatial_ch[-1],
relu=False) #64
self.decoder2 = DecoderBlock(spatial_ch[-1], spatial_ch[-2]) #32
self.mce_2 = GPG_2([spatial_ch[-1], expan[0], expan[1], expan[2]],
width=spatial_ch[-1],
up_kwargs=up_kwargs)
self.mce_3 = GPG_3([expan[0], expan[1], expan[2]],
width=expan[0],
up_kwargs=up_kwargs)
self.mce_4 = GPG_4([expan[1], expan[2]],
width=expan[1],
up_kwargs=up_kwargs)
self.main_head = BaseNetHead(spatial_ch[0],
out_planes,
2,
is_aux=False,
norm_layer=norm_layer)
self.relu = nn.ReLU()
def __init__(self, classNum, pretrained=True):
super(MA_quality_check_res34, self).__init__()
self.base = resnet.resnet34(pretrained=pretrained)
self.num_att = classNum
# print ((self.base))
# exit()
self.classifier = nn.Linear(512, self.num_att)
init.normal(self.classifier.weight, std=0.001)
init.constant(self.classifier.bias, 0)
def __init__(self, layers=50, dropout=0.1, classes=2, use_dct=True, use_bise=True, vec_dim=300):
super(DCTNet, self).__init__()
assert layers in [18, 34, 50, 101]
assert classes > 1
self.use_dct = use_dct
self.use_bise = use_bise
self.vec_dim = vec_dim
if layers == 18:
resnet = models.resnet18(pretrained=False, deep_base=False, strides=(1, 2, 2, 2), dilations=(1, 1, 1, 1))
elif layers == 34:
resnet = models.resnet34(pretrained=True, deep_base=False, strides=(1, 2, 2, 2), dilations=(1, 1, 1, 1))
elif layers == 50:
resnet = models.resnet50_semseg(pretrained=True, deep_base=True, strides=(1, 2, 1, 1), dilations=(1, 1, 2, 4))
elif layers == 101:
resnet = models.resnet101_semseg(pretrained=True, deep_base=True, strides=(1, 2, 1, 1), dilations=(1, 1, 2, 4))
if layers == 18 or layers == 34:
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool)
else:
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
if layers == 18 or layers == 34:
fea_dim = 512
aux_dim = 256
else:
fea_dim = 2048
aux_dim = 1024
down_dim = fea_dim // 4
if use_dct:
self.dct_encoding = DCTModule(vec_dim=self.vec_dim)
if use_bise:
self.ffm = FeatureFusionModule(
in_channels=self.vec_dim + 128, out_channels=fea_dim) # concat: 128+128
self.cls = nn.Sequential(
nn.Conv2d(fea_dim, down_dim, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(down_dim),
nn.ReLU(inplace=True),
nn.Dropout2d(p=dropout),
nn.Conv2d(down_dim, classes, kernel_size=1)
)
if self.training:
self.aux = nn.Sequential(
nn.Conv2d(aux_dim, aux_dim // 4, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(aux_dim // 4),
nn.ReLU(inplace=True),
nn.Dropout2d(p=dropout),
nn.Conv2d(aux_dim // 4, classes, kernel_size=1)
)
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=2,
use_dct=True,
vec_dim=300):
super(DCTNet, self).__init__()
assert layers in [18, 34, 50, 101]
assert classes > 1
self.use_dct = use_dct
self.vec_dim = vec_dim
if layers == 18:
resnet = models.resnet18(pretrained=False,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
elif layers == 34:
resnet = models.resnet34(pretrained=True,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
if layers == 18 or layers == 34:
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu,
resnet.maxpool)
else:
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 = resnet.layer1, resnet.layer2, resnet.layer3
if layers == 18 or layers == 34:
fea_dim = 256
down_dim = fea_dim // 4
if use_dct:
self.dct_encoding = DCTModule(vec_dim=self.vec_dim)
self.up_conv = ConvBNReLU(fea_dim,
fea_dim,
ks=1,
stride=1,
padding=0)
self.cls = nn.Sequential(
nn.Conv2d(fea_dim, down_dim, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(down_dim), nn.ReLU(inplace=True),
nn.Dropout2d(p=dropout), nn.Conv2d(down_dim,
classes,
kernel_size=1))
def main():
pth_path = '/home/zhaoliu/car_brand/car_mid/results_perspective/per_1/save_50.pth'
# test_result=Path('/home/zhaoliu/car_class+ori/test_result')
# test_path = '/home/zhaoliu/car_data/训练数据/4.9新加测试集/val_lmdb'
# keys_path = '/home/zhaoliu/car_data/训练数据/4.9新加测试集/new_val.npy'
test_path = "/mnt/disk/zhaoliu_data/perspective/lmdb/test"
keys_path = '/home/zhaoliu/car_brand/perspective_data/prespective_test.npy'
model = resnet34(num_classes=1189)
model = resume_model(pth_path, model)
test_loader = get_test_utils(test_path, keys_path)
print('数据加载完毕...')
test(model, test_loader)
def main():
pth_path = '/home/zhaoliu/car_brand/car_mid/results_perspective/per_1/save_50.pth'
test_result = '/home/zhaoliu/car_brand/car_mid/badcase_per/'
test_path = '/mnt/disk/zhaoliu_data/perspective/lmdb/test'
keys_path = '/home/zhaoliu/car_brand/perspective_data/prespective_test.npy'
midnpy_path = test_result + 'mid_badkeys.npy'
midtxt_path = test_result + 'mid_badcase_info.txt'
model = resnet34(num_classes=1189)
model = resume_model(pth_path, model)
test_loader = get_test_utils(test_path, keys_path)
print('数据加载完毕...')
