def __init__(self, backbone, hyperpixel_ids, benchmark, device):
r"""Constructor for Hyperpixel Flow framework"""
# Feature extraction network initialization.
if backbone == 'resnet50':
self.backbone = resnet.resnet50(pretrained=True).to(device)
nbottlenecks = [3, 4, 6, 3]
elif backbone == 'resnet101':
self.backbone = resnet.resnet101(pretrained=True).to(device)
nbottlenecks = [3, 4, 23, 3]
elif backbone == 'fcn101':
self.backbone = gcv.models.get_fcn_resnet101_voc(
pretrained=True).to(device).pretrained
nbottlenecks = [3, 4, 23, 3]
else:
raise Exception('Unavailable backbone: %s' % backbone)
self.bottleneck_ids = reduce(
add, list(map(lambda x: list(range(x)), nbottlenecks)))
self.layer_ids = reduce(add, [[i + 1] * x
for i, x in enumerate(nbottlenecks)])
self.backbone.eval()
# Hyperpixel id and pre-computed jump and receptive field size initialization
# Reference: https://fomoro.com/research/article/receptive-field-calculator
# (the jump and receptive field sizes for 'fcn101' are heuristic values)
self.hyperpixel_ids = util.parse_hyperpixel(hyperpixel_ids)
self.jsz = torch.tensor([4, 4, 4, 4, 8, 8, 8, 8, 16, 16]).to(device)
self.rfsz = torch.tensor([11, 19, 27, 35, 43, 59, 75, 91, 107,
139]).to(device)
# Miscellaneous
self.hsfilter = geometry.gaussian2d(7).to(device)
self.device = device
self.benchmark = benchmark
def __init__(self, backbone='resnet50'):
super().__init__()
if backbone == 'resnet18':
backbone = resnet18(pretrained=True)
self.out_channels = [256, 512, 512, 256, 256, 128]
elif backbone == 'resnet34':
backbone = resnet34(pretrained=True)
self.out_channels = [256, 512, 512, 256, 256, 256]
elif backbone == 'resnet50':
backbone = resnet50(pretrained=True)
self.out_channels = [1024, 512, 512, 256, 256, 256]
elif backbone == 'resnet101':
backbone = resnet101(pretrained=True)
self.out_channels = [1024, 512, 512, 256, 256, 256]
else: # backbone == 'resnet152':
backbone = resnet152(pretrained=True)
self.out_channels = [1024, 512, 512, 256, 256, 256]
self.feature_extractor = nn.Sequential(*list(backbone.children())[:7])
conv4_block1 = self.feature_extractor[-1][0]
conv4_block1.conv1.stride = (1, 1)
conv4_block1.conv2.stride = (1, 1)
conv4_block1.downsample[0].stride = (1, 1)
def __init__(self,
pretrained=True,
output_channel=1,
rate=1,
fixed_feature_weights=False):
super(SANet, self).__init__()
self.output_channel = output_channel
resnet = resnet101(pretrained=pretrained)
# Freeze resnet weights
if fixed_feature_weights:
for p in resnet.parameters():
p.requires_grad = False
c = [i // rate for i in [64, 256, 512, 1024, 2048]]
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu)
self.layer1 = nn.Sequential(resnet.maxpool, resnet.layer1)
self.layer2 = nn.Sequential(resnet.layer2)
self.layer3 = nn.Sequential(resnet.layer3)
self.layer4 = nn.Sequential(resnet.layer4)
# previous layers
self.mid1 = conv_bn(c[0], c[0], 1)
self.mid2 = conv_bn(c[1], c[1], 1)
self.mid3 = conv_bn(c[2], c[2], 1)
self.mid4 = conv_bn(c[3], c[3], 1)
# Smooth layers
self.smooth1 = conv_bn(c[4], c[3], 1)
self.smooth2 = conv_bn(c[3], c[2], 1)
self.smooth3 = conv_bn(c[2], c[1], 1)
self.smooth4 = conv_bn(c[0] * 2, c[0], 1)
self.smooth5 = conv(c[0] * 4, c[0] * 2, 1)
self.upconv1 = conv_bn(c[4], c[3], 1)