test(model, test_loader, midnpy_path, midtxt_path)
def __init__(self, layers=50, dropout=0.1, classes=2, fuse=8):
super(TriSeNet1, self).__init__()
assert layers in [18, 34, 50, 101]
assert classes > 1
self.fuse = fuse
# Backbone
if layers == 18:
resnet = models.resnet18(pretrained=False, deep_base=False, strides=(1, 2, 2, 2), dilations=(1, 1, 1, 1))
elif layers == 34:
resnet = models.resnet34(pretrained=True, deep_base=False, strides=(1, 2, 2, 2), dilations=(1, 1, 1, 1))
elif layers == 50:
resnet = models.resnet50_semseg(pretrained=True, deep_base=True, strides=(1, 2, 1, 1), dilations=(1, 1, 2, 4))
elif layers == 101:
resnet = models.resnet101_semseg(pretrained=True, deep_base=True, strides=(1, 2, 1, 1), dilations=(1, 1, 2, 4))
if layers == 18 or layers == 34:
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool)
else:
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
if layers == 18 or layers == 34:
fea_dim = 512
aux_dim = 256
else:
fea_dim = 2048
aux_dim = 1024
down_dim = fea_dim // 4
self.cls = nn.Sequential(
nn.Conv2d(fea_dim, down_dim, kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(down_dim),
nn.ReLU(inplace=True),
nn.Dropout2d(p=dropout),
nn.Conv2d(down_dim, classes, kernel_size=1)
)
if self.fuse == 16 or self.fuse == 8:
self.fuse_16 = nn.Conv2d(fea_dim//2, classes, kernel_size=1)
if self.fuse == 8:
self.fuse_8 = nn.Conv2d(fea_dim//4, classes, kernel_size=1)
def __init__(self, layers=18, classes=2):
super(TriSeNet, self).__init__()
if layers == 18:
backbone = models.resnet18(pretrained=True,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
elif layers == 34:
backbone = models.resnet34(pretrained=True,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
# the initial layer conv is 7x7, instead of three 3x3
self.layer0 = nn.Sequential(backbone.conv1, backbone.bn1,
backbone.relu, backbone.maxpool)
# stage channels for resnet18 and resnet34 are:(64, 128, 256, 512)
self.layer1, self.layer2, self.layer3, self.layer4 \
= backbone.layer1, backbone.layer2, backbone.layer3, backbone.layer4
# self.gap = nn.AdaptiveAvgPool2d(1) # Global Average Pooling
# self.up_16_8 = UpModule(in_channels=256, out_channels=128, up_scale=2) # feat_16 up to the size of feat_8
# self.up_32_8 = UpModule(in_channels=512, out_channels=128, up_scale=4)
# self.down_8_16 = DownModule(in_channels=128, out_channels=256, down_scale=2) # feat_8 down to the size of feat_16
# self.up_32_16 = UpModule(in_channels=512, out_channels=256, up_scale=2)
self.down_8_32 = DownModule(in_channels=128,
out_channels=512,
down_scale=4)
# self.down_16_32 = DownModule(in_channels=256, out_channels=512, down_scale=2)
self.relu = nn.ReLU(inplace=True)
self.sa_8_32 = SelfAttentionBlock(512)
# self.ca_32_8 = ChannelAttentionModule(in_channels=128, reduction=4)
# self.cp = ContextPath(in_channels=3, out_channels=128, backbone=resnet)
self.seg_head = SegHead(in_channels=640,
mid_channels=256,
classes=classes)
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 __init__(self,
in_channels=3,
n_classes=1,
feature_scale=2,
is_deconv=True,
is_batchnorm=True):
super(UNet, self).__init__()
self.backbone = resnet34(pretrained=True)
self.in_channels = in_channels
self.feature_scale = feature_scale
self.is_deconv = is_deconv
self.is_batchnorm = is_batchnorm
filters = [64, 128, 256, 512, 1024]
filters = [int(x / self.feature_scale) for x in filters]
# downsampling
self.maxpool = nn.MaxPool2d(kernel_size=2)
self.conv1 = unetConv2(self.in_channels, filters[0], self.is_batchnorm)
self.conv2 = unetConv2(filters[0], filters[1], self.is_batchnorm)
self.conv3 = unetConv2(filters[1], filters[2], self.is_batchnorm)
self.conv4 = unetConv2(filters[2], filters[3], self.is_batchnorm)
self.center = unetConv2(filters[3], filters[4], self.is_batchnorm)
# upsampling
self.up_concat4 = unetUp(filters[4], filters[3], self.is_deconv)
self.up_concat3 = unetUp(filters[3], filters[2], self.is_deconv)
self.up_concat2 = unetUp(filters[2], filters[1], self.is_deconv)
self.up_concat1 = unetUp(filters[1], filters[0], self.is_deconv)
# final conv (without any concat)
self.final = nn.Conv2d(filters[0], n_classes, 1)
# initialise weights
for m in self.modules():
if isinstance(m, nn.Conv2d):
init_weights(m, init_type='kaiming')
elif isinstance(m, nn.BatchNorm2d):
init_weights(m, init_type='kaiming')
def main():
# Training settings
parser = argparse.ArgumentParser(description='SSDA Classification')
parser.add_argument('--steps',
type=int,
default=50000,
metavar='N',
help='maximum number of iterations '
'to train (default: 50000)')
parser.add_argument(
'--method',
type=str,
default='MME',
choices=['S+T', 'ENT', 'MME'],
help='MME is proposed method, ENT is entropy minimization,'
' S+T is training only on labeled examples')
parser.add_argument('--lr',
type=float,
default=0.01,
metavar='LR',
help='learning rate (default: 0.001)')
parser.add_argument('--multi',
type=float,
default=0.1,
metavar='MLT',
help='learning rate multiplication')
parser.add_argument('--T',
type=float,
default=0.05,
metavar='T',
help='temperature (default: 0.05)')
parser.add_argument('--lamda',
type=float,
default=0.1,
metavar='LAM',
help='value of lamda')
parser.add_argument('--save_check',
action='store_true',
default=False,
help='save checkpoint or not')
parser.add_argument('--checkpath',
type=str,
default='./save_model_ssda',
help='dir to save checkpoint')
parser.add_argument('--seed',
type=int,
default=1,
metavar='S',
help='random seed (default: 1)')
parser.add_argument('--log-interval',
type=int,
default=100,
metavar='N',
help='how many batches to wait before logging '
'training status')
parser.add_argument('--save_interval',
type=int,
default=500,
metavar='N',
help='how many batches to wait before saving a model')
parser.add_argument('--net',
type=str,
default='al exnet',
help='which network to use')
parser.add_argument('--source',
type=str,
default='real',
help='source domain')
parser.add_argument('--target',
type=str,
default='sketch',
help='target domain')
parser.add_argument('--dataset',
type=str,
default='multi',
help='the name of dataset')
parser.add_argument('--num',
type=int,
default=3,
help='number of labeled examples in the target')
parser.add_argument('--patience',
type=int,
default=5,
metavar='S',
help='early stopping to wait for improvment '
'before terminating. (default: 5 (5000 iterations))')
parser.add_argument('--early',
action='store_false',
default=True,
help='early stopping on validation or not')
parser.add_argument('--loss',
type=str,
default='CE',
choices=['CE', 'FL', 'CBFL'],
help='classifier loss function')
parser.add_argument('--beta',
type=float,
default=0.99,
required=False,
help='beta value in CBFL loss')
parser.add_argument('--gamma',
type=float,
default=0.5,
required=False,
help='gamma value in CBFL or FL')
parser.add_argument('--reg',
type=float,
default=0.1,
required=False,
help='weight of semantic regularizer')
parser.add_argument('--attribute',
type=str,
default=None,
help='semantic attribute feature vector to be used')
parser.add_argument(
'--dim',
type=int,
default=50,
help=
'dimensionality of the feature vector - make sure this in sync with the dim of the semantic attribute vector'
)
parser.add_argument('--mode',
type=str,
default='train',
choices=['train', 'infer'],
help='mode of script train or infer')
# this argument is valid only if the mode is infer
parser.add_argument('--model_path',