self.upconv2 = conv_bn(c[3], c[2], 1)
self.upconv3 = conv_bn(c[2], c[1], 1)
self.upconv4 = conv_bn(c[1], c[0], 1)
self.upconv5 = conv(c[0] * 2, c[0], 1)
self.att1 = SAB(c[3], c[3])
self.att2 = SAB(c[2], c[2])
self.att3 = SAB(c[1], c[1])
self.att4 = SAB(c[0], c[0])
# upshuffle layers
self.up1 = upshuffle(c[3], c[0], 8)
self.up2 = upshuffle(c[2], c[0], 4)
self.up3 = upshuffle(c[1], c[0], 2)
# Context block
self.gc2 = ContextBlock(c[1], 1 / 4)
self.gc3 = ContextBlock(c[2], 1 / 4)
self.gc4 = ContextBlock(c[3], 1 / 4)
self.gc5 = ContextBlock(c[4], 1 / 4)
# Depth prediction
self.predict = predict(c[0], self.output_channel)
def __init__(self, backbone, use_original_imgsize):
super(HypercorrSqueezeNetwork, self).__init__()
# 1. Backbone network initialization
self.backbone_type = backbone
self.use_original_imgsize = use_original_imgsize
if backbone == 'vgg16':
self.backbone = vgg.vgg16(pretrained=True)
self.feat_ids = [17, 19, 21, 24, 26, 28, 30]
self.extract_feats = extract_feat_vgg
nbottlenecks = [2, 2, 3, 3, 3, 1]
elif backbone == 'resnet50':
self.backbone = resnet.resnet50(pretrained=True)
self.feat_ids = list(range(4, 17))
self.extract_feats = extract_feat_res
nbottlenecks = [3, 4, 6, 3]
elif backbone == 'resnet101':
self.backbone = resnet.resnet101(pretrained=True)
self.feat_ids = list(range(4, 34))
self.extract_feats = extract_feat_res
nbottlenecks = [3, 4, 23, 3]
else:
raise Exception('Unavailable backbone: %s' % backbone)
self.bottleneck_ids = reduce(
add, list(map(lambda x: list(range(x)), nbottlenecks)))
self.lids = reduce(add,
[[i + 1] * x for i, x in enumerate(nbottlenecks)])
self.stack_ids = torch.tensor(
self.lids).bincount().__reversed__().cumsum(dim=0)[:3]
self.backbone.eval()
self.hpn_learner = HPNLearner(list(reversed(nbottlenecks[-3:])))
self.cross_entropy_loss = nn.CrossEntropyLoss()
def create_dataset(split):
batch_size = 50
resnet = resnet101(True).cuda()
resnet.eval()
resnet.forward = forward.__get__(resnet, ResNet)
dataloader = DataLoader(CLEVR(sys.argv[1], split),
batch_size=batch_size,
num_workers=4)
size = len(dataloader)
print(split, 'total', size * batch_size)
f = h5py.File('data/{}_features.hdf5'.format(split), 'w', libver='latest')
dset = f.create_dataset('data', (size * batch_size, 1024, 14, 14),
dtype='f4')
with torch.no_grad():
for i, image in tqdm(enumerate(dataloader)):
image = image.to(device)
features = resnet(image).detach().cpu().numpy()
dset[i * batch_size:(i + 1) * batch_size] = features
f.close()
def __init__(self, mode_name='wide_resnet50_2'):
super(Backbone, self).__init__()
assert mode_name in ('resnet18', 'resnet34', 'resnet50', 'resnet101',
'resnet152', 'resnext50_32x4d',
'resnext101_32x8d', 'wide_resnet50_2',
'wide_resnet101_2')
if mode_name == 'resnet18':
self.res_back_bone = resnet.resnet18()
elif mode_name == 'resnet34':
self.res_back_bone = resnet.resnet34()
elif mode_name == 'resnet50':
self.res_back_bone = resnet.resnet50()
elif mode_name == 'resnet101':
self.res_back_bone = resnet.resnet101()
elif mode_name == 'resnet152':
self.res_back_bone = resnet.resnet152()
elif mode_name == 'resnext50_32x4d':
self.res_back_bone = resnet.resnext50_32x4d()
elif mode_name == 'resnext101_32x8d':
self.res_back_bone = resnet.resnext101_32x8d()
elif mode_name == 'wide_resnet50_2':
self.res_back_bone = resnet.wide_resnet50_2()
else:
self.res_back_bone = resnet.wide_resnet101_2()
self.backbone = nn.Module()
layer0 = nn.Sequential(*list(self.res_back_bone.children())[:4])
self.backbone.add_module('layer0', layer0)
self.backbone.add_module('layer1', self.res_back_bone.layer1)
self.backbone.add_module('layer2', self.res_back_bone.layer2)
self.backbone.add_module('layer3', self.res_back_bone.layer3)
self.backbone.add_module('layer4', self.res_back_bone.layer4)
def __init__(self, backbone="resnet50", backbone_path=None):
super().__init__()
if backbone == "resnet18":
backbone = resnet18(pretrained=not backbone_path)
self.out_channels = [256, 512, 512, 256, 256, 128]
elif backbone == "resnet34":
backbone = resnet34(pretrained=not backbone_path)
self.out_channels = [256, 512, 512, 256, 256, 256]
elif backbone == "resnet50":
backbone = resnet50(pretrained=not backbone_path)
self.out_channels = [1024, 512, 512, 256, 256, 256]
elif backbone == "resnet101":
backbone = resnet101(pretrained=not backbone_path)
self.out_channels = [1024, 512, 512, 256, 256, 256]
else:
backbone = resnet152(pretrained=not backbone_path)
self.out_channels = [1024, 512, 512, 256, 256, 256]
if backbone_path:
backbone.load_state_dict(torch.load(backbone_path))
self.feature_extractor = nn.Sequential(*list(backbone.children())[:7])
conv4_block1 = self.feature_extractor[-1][0]
conv4_block1.conv1.stride = (1, 1)
conv4_block1.conv2.stride = (1, 1)
conv4_block1.downsample[0].stride = (1, 1)
def __init__(self, name="resnet50", pretrain=True):
super().__init__()
if name == "resnet50":
base_net = resnet.resnet50(pretrained=False)
elif name == "resnet101":
base_net = resnet.resnet101(pretrained=False)
else:
print(" base model is not support !")
if pretrain:
print("load the {} weight from ./cache".format(name))
base_net.load_state_dict(
model_zoo.load_url(model_urls["resnet50"],
model_dir="../cache"))
# print(base_net)
self.stage1 = nn.Sequential(base_net.conv1, base_net.bn1,
base_net.relu, base_net.maxpool)
self.stage2 = base_net.layer1
self.stage3 = base_net.layer2
self.stage4 = base_net.layer3
self.stage5 = base_net.layer4
self.up2 = nn.ConvTranspose2d(64,
64,
kernel_size=4,
stride=2,
padding=1)
def __init__(self, backbone, num_classes=21):
super().__init__()
if backbone == 'resnet-18':
self.backbone = resnet18()
self.in_channel = [512, 128, 64, 64]
self.out_channel = [256, 64, 64, 64]
elif backbone == 'resnet-34':
self.backbone = resnet34()
self.in_channel = [512, 128, 64, 64]
self.out_channel = [256, 64, 64, 64]
elif backbone == 'resnet-50':
self.backbone = resnet50()
self.in_channel = [2048, 512, 256, 64]
self.out_channel = [1024, 256, 64, 64]
elif backbone == 'resnet-101':
self.backbone = resnet101()
self.in_channel = [2048, 512, 256, 64]
self.out_channel = [1024, 256, 64, 64]
elif backbone == 'resnet-152':
self.backbone = resnet152()
self.in_channel = [2048, 512, 256, 64]
self.out_channel = [1024, 256, 64, 64]
else:
raise NotImplementedError
self.encoder = Encoder(self.backbone)
self.decoder = Decoder(self.in_channel, self.out_channel)
self.out = nn.Conv2d(64, num_classes, 1)
self._init_weight()
def __init__(self, backbone='resnet50', pretrained_path=None):