type=str,
help='path to the checkpoint of the model')
parser.add_argument(
'--uda',
type=int,
default=0,
help='unsupervised domain adaptation or not - 0 for ssda and 1 for uda'
)
args = parser.parse_args()
print('Dataset %s Source %s Target %s Labeled num perclass %s Network %s' %
(args.dataset, args.source, args.target, args.num, args.net))
source_loader, target_loader, target_loader_unl, target_loader_val, \
target_loader_test, class_num_list, class_list = return_dataset(args) # class num list is returned for CBFL
use_gpu = torch.cuda.is_available()
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))
if use_gpu:
device = 'cuda'
else:
device = 'cpu'
print("Device: %s Loss: %s Attributes: %s" %
(device, args.loss, args.attribute))
if use_gpu:
torch.cuda.manual_seed(args.seed)
else:
torch.manual_seed(args.seed)
if args.net == 'resnet34':
G = resnet34()
inc = 512
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 += [{
'params': [value],
'lr': 0.1,
'weight_decay': 0.0005
}]
else:
params += [{
'params': [value],
'lr': 1,
'weight_decay': 0.0005
}]
# Setting the predictor layer
if args.attribute is not None:
if args.net == 'resnet34':
F1 = Predictor_deep_attributes(num_class=len(class_list),
inc=inc,
feat_dim=args.dim)
print("Using: Predictor_deep_attributes")
else:
F1 = Predictor_attributes(num_class=len(class_list),
inc=inc,
feat_dim=args.dim)
print("Using: Predictor_attributes")
else:
if args.net == 'resnet34':
F1 = Predictor_deep(num_class=len(class_list), inc=inc)
print("Using: Predictor_deep")
else:
F1 = Predictor(num_class=len(class_list), inc=inc, temp=args.T)
print("Using: Predictor")
# Initializing the weights of the prediction layer
weights_init(F1)
# Setting the prediction layer weights as the semantic attributes
if args.attribute is not None:
att = np.load('attributes/%s_%s.npy' % (args.dataset, args.attribute))
#att = np.load('attributes/multi_%s.npy'%(args.attribute))
if use_gpu:
att = nn.Parameter(torch.cuda.FloatTensor(att))
else:
att = nn.Parameter(torch.FloatTensor(att, device="cpu"))
if args.net == 'resnet34':
F1.fc3.weight = att
else:
F1.fc2.weight = att
print("attribute shape is: ", att.shape)
lr = args.lr
# If the mode is inference then load the pretrained network
if args.mode == 'infer':
# loading the model checkpoint
main_dict = torch.load(args.model_path)
G.load_state_dict(main_dict['G_state_dict'])
F1.load_state_dict(main_dict['F_state_dict'])
print("Loaded pretrained model weights")
G.to(device)
F1.to(device)
if args.uda == 1:
print("Using: Unsupervised domain adaptation")
im_data_s = torch.FloatTensor(1)
im_data_t = torch.FloatTensor(1)
im_data_tu = torch.FloatTensor(1)
gt_labels_t = torch.LongTensor(1)
gt_labels_s = torch.LongTensor(1)
sample_labels_t = torch.LongTensor(1)
sample_labels_s = torch.LongTensor(1)
im_data_s = im_data_s.to(device)
im_data_t = im_data_t.to(device)
im_data_tu = im_data_tu.to(device)
gt_labels_s = gt_labels_s.to(device)
gt_labels_t = gt_labels_t.to(device)
sample_labels_t = sample_labels_t.to(device)
sample_labels_s = sample_labels_s.to(device)
im_data_s = Variable(im_data_s)
im_data_t = Variable(im_data_t)
im_data_tu = Variable(im_data_tu)
gt_labels_s = Variable(gt_labels_s)
gt_labels_t = Variable(gt_labels_t)
sample_labels_t = Variable(sample_labels_t)
sample_labels_s = Variable(sample_labels_s)
if os.path.exists(args.checkpath) == False:
os.mkdir(args.checkpath)
time_stamp = datetime.now()
print(time_stamp)
def train(class_dist_threshold_list):
G.train()
F1.train()
optimizer_g = optim.SGD(params,
momentum=0.9,
weight_decay=0.0005,
nesterov=True)
optimizer_f = optim.SGD(list(F1.parameters()),
lr=1.0,
momentum=0.9,
weight_decay=0.0005,
nesterov=True)
def zero_grad_all():
optimizer_g.zero_grad()
optimizer_f.zero_grad()
param_lr_g = []
for param_group in optimizer_g.param_groups:
param_lr_g.append(param_group["lr"])
param_lr_f = []
for param_group in optimizer_f.param_groups:
param_lr_f.append(param_group["lr"])
# Setting the loss function to be used for the classification loss
if args.loss == 'CE':
criterion = nn.CrossEntropyLoss().to(device)
if args.loss == 'FL':
criterion = FocalLoss(alpha=1, gamma=args.gamma).to(device)
if args.loss == 'CBFL':
# Calculating the list having the number of examples per class which is going to be used in the CB focal loss
beta = args.beta
effective_num = 1.0 - np.power(beta, class_num_list)
per_cls_weights = (1.0 - beta) / np.array(effective_num)
per_cls_weights = per_cls_weights / np.sum(per_cls_weights) * len(
class_num_list)
per_cls_weights = torch.FloatTensor(per_cls_weights).to(device)
criterion = CBFocalLoss(weight=per_cls_weights,
gamma=args.gamma).to(device)
all_step = args.steps
data_iter_s = iter(source_loader)
data_iter_t = iter(target_loader)
data_iter_t_unl = iter(target_loader_unl)
len_train_source = len(source_loader)
len_train_target = len(target_loader)
len_train_target_semi = len(target_loader_unl)
best_acc = 0
counter = 0
"""
x = torch.load("./freezed_models/alexnet_p2r.ckpt.best.pth.tar")
G.load_state_dict(x['G_state_dict'])
F1.load_state_dict(x['F1_state_dict'])
optimizer_f.load_state_dict(x['optimizer_f'])
optimizer_g.load_state_dict(x['optimizer_g'])
"""
reg_weight = args.reg
for step in range(all_step):
optimizer_g = inv_lr_scheduler(param_lr_g,
optimizer_g,
step,
init_lr=args.lr)
optimizer_f = inv_lr_scheduler(param_lr_f,
optimizer_f,
step,
init_lr=args.lr)
lr = optimizer_f.param_groups[0]['lr']
# condition for restarting the iteration for each of the data loaders
if step % len_train_target == 0:
data_iter_t = iter(target_loader)
if step % len_train_target_semi == 0:
data_iter_t_unl = iter(target_loader_unl)
if step % len_train_source == 0:
data_iter_s = iter(source_loader)
data_t = next(data_iter_t)
data_t_unl = next(data_iter_t_unl)
data_s = next(data_iter_s)
with torch.no_grad():
im_data_s.resize_(data_s[0].size()).copy_(data_s[0])
gt_labels_s.resize_(data_s[1].size()).copy_(data_s[1])
im_data_t.resize_(data_t[0].size()).copy_(data_t[0])
gt_labels_t.resize_(data_t[1].size()).copy_(data_t[1])
im_data_tu.resize_(data_t_unl[0].size()).copy_(data_t_unl[0])
zero_grad_all()
if args.uda == 1:
data = im_data_s
target = gt_labels_s
else:
data = torch.cat((im_data_s, im_data_t), 0)
target = torch.cat((gt_labels_s, gt_labels_t), 0)
#print(data.shape)
output = G(data)
out1 = F1(output)
if args.attribute is not None:
if args.net == 'resnet34':
reg_loss = regularizer(F1.fc3.weight, att)
loss = criterion(out1, target) + reg_weight * reg_loss
else:
reg_loss = regularizer(F1.fc2.weight, att)
loss = criterion(out1, target) + reg_weight * reg_loss
else:
reg_loss = torch.tensor(0)
loss = criterion(out1, target)
if args.attribute is not None:
if step % args.save_interval == 0 and step != 0:
reg_weight = 0.5 * reg_weight
print("Reduced Reg weight to: ", reg_weight)
loss.backward(retain_graph=True)
optimizer_g.step()
optimizer_f.step()
zero_grad_all()
if not args.method == 'S+T':
output = G(im_data_tu)
if args.method == 'ENT':
loss_t = entropy(F1, output, args.lamda)
#print(loss_t.cpu().data.item())
loss_t.backward()
optimizer_f.step()
optimizer_g.step()
elif args.method == 'MME':
loss_t = adentropy(F1, output, args.lamda,
class_dist_threshold_list)
loss_t.backward()
optimizer_f.step()
optimizer_g.step()
else:
raise ValueError('Method cannot be recognized.')