super().__init__()
if backbone == 'resnet18':
backbone = resnet18(pretrained=not pretrained_path)
self.final_out_channels = 256
self.low_level_inplanes = 64
elif backbone == 'resnet34':
backbone = resnet34(pretrained=not pretrained_path)
self.final_out_channels = 256
self.low_level_inplanes = 64
elif backbone == 'resnet50':
backbone = resnet50(pretrained=not pretrained_path)
self.final_out_channels = 1024
self.low_level_inplanes = 256
elif backbone == 'resnet101':
backbone = resnet101(pretrained=not pretrained_path)
self.final_out_channels = 1024
self.low_level_inplanes = 256
else: # backbone == 'resnet152':
backbone = resnet152(pretrained=not pretrained_path)
self.final_out_channels = 1024
self.low_level_inplanes = 256
if pretrained_path:
backbone.load_state_dict(torch.load(pretrained_path))
self.early_extractor = nn.Sequential(*list(backbone.children())[:5])
self.later_extractor = nn.Sequential(*list(backbone.children())[5:7])
conv4_block1 = self.later_extractor[-1][0]
conv4_block1.conv1.stride = (1, 1)
conv4_block1.conv2.stride = (1, 1)
conv4_block1.downsample[0].stride = (1, 1)
def main():
device = torch.device(
'cuda:0' if args.gpu and torch.cuda.is_available() else 'cpu')
dataset_dir = '../' + args.dataset_dir + '/Data/RGB/'
dataloader, dataset_size = load_data(dataset_dir, args.batch_size,
args.num_workers)
model = resnet101(num_classes=51)
model.to(device)
model.load_state_dict(torch.load(args.model_path, map_location='cpu'))
vis_results = visualize(model, dataloader, dataset_size, device,
args.selected_layer)
layer_output = vis_results[0][0]
print(layer_output.shape)
num_features = layer_output.shape[0]
num_row = int(num_features**0.5)
for i in range(num_features):
feature = layer_output[i].data.numpy()
feature = 1.0 / (1 + np.exp(-1 * feature))
feature = np.round(feature * 255)
ax = plt.subplot(num_row, num_row, i + 1)
ax.axis('off')
plt.imshow(feature, cmap=plt.cm.gray)
plt.subplots_adjust(wspace=0.2, hspace=0)
plt.savefig('{0}.jpg'.format(args.selected_layer))
def __init__(self, pretrained=None):
super(HeadCommon, self).__init__()
self.config = ConfigParser()
config_path = os.path.abspath(
os.path.join(__file__, "../../", "config.ini"))
assert os.path.exists(config_path), "config.ini not exists!"
self.config.read(config_path)
self.backbone_type = self.config['BACKBONE']['BACKBONE_TYPE']
_pretrained = True if pretrained is not None else False
assert self.backbone_type in ['resnet', 'vgg16']
if self.backbone_type == 'resnet':
resnet_layer = int(self.config['BACKBONE']['RESNET_LAYER'])
assert resnet_layer in [18, 34, 50, 101, 152]
if resnet_layer == 18:
_resnet = resnet.resnet18(_pretrained)
elif resnet_layer == 34:
_resnet = resnet.resnet34(_pretrained)
elif resnet_layer == 50:
_resnet = resnet.resnet50(_pretrained)
elif resnet_layer == 101:
_resnet = resnet.resnet101(_pretrained)
else:
_resnet = resnet.resnet152(_pretrained)
# using resnet_c5 the last bottle neck of resnet
_resnet.layer4[0].conv2.stride = (1, 1)
_resnet.layer4[0].downsample[0].stride = (1, 1)
self.resnet_c5 = _resnet.layer4
self.resnet_c5_avg = _resnet.avgpool
elif self.backbone_type == 'vgg16':
assert not bool(int(self.config['HEAD']['MASK_HEAD_ON'])), (
"When mask head on, not support vgg16 backbone.")