log_train = 'S {} T {} Train Ep: {} lr{} \t ' \
'Loss Classification: {:.6f} Reg: {:.6f} Loss T {:.6f} ' \
'Method {}\n'.format(args.source, args.target,
step, lr, loss.data, reg_weight*reg_loss.data,
-loss_t.data, args.method)
else:
log_train = 'S {} T {} Train Ep: {} lr{} \t ' \
'Loss Classification: {:.6f} Reg: {:.6f} Method {}\n'.\
format(args.source, args.target,
step, lr, loss.data, reg_weight * reg_loss.data,
args.method)
G.zero_grad()
F1.zero_grad()
zero_grad_all()
if step % args.log_interval == 0:
print(log_train)
if step % args.save_interval == 0 and step > 0:
loss_val, acc_val = test(target_loader_val)
loss_test, acc_test = test(target_loader_test)
G.train()
F1.train()
if acc_val >= best_acc:
best_acc = acc_val
best_acc_test = acc_test
counter = 0
else:
counter += 1
if args.early:
if counter > args.patience:
break
print('best acc test %f best acc val %f' %
(best_acc_test, acc_val))
print('record %s' % record_file)
with open(record_file, 'a') as f:
f.write('step %d best %f final %f \n' %
(step, best_acc_test, acc_val))
G.train()
F1.train()
#saving model as a checkpoint dict having many things
if args.save_check:
print('saving model')
is_best = True if counter == 0 else False
save_mymodel(
args, {
'step': step,
'arch': args.net,
'G_state_dict': G.state_dict(),
'F1_state_dict': F1.state_dict(),
'best_acc_test': best_acc_test,
'optimizer_g': optimizer_g.state_dict(),
'optimizer_f': optimizer_f.state_dict(),
}, is_best, time_stamp)
# defining the function for in training validation and testing
def test(loader):
G.eval()
F1.eval()
test_loss = 0
correct = 0
size = 0
num_class = len(class_list)
output_all = np.zeros((0, num_class))
# Setting the loss function to be used for the classification loss
if args.loss == 'CE':
criterion = nn.CrossEntropyLoss().to(device)
if args.loss == 'FL':
criterion = FocalLoss(alpha=1, gamma=args.gamma).to(device)
if args.loss == 'CBFL':
# Calculating the list having the number of examples per class which is going to be used in the CB focal loss
beta = args.beta
effective_num = 1.0 - np.power(beta, class_num_list)
per_cls_weights = (1.0 - beta) / np.array(effective_num)
per_cls_weights = per_cls_weights / np.sum(per_cls_weights) * len(
class_num_list)
per_cls_weights = torch.FloatTensor(per_cls_weights).to(device)
criterion = CBFocalLoss(weight=per_cls_weights,
gamma=args.gamma).to(device)
confusion_matrix = torch.zeros(num_class, num_class)
with torch.no_grad():
for batch_idx, data_t in enumerate(loader):
im_data_t.data.resize_(data_t[0].size()).copy_(data_t[0])
gt_labels_t.data.resize_(data_t[1].size()).copy_(data_t[1])
feat = G(im_data_t)
output1 = F1(feat)
output_all = np.r_[output_all, output1.data.cpu().numpy()]
size += im_data_t.size(0)
pred1 = output1.data.max(1)[1]
for t, p in zip(gt_labels_t.view(-1), pred1.view(-1)):
confusion_matrix[t.long(), p.long()] += 1
correct += pred1.eq(gt_labels_t.data).cpu().sum()
test_loss += criterion(output1, gt_labels_t) / len(loader)
np.save("cf_target.npy", confusion_matrix)
#print(confusion_matrix)
print('\nTest set: Average loss: {:.4f}, '
'Accuracy: {}/{} F1 ({:.0f}%)\n'.format(test_loss, correct, size,
100. * correct / size))
return test_loss.data, 100. * float(correct) / size
# defining the function for inference which is similar to the testing function as above but with some additional functionality for calculating the distances between the class prototypes and the predicted testing samples
def infer(loader):
G.eval()
F1.eval()
test_loss = 0
correct = 0
size = 0
num_class = len(class_list)
output_all = np.zeros((0, num_class))
# Setting the loss function to be used for the classification loss
if args.loss == 'CE':
criterion = nn.CrossEntropyLoss().to(device)
if args.loss == 'FL':
criterion = FocalLoss(alpha=1, gamma=1).to(device)
if args.loss == 'CBFL':
# Calculating the list having the number of examples per class which is going to be used in the CB focal loss
beta = 0.99
effective_num = 1.0 - np.power(beta, class_num_list)
per_cls_weights = (1.0 - beta) / np.array(effective_num)
per_cls_weights = per_cls_weights / np.sum(per_cls_weights) * len(
class_num_list)
per_cls_weights = torch.FloatTensor(per_cls_weights).to(device)
criterion = CBFocalLoss(weight=per_cls_weights,
gamma=0.5).to(device)
# defining a nested list to store the cosine similarity (or distances) of the vectors from the class prototypes
class_dist_list = []
for i in range(num_class):
empty_dists = []
class_dist_list.append(empty_dists)
confusion_matrix = torch.zeros(num_class, num_class)
# iterating through the elements of the batch in the dataloader
with torch.no_grad():
for batch_idx, data_t in enumerate(loader):
im_data_t.data.resize_(data_t[0].size()).copy_(data_t[0])
gt_labels_t.data.resize_(data_t[1].size()).copy_(data_t[1])
feat = G(im_data_t)
output1 = F1(feat)
output_all = np.r_[output_all, output1.data.cpu().numpy()]
size += im_data_t.size(0)
pred1 = output1.data.max(1)[1]
# filling the elements of the confusion matrix
for t, p in zip(gt_labels_t.view(-1), pred1.view(-1)):
confusion_matrix[t.long(), p.long()] += 1
correct += pred1.eq(gt_labels_t.data).cpu().sum()
test_loss += criterion(output1, gt_labels_t) / len(loader)
pred1 = pred1.cpu().numpy()
dists = output1.data.max(1)[0]
dists = dists.cpu().numpy()
# forming the lists of the distances of the predicted labels and the class prototype
for label, dist in zip(pred1, dists):
label = int(label)
class_dist_list[label].append(dist)
# sorting the distances in ascending order for each of the classes, also finding a threshold for similarity of each class
summ = 0
class_dist_threshold_list = []
for class_ in range(len(class_dist_list)):
class_dist_list[class_].sort()
l = len(class_dist_list[class_])
tenth = l / 10
idx_tenth = math.ceil(tenth)
class_dist_threshold_list.append(
class_dist_list[class_][idx_tenth])
print('\nTest set: Average loss: {:.4f}, '
'Accuracy: {}/{} F1 ({:.2f}%)\n'.format(test_loss, correct, size,
100. * correct / size))
return test_loss.data, 100. * float(
correct) / size, class_dist_threshold_list
# choosing the mode of the model - whether to be used for training or for inference
if args.mode == 'train':
print("Training the model...")