vgg = vgg16(pretrained=True)
self.vgg_fc = nn.Sequential(
*list(vgg.classifier._modules.values())[:-1])
def __init__(self, backbone='resnet50', backbone_path=None):
super().__init__()
if backbone == 'resnet18':
backbone = resnet18(pretrained=not backbone_path)
self.out_channels = [256, 512, 512, 256, 256, 128]
elif backbone == 'resnet34':
backbone = resnet34(pretrained=not backbone_path)
self.out_channels = [256, 512, 512, 256, 256, 256]
elif backbone == 'resnet50':
backbone = resnet50(pretrained=not backbone_path)
self.out_channels = [1024, 512, 512, 256, 256, 256]
elif backbone == 'resnet101':
backbone = resnet101(pretrained=not backbone_path)
self.out_channels = [1024, 512, 512, 256, 256, 256]
else: # backbone == 'resnet152':
backbone = resnet152(pretrained=not backbone_path)
self.out_channels = [1024, 512, 512, 256, 256, 256]
if backbone_path:
backbone.load_state_dict(torch.load(backbone_path))
for name, parameter in backbone.named_parameters():
if 'layer2' not in name and 'layer3' not in name and 'layer4' not in name:
parameter.requires_grad_(False)
self.feature_extractor = nn.Sequential(*list(backbone.children())[:7])
conv4_block1 = self.feature_extractor[-1][0]
conv4_block1.conv1.stride = (1, 1)
conv4_block1.conv2.stride = (1, 1)
conv4_block1.downsample[0].stride = (1, 1)
def __init__(self, resnet_layer, pretrained=None):
super(ResNet, self).__init__()
assert resnet_layer in [18, 34, 50, 101, 152]
pretrained = True if pretrained is not None else False
if resnet_layer == 18:
self.resnet = resnet.resnet18(pretrained)
elif resnet_layer == 34:
self.resnet = resnet.resnet34(pretrained)
elif resnet_layer == 50:
self.resnet = resnet.resnet50(pretrained)
elif resnet_layer == 101:
self.resnet = resnet.resnet101(pretrained)
else:
self.resnet = resnet.resnet152(pretrained)
# fix layer grad
for p in self.resnet.conv1.parameters():
p.requires_grad = False
for p in self.resnet.layer1.parameters():
p.requires_grad = False
for p in self.resnet.layer2.parameters():
p.requires_grad = True
for p in self.resnet.layer3.parameters():
p.requires_grad = True
# fix batch norm layer
for m in self.resnet.modules():
if isinstance(m, nn.BatchNorm2d):
for p in m.parameters():
p.requires_grad = False
def __init__(self, pretrained=True):
super(I2D, self).__init__()
resnet = resnet101(pretrained=pretrained)
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool)
self.layer1 = nn.Sequential(resnet.layer1) # 256
self.layer2 = nn.Sequential(resnet.layer2) # 512
self.layer3 = nn.Sequential(resnet.layer3) # 1024
self.layer4 = nn.Sequential(resnet.layer4) # 2048
# Top layer
self.toplayer = nn.Conv2d(2048, 256, kernel_size=1, stride=1, padding=0) # Reduce channels
# Lateral layers
self.latlayer1 = nn.Conv2d(1024, 256, kernel_size=1, stride=1, padding=0)
self.latlayer2 = nn.Conv2d( 512, 256, kernel_size=1, stride=1, padding=0)
self.latlayer3 = nn.Conv2d( 256, 256, kernel_size=1, stride=1, padding=0)
# Smooth layers
self.smooth1 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
self.smooth2 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
self.smooth3 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
# Depth prediction
self.predict1 = smooth(256, 64)
self.predict2 = predict(64, 1)
def __init__(self, pretrained=True, fixed_feature_weights=False):
super(I2D, self).__init__()
resnet = resnet101(pretrained=pretrained)
# Freeze those weights
if fixed_feature_weights:
for p in resnet.parameters():
p.requires_grad = False
self.layer0 = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu,
resnet.maxpool)
self.layer1 = nn.Sequential(resnet.layer1)
self.layer2 = nn.Sequential(resnet.layer2)
self.layer3 = nn.Sequential(resnet.layer3)
self.layer4 = nn.Sequential(resnet.layer4)
# Top layer
self.toplayer = nn.Conv2d(2048,
256,
kernel_size=1,
stride=1,
padding=0) # Reduce channels
# Lateral layers
self.latlayer1 = nn.Conv2d(1024,
256,
kernel_size=1,
stride=1,
padding=0)
self.latlayer2 = nn.Conv2d(512,
256,
kernel_size=1,
stride=1,
padding=0)
self.latlayer3 = nn.Conv2d(256,
256,
kernel_size=1,
stride=1,
padding=0)
# Smooth layers
self.smooth1 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
self.smooth2 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
self.smooth3 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
# Aggregate layers
self.agg1 = agg_node(256, 128)
self.agg2 = agg_node(256, 128)
self.agg3 = agg_node(256, 128)
self.agg4 = agg_node(256, 128)
# Upshuffle layers
self.up1 = upshuffle(128, 128, 8)
self.up2 = upshuffle(128, 128, 4)
self.up3 = upshuffle(128, 128, 2)
# Depth prediction
self.predict1 = smooth(512, 128)
self.predict2 = predict(128, 1)
def get(cls, args):
model = ResNet3D(Bottleneck3D, [3, 8, 36, 3]) # 101
if args.pretrained:
from torchvision.models.resnet import resnet101
model2d = resnet101(pretrained=True)
model.load_2d(model2d)
return model
def get_classification_model(classifier_name='resnet18',
num_classes=1000,
pretrained=True):
"""
Get the detection model
:param pretrained:
:param classifier_name:
:param num_classes:
:return:
"""
if classifier_name == 'resnet18':
model = resnet18(pretrained, num_classes=num_classes)
elif classifier_name == 'resnet34':
model = resnet34(pretrained, num_classes=num_classes)
elif classifier_name == 'resnet50':
model = resnet50(pretrained, num_classes=num_classes)
elif classifier_name == 'resnet101':
model = resnet101(pretrained, num_classes=num_classes)
elif classifier_name == 'resnet152':
model = resnet152(pretrained, num_classes=num_classes)
else:
raise ValueError('Unsupported resnet type.')