train(None)
if args.mode == 'infer':
print("Infering from the model...")
_, _, class_dist_threshold_list = infer(target_loader_test)
print(
"Starting model retraining using weights for entropy maximization..."
)
train(class_dist_threshold_list)
if not os.path.exists(tf_record):
os.mkdir(tf_record)
tf_record = os.path.join(tf_record, args.arch + '_' + args.method)
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,
def __init__(self,
layers=50,
dropout=0.1,
classes=2,
zoom_factor=8,
reduction=2,
use_scale=True,
mode='embedded_gaussian'):
super(Nonlocal, self).__init__()
assert layers in [18, 34, 50, 101]
assert classes > 1
assert zoom_factor in [1, 2, 4, 8]
self.zoom_factor = zoom_factor
self.reduction = reduction
self.use_scale = use_scale
self.mode = mode
if layers == 18:
resnet = models.resnet18(pretrained=True,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
elif layers == 34:
resnet = models.resnet34(pretrained=True,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
elif layers == 50:
resnet = models.resnet50_semseg(pretrained=True,
deep_base=True,
strides=(1, 2, 1, 1),
dilations=(1, 1, 2, 4))
elif layers == 101:
resnet = models.resnet101_semseg(pretrained=True,
deep_base=True,
strides=(1, 2, 1, 1),
dilations=(1, 1, 2, 4))
if layers == 18 or layers == 34:
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu,
resnet.maxpool)
else:
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
if layers == 18 or layers == 34:
fea_dim = 512
aux_dim = 256
else:
fea_dim = 2048
aux_dim = 1024
down_dim = fea_dim // 4
self.conv1 = ConvBNReLU(fea_dim, down_dim, ks=3, stride=1, padding=1)
self.nl_block = NonLocal2d(in_channels=down_dim,
reduction=self.reduction,
use_scale=self.use_scale,
norm_cfg=dict(type='SyncBN',
requires_grad=True),
mode=self.mode)
self.conv2 = ConvBNReLU(down_dim, down_dim, ks=3, stride=1, padding=1)
self.cls = nn.Sequential(
nn.Conv2d(fea_dim + down_dim,
down_dim,
kernel_size=3,
padding=1,
bias=False), nn.BatchNorm2d(down_dim),
nn.ReLU(inplace=True), nn.Dropout2d(p=dropout),
nn.Conv2d(down_dim, classes, kernel_size=1))
if self.training:
self.aux = nn.Sequential(
nn.Conv2d(aux_dim,
aux_dim // 4,
kernel_size=3,
padding=1,
bias=False), nn.BatchNorm2d(aux_dim // 4),
nn.ReLU(inplace=True), nn.Dropout2d(p=dropout),
nn.Conv2d(aux_dim // 4, classes, kernel_size=1))
def __init__(self, layers=50, dropout=0.1, classes=2, vec_dim=300):
super(DCTNet, self).__init__()
assert layers in [18, 34, 50, 101]
assert classes > 1
self.vec_dim = vec_dim
# Backbone
if layers == 18:
resnet = models.resnet18(pretrained=False,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
elif layers == 34:
resnet = models.resnet34(pretrained=True,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
elif layers == 50:
resnet = models.resnet50_semseg(pretrained=True,
deep_base=True,
strides=(1, 2, 1, 1),
dilations=(1, 1, 2, 4))
elif layers == 101:
resnet = models.resnet101_semseg(pretrained=True,
deep_base=True,
strides=(1, 2, 1, 1),
dilations=(1, 1, 2, 4))
if layers == 18 or layers == 34:
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu,
resnet.maxpool)
else:
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
if layers == 18 or layers == 34:
fea_dim = 512
aux_dim = 256
else:
fea_dim = 2048
aux_dim = 1024
down_dim = fea_dim // 4
self.dct = nn.ModuleList()
for i in range(6, 10): # the number of in_channels is 2^i
self.dct.append(
DCTBlock(
in_channels=2**i,
mid_channels=32, # channels can be changed if you want.
up_flag=False if i == 9 else True,
up_channels=2**i + 2**(i + 1),
out_channels=2**i,
vec_dim=self.vec_dim))
self.cls = nn.Sequential(
nn.Conv2d(self.vec_dim,
down_dim,
kernel_size=3,
padding=1,
bias=False), nn.BatchNorm2d(down_dim),
nn.ReLU(inplace=True), nn.Dropout2d(p=dropout),
nn.Conv2d(down_dim, classes, kernel_size=1))
if self.training:
self.aux = nn.Sequential(
nn.Conv2d(aux_dim,
aux_dim // 4,
kernel_size=3,
padding=1,
bias=False), nn.BatchNorm2d(aux_dim // 4),
nn.ReLU(inplace=True), nn.Dropout2d(p=dropout),
nn.Conv2d(aux_dim // 4, classes, kernel_size=1))
self.gamma = gamma
self.weight = weight
def forward(self, input, target):
"""
input: [N, C]
target: [N, ]
"""
logpt = F.log_softmax(input, dim=1)
pt = torch.exp(logpt)
logpt = (1 - pt)**self.gamma * logpt
loss = F.nll_loss(logpt, target, self.weight)
return loss
model = nn.DataParallel(resnet34()).to(device)
print(model)
#criterion1 = nn.CrossEntropyLoss()
criterion1 = FocalLoss()
criterion2 = nn.SmoothL1Loss()
optimizer = optim.SGD(model.parameters(),
lr=learning_rate,
momentum=0.9,
weight_decay=5e-4)
exp_lr_scheduler = lr_scheduler.MultiStepLR(optimizer, [20, 45, 60], gamma=0.2)
#Train
total_step = len(train_loader)
curr_lr = learning_rate
Returns:
return a bilinear filter tensor
"""
factor = (kernel_size + 1) // 2
if kernel_size % 2 == 1:
center = factor - 1
else:
center = factor - 0.5
og = np.ogrid[:kernel_size, :kernel_size]
bilinear_filter = (1 - abs(og[0] - center) / factor) * (1 - abs(og[1] - center) / factor)
weight = np.zeros((in_channels, out_channels, kernel_size, kernel_size), dtype=np.float32)
weight[range(in_channels), range(out_channels), :, :] = bilinear_filter
return torch.from_numpy(weight)
pretrained_net = resnet.resnet34(pretrained=True)
pretrained_net.load_state_dict(torch.load(r'./pth/resnet34-333f7ec4.pth'))
class FcnResNet(nn.Module):
def __init__(self, num_classes):
super().__init__()
self.stage1 = nn.Sequential(*list(pretrained_net.children())[:-4])
self.stage2 = list(pretrained_net.children())[-4]
self.stage3 = list(pretrained_net.children())[-3]
self.scores1 = nn.Conv2d(512, num_classes, 1)
self.scores2 = nn.Conv2d(256, num_classes, 1)
self.scores3 = nn.Conv2d(128, num_classes, 1)
def __init__(self, layers=50, dropout=0.1, classes=2, zoom_factor=8):
super(DANet, self).__init__()
assert layers in [18, 34, 50, 101]
assert classes > 1
assert zoom_factor in [1, 2, 4, 8]
self.zoom_factor = zoom_factor
if layers == 18:
resnet = models.resnet18(pretrained=True,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
elif layers == 34:
resnet = models.resnet34(pretrained=True,
deep_base=False,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1))
elif layers == 50:
resnet = models.resnet50_semseg(pretrained=True,
deep_base=True,
strides=(1, 2, 1, 1),
dilations=(1, 1, 2, 4))
elif layers == 101:
resnet = models.resnet101_semseg(pretrained=True,
deep_base=True,
strides=(1, 2, 1, 1),
dilations=(1, 1, 2, 4))
if layers == 18 or layers == 34:
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu,
resnet.maxpool)
else:
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
if layers == 18 or layers == 34:
fea_dim = 512
aux_dim = 256
else:
fea_dim = 2048
aux_dim = 1024
down_dim = fea_dim // 4
self.pam_in_conv = ConvBNReLU(fea_dim,
down_dim,
ks=3,
stride=1,
padding=1)
self.pam = PAM(down_dim)
self.pam_out_conv = ConvBNReLU(down_dim,
down_dim,
ks=3,
stride=1,
padding=1)
self.pam_cls_seg = nn.Sequential(
nn.Dropout2d(p=dropout), nn.Conv2d(down_dim,
classes,
kernel_size=1))
self.cam_in_conv = ConvBNReLU(fea_dim,
down_dim,
ks=3,
stride=1,
padding=1)
self.cam = CAM(down_dim)
self.cam_out_conv = ConvBNReLU(down_dim,
down_dim,
ks=3,
stride=1,
padding=1)
self.cam_cls_seg = nn.Sequential(
nn.Dropout2d(p=dropout), nn.Conv2d(down_dim,
classes,
kernel_size=1))
self.cls_seg = nn.Sequential(
nn.Dropout2d(p=dropout), nn.Conv2d(down_dim,
classes,
kernel_size=1))
if self.training:
self.aux = nn.Sequential(
nn.Conv2d(aux_dim,
aux_dim // 4,
kernel_size=3,
padding=1,
bias=False), nn.BatchNorm2d(aux_dim // 4),
nn.ReLU(inplace=True), nn.Dropout2d(p=dropout),
nn.Conv2d(aux_dim // 4, classes, kernel_size=1))
'activation should be relu/elu/leakyrelu/rrelu/sigmoid/tanh')
# 设置是否使用多一层隐藏层+dropout
hidden = int(args.hidden)
# 选择模型
if args.layer == '18':
net = resnet18(pretrained=False,
progress=True,
activate=activate,
hidden=hidden,
num_classes=10)
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':
record_dir = 'record/'
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)
print('Source %s Target %s Labeled num perclass %s Network %s' %
(args.source, args.target, args.num, args.net))
source_loader, target_loader, target_loader_unl, target_loader_val, \
target_loader_test, class_list = return_dataset(args)
G = resnet34() # feature generator
inc = 512
params = []
for key, value in dict(G.named_parameters()).items():
if value.requires_grad:
if 'classifier' not in key:
params += [{
'params': [value],
'lr': args.multi,
'weight_decay': 0.0005
}]
else:
params += [{
'params': [value],
'lr': args.multi * 10,
'weight_decay': 0.0005
train_sampler = SubsetRandomSampler(train_idx)
valid_sampler = SubsetRandomSampler(valid_idx)
# prepare data loaders (combine dataset and smapler)
train_loader = DataLoader(dataset_train,
batch_size=64,
sampler=train_sampler,
num_workers=4)
valid_loader = DataLoader(dataset_train,
batch_size=64,
sampler=valid_sampler,
num_workers=4)
model = resnet34().to(device)
print(model)
criterion1 = nn.CrossEntropyLoss()
criterion2 = nn.MSELoss()
optimizer = optim.SGD(model.parameters(),
lr=learning_rate,
momentum=0.9,
weight_decay=5e-4)
exp_lr_scheduler = lr_scheduler.MultiStepLR(optimizer, [10, 25, 35, 45],
gamma=0.2)
#Train
total_step = len(train_loader)
curr_lr = learning_rate
print('Dataset %s Source %s Target %s Labeled num perclass %s Network %s' %
(args.dataset, args.source, args.target, args.num, args.net))
source_loader, target_loader, target_loader_unl, target_loader_val, \
target_loader_test, class_list = return_dataset(args)
use_gpu = torch.cuda.is_available()
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 == "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 += [{'params': [value], 'lr': args.multi,
def main():
# Training settings
parser = argparse.ArgumentParser(description='SSDA Classification')
parser.add_argument('--steps',
type=int,
default=50000,
metavar='N',
help='maximum number of iterations '
'to train (default: 50000)')
parser.add_argument(
'--method',
type=str,
default='MME',
choices=['S+T', 'ENT', 'MME'],
help='MME is proposed method, ENT is entropy minimization,'
' S+T is training only on labeled examples')
parser.add_argument('--lr',
type=float,
default=0.01,
metavar='LR',
help='learning rate (default: 0.001)')
parser.add_argument('--multi',
type=float,
default=0.1,
metavar='MLT',
help='learning rate multiplication')
parser.add_argument('--T',
type=float,
default=0.05,
metavar='T',
help='temperature (default: 0.05)')
parser.add_argument('--lamda',
type=float,
default=0.1,
metavar='LAM',
help='value of lamda')
parser.add_argument('--save_check',
action='store_true',
default=False,
help='save checkpoint or not')
parser.add_argument('--checkpath',
type=str,
default='./save_model_ssda',
help='dir to save checkpoint')
parser.add_argument('--seed',
type=int,
default=1,
metavar='S',
help='random seed (default: 1)')
parser.add_argument('--log-interval',
type=int,
default=100,
metavar='N',
help='how many batches to wait before logging '
'training status')
parser.add_argument('--save_interval',
type=int,
default=500,
metavar='N',
help='how many batches to wait before saving a model')
parser.add_argument('--net',
type=str,
default='alexnet',
help='which network to use')
parser.add_argument('--source',
type=str,
default='real',
help='source domain')
parser.add_argument('--target',
type=str,
default='sketch',
help='target domain')
parser.add_argument('--dataset',
type=str,
default='multi',
choices=['multi', 'office', 'office_home'],
help='the name of dataset')
parser.add_argument('--num',
type=int,
default=3,
help='number of labeled examples in the target')
parser.add_argument('--patience',
type=int,
default=5,
metavar='S',
help='early stopping to wait for improvment '
'before terminating. (default: 5 (5000 iterations))')
parser.add_argument('--early',
action='store_false',
default=True,
help='early stopping on validation or not')
parser.add_argument('--loss',
type=str,
default='CE',
choices=['CE', 'FL', 'CBFL'],
help='classifier loss function')
parser.add_argument(
'--attribute',
type=str,
default=None,
choices=[
'word2vec', 'glove.6B.100d.txt', 'glove.6B.300d.txt',
'glove_anurag', 'fasttext_anurag', 'glove.840B.300d.txt',
'glove.twitter.27B.200d.txt', 'glove.twitter.27B.50d.txt',
'glove.42B.300d.txt', 'glove.6B.200d.txt', 'glove.6B.50d.txt',
'glove.twitter.27B.100d.txt', 'glove.twitter.27B.25d.txt'
],
help='semantic attribute feature vector to be used')
parser.add_argument(
'--dim',
type=int,
default=300,
help=
'dimensionality of the feature vector - make sure this in sync with the dim of the semantic attribute vector'
)
parser.add_argument(
'--deep',
type=int,
default=0,
help='type of classification predictor - 0 for shallow, 1 for deep')
parser.add_argument('--mode',
type=str,
default='train',
choices=['train', 'infer'],
help='mode of script train or infer')
args = parser.parse_args()
print('Dataset %s Source %s Target %s Labeled num perclass %s Network %s' %
(args.dataset, args.source, args.target, args.num, args.net))
source_loader, target_loader, target_loader_unl, target_loader_val, \
target_loader_test, class_num_list, class_list = return_dataset(args) # class num list is returned for CBFL
use_gpu = torch.cuda.is_available()
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))
if use_gpu:
device = 'cuda'
else:
device = 'cpu'
print("Device: %s Loss: %s Attributes: %s" %
(device, args.loss, args.attribute))
if use_gpu:
torch.cuda.manual_seed(args.seed)
else:
torch.manual_seed(args.seed)
if args.net == 'resnet34':
G = resnet34()
inc = 512
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 += [{
'params': [value],
'lr': 0.1,
'weight_decay': 0.0005
}]
else:
params += [{
'params': [value],
'lr': 1,
'weight_decay': 0.0005
}]
# Setting the predictor layer
if args.attribute is not None:
if args.deep:
F1 = Predictor_deep_attributes(num_class=len(class_list),
inc=inc,
feat_dim=args.dim)
print("Using: Predictor_deep_attributes")
else:
F1 = Predictor_attributes(num_class=len(class_list),
inc=inc,
feat_dim=args.dim)
print("Using: Predictor_attributes")
else:
if args.deep:
F1 = Predictor_deep(num_class=len(class_list), inc=inc)
print("Using: Predictor_deep")
else:
F1 = Predictor(num_class=len(class_list), inc=inc, temp=args.T)
print("Using: Predictor")
# Initializing the weights of the prediction layer
weights_init(F1)
# Setting the prediction layer weights as the semantic attributes
if args.attribute is not None:
att = np.load('attributes/%s_%s.npy' % (args.dataset, args.attribute))
if use_gpu:
att = nn.Parameter(torch.cuda.FloatTensor(att))
else:
att = nn.Parameter(torch.FloatTensor(att, device="cpu"))
if args.deep:
F1.fc3.weight = att
else:
F1.fc2.weight = att
print("attribute shape is: ", att.shape)
lr = args.lr
# loading the model checkpoint - and printing some parameters relating to the checkpoints
main_dict = torch.load(args.checkpath + "/" + args.net + "_" +
args.method + "_" + args.source + "_" +
args.target + ".ckpt.best.pth.tar")
G.load_state_dict(main_dict['G_state_dict'])
F1.load_state_dict(main_dict['F1_state_dict'])
print("Loaded pretrained model weights")
print("Loaded weights from step: ", main_dict['step'])
print("Current best test acc is: ", main_dict['best_acc_test'])
G.to(device)
F1.to(device)
# Loading the txt file having the weights and paths of the image file as a data frame
df = pd.read_csv(args.method + '_' + main_dict['arch'] + '_' +
str(main_dict['step']) + '.txt',
sep=" ",
header=None)
df = df[[3, 0]]
df = df.rename(columns={3: "img", 0: "weight"})
im_data_s = torch.FloatTensor(1)
im_data_t = torch.FloatTensor(1)
im_data_tu = torch.FloatTensor(1)
gt_labels_t = torch.LongTensor(1)
gt_labels_s = torch.LongTensor(1)
sample_labels_t = torch.LongTensor(1)
sample_labels_s = torch.LongTensor(1)
im_data_s = im_data_s.to(device)
im_data_t = im_data_t.to(device)
im_data_tu = im_data_tu.to(device)
gt_labels_s = gt_labels_s.to(device)
gt_labels_t = gt_labels_t.to(device)
sample_labels_t = sample_labels_t.to(device)
sample_labels_s = sample_labels_s.to(device)
im_data_s = Variable(im_data_s)
im_data_t = Variable(im_data_t)
im_data_tu = Variable(im_data_tu)
gt_labels_s = Variable(gt_labels_s)
gt_labels_t = Variable(gt_labels_t)
sample_labels_t = Variable(sample_labels_t)
sample_labels_s = Variable(sample_labels_s)
if os.path.exists(args.checkpath) == False:
os.mkdir(args.checkpath)
def train():
G.train()
F1.train()
optimizer_g = optim.SGD(params,
momentum=0.9,
weight_decay=0.0005,
nesterov=True)
optimizer_f = optim.SGD(list(F1.parameters()),
lr=1.0,
momentum=0.9,
weight_decay=0.0005,
nesterov=True)
# Loading the states of the two optmizers
optimizer_g.load_state_dict(main_dict['optimizer_g'])
optimizer_f.load_state_dict(main_dict['optimizer_f'])
print("Loaded optimizer states")
def zero_grad_all():
optimizer_g.zero_grad()
optimizer_f.zero_grad()
param_lr_g = []
for param_group in optimizer_g.param_groups:
param_lr_g.append(param_group["lr"])
param_lr_f = []
for param_group in optimizer_f.param_groups:
param_lr_f.append(param_group["lr"])
# Setting the loss function to be used for the classification loss
if args.loss == 'CE':
criterion = nn.CrossEntropyLoss().to(device)
if args.loss == 'FL':
criterion = FocalLoss(alpha=1, gamma=1).to(device)
if args.loss == 'CBFL':
# Calculating the list having the number of examples per class which is going to be used in the CB focal loss
beta = 0.99
effective_num = 1.0 - np.power(beta, class_num_list)
per_cls_weights = (1.0 - beta) / np.array(effective_num)
per_cls_weights = per_cls_weights / np.sum(per_cls_weights) * len(
class_num_list)
per_cls_weights = torch.FloatTensor(per_cls_weights).to(device)
criterion = CBFocalLoss(weight=per_cls_weights,
gamma=0.5).to(device)
all_step = args.steps
data_iter_s = iter(source_loader)
data_iter_t = iter(target_loader)
data_iter_t_unl = iter(target_loader_unl)
len_train_source = len(source_loader)
len_train_target = len(target_loader)
len_train_target_semi = len(target_loader_unl)
best_acc = 0
counter = 0
for step in range(all_step):
optimizer_g = inv_lr_scheduler(param_lr_g,
optimizer_g,
step,
init_lr=args.lr)
optimizer_f = inv_lr_scheduler(param_lr_f,
optimizer_f,
step,
init_lr=args.lr)
lr = optimizer_f.param_groups[0]['lr']
# condition for restarting the iteration for each of the data loaders
if step % len_train_target == 0:
data_iter_t = iter(target_loader)
if step % len_train_target_semi == 0:
data_iter_t_unl = iter(target_loader_unl)
if step % len_train_source == 0:
data_iter_s = iter(source_loader)
data_t = next(data_iter_t)
data_t_unl = next(data_iter_t_unl)
data_s = next(data_iter_s)
with torch.no_grad():
im_data_s.resize_(data_s[0].size()).copy_(data_s[0])
gt_labels_s.resize_(data_s[1].size()).copy_(data_s[1])
im_data_t.resize_(data_t[0].size()).copy_(data_t[0])
gt_labels_t.resize_(data_t[1].size()).copy_(data_t[1])
im_data_tu.resize_(data_t_unl[0].size()).copy_(data_t_unl[0])
zero_grad_all()
data = torch.cat((im_data_s, im_data_t), 0)
target = torch.cat((gt_labels_s, gt_labels_t), 0)
output = G(data)
out1 = F1(output)
loss = criterion(out1, target)
loss.backward(retain_graph=True)
optimizer_g.step()
optimizer_f.step()
zero_grad_all()
# list of the weights and image paths in this batch
img_paths = list(data_t_unl[2])
df1 = df.loc[df['img'].isin(img_paths)]
df1 = df1['weight']
weight_list = list(df1)
if not args.method == 'S+T':
output = G(im_data_tu)
if args.method == 'ENT':
loss_t = entropy(F1, output, args.lamda)
loss_t.backward()
optimizer_f.step()
optimizer_g.step()
elif args.method == 'MME':
loss_t = adentropy(F1, output, args.lamda, weight_list)
loss_t.backward()
optimizer_f.step()
optimizer_g.step()
else:
raise ValueError('Method cannot be recognized.')
log_train = 'S {} T {} Train Ep: {} lr{} \t ' \
'Loss Classification: {:.6f} Loss T {:.6f} ' \
'Method {}\n'.format(args.source, args.target,
step, lr, loss.data,
-loss_t.data, args.method)
else:
log_train = 'S {} T {} Train Ep: {} lr{} \t ' \
'Loss Classification: {:.6f} Method {}\n'.\
format(args.source, args.target,
step, lr, loss.data,
args.method)
G.zero_grad()
F1.zero_grad()
zero_grad_all()
if step % args.log_interval == 0:
print(log_train)
if step % args.save_interval == 0 and step > 0:
loss_val, acc_val = test(target_loader_val)
loss_test, acc_test = test(target_loader_test)
G.train()
F1.train()
if acc_test >= best_acc:
best_acc = acc_test
best_acc_test = acc_test
counter = 0
else:
counter += 1
if args.early:
if counter > args.patience:
break
print('best acc test %f best acc val %f' %
(best_acc_test, acc_val))
print('record %s' % record_file)
with open(record_file, 'a') as f:
f.write('step %d best %f final %f \n' %
(step, best_acc_test, acc_val))
G.train()
F1.train()
#saving model as a checkpoint dict having many things
if args.save_check:
print('saving model')
is_best = True if counter == 0 else False
save_mymodel(
args, {
'step': step,
'arch': args.net,
'G_state_dict': G.state_dict(),
'F1_state_dict': F1.state_dict(),
'best_acc_test': best_acc_test,
'optimizer_g': optimizer_g.state_dict(),
'optimizer_f': optimizer_f.state_dict(),
}, is_best)
# defining the function for in training validation and testing
def test(loader):
G.eval()
F1.eval()
test_loss = 0
correct = 0
size = 0
num_class = len(class_list)
output_all = np.zeros((0, num_class))
# Setting the loss function to be used for the classification loss
if args.loss == 'CE':
criterion = nn.CrossEntropyLoss().to(device)
if args.loss == 'FL':
criterion = FocalLoss(alpha=1, gamma=1).to(device)
if args.loss == 'CBFL':
# Calculating the list having the number of examples per class which is going to be used in the CB focal loss
beta = 0.99
effective_num = 1.0 - np.power(beta, class_num_list)
per_cls_weights = (1.0 - beta) / np.array(effective_num)
per_cls_weights = per_cls_weights / np.sum(per_cls_weights) * len(
class_num_list)
per_cls_weights = torch.FloatTensor(per_cls_weights).to(device)
criterion = CBFocalLoss(weight=per_cls_weights,
gamma=0.5).to(device)
confusion_matrix = torch.zeros(num_class, num_class)
with torch.no_grad():
for batch_idx, data_t in enumerate(loader):
im_data_t.data.resize_(data_t[0].size()).copy_(data_t[0])
gt_labels_t.data.resize_(data_t[1].size()).copy_(data_t[1])
feat = G(im_data_t)
output1 = F1(feat)
output_all = np.r_[output_all, output1.data.cpu().numpy()]
size += im_data_t.size(0)
pred1 = output1.data.max(1)[1]
for t, p in zip(gt_labels_t.view(-1), pred1.view(-1)):
confusion_matrix[t.long(), p.long()] += 1
correct += pred1.eq(gt_labels_t.data).cpu().sum()
test_loss += criterion(output1, gt_labels_t) / len(loader)
np.save("cf_target.npy", confusion_matrix)
#print(confusion_matrix)
print('\nTest set: Average loss: {:.4f}, '
'Accuracy: {}/{} F1 ({:.0f}%)\n'.format(test_loss, correct, size,
100. * correct / size))
return test_loss.data, 100. * float(correct) / size
print("Starting stage 2 training of the model ...")
train()