return model
def __init__(self, resnet_layer, pretrained=None):
super(ResNetFPN, self).__init__()
self.lateral_conv1 = nn.Conv2d(2048, 256, kernel_size=1)
self.lateral_conv2 = nn.Conv2d(1024, 256, kernel_size=1)
self.lateral_conv3 = nn.Conv2d(512, 256, kernel_size=1)
self.lateral_conv4 = nn.Conv2d(256, 256, kernel_size=1)
self.anti_aliasing_conv1 = nn.Conv2d(256,
256,
kernel_size=3,
padding=1)
self.anti_aliasing_conv2 = nn.Conv2d(256,
256,
kernel_size=3,
padding=1)
self.anti_aliasing_conv3 = nn.Conv2d(256,
256,
kernel_size=3,
padding=1)
self._init_parameters()
assert resnet_layer in [18, 34, 50, 101, 152]
pretrained = True if pretrained is not None else False
if resnet_layer == 18:
self.resnet = resnet.resnet18(pretrained)
elif resnet_layer == 34:
self.resnet = resnet.resnet34(pretrained)
elif resnet_layer == 50:
self.resnet = resnet.resnet50(pretrained)
elif resnet_layer == 101:
self.resnet = resnet.resnet101(pretrained)
elif resnet_layer == 152:
self.resnet = resnet.resnet152(pretrained)
def load_resnet():
model = resnet101(pretrained=True)
del model.layer4
del model.avgpool
del model.fc
return model
def __init__(self):
super(MGN_PTL, self).__init__()
num_classes = opt.classn
feats = 256
self.val = False
self.ptl = PTL()
self.ptl.init_param()
if opt.backbone == 'resnet50':
self.backbone = resnet50(pretrained=True)
elif opt.backbone == 'resnet101':
self.backbone = resnet101(pretrained=True)
res_conv4 = nn.Sequential(*self.backbone.layer3[1:])
res_g_conv5 = self.backbone.layer4
res_p_conv5 = nn.Sequential(
Bottleneck(1024,
512,
downsample=nn.Sequential(
nn.Conv2d(1024, 2048, 1, bias=False),
nn.BatchNorm2d(2048))), Bottleneck(2048, 512),
Bottleneck(2048, 512))
res_p_conv5.load_state_dict(self.backbone.layer4.state_dict())
self.p1 = nn.Sequential(copy.deepcopy(res_conv4),
copy.deepcopy(res_g_conv5))
self.p2 = nn.Sequential(copy.deepcopy(res_conv4),
copy.deepcopy(res_p_conv5))
self.p3 = nn.Sequential(copy.deepcopy(res_conv4),
copy.deepcopy(res_p_conv5))
self.maxpool_zg_p1 = nn.MaxPool2d(kernel_size=(12, 4))
self.maxpool_zg_p2 = nn.MaxPool2d(kernel_size=(24, 8))
self.maxpool_zg_p3 = nn.MaxPool2d(kernel_size=(24, 8))
self.maxpool_zp2 = nn.MaxPool2d(kernel_size=(12, 8))
self.maxpool_zp3 = nn.MaxPool2d(kernel_size=(8, 8))
self.reduction = nn.Sequential(
nn.Conv2d(2048 + 128, feats, 1, bias=True), nn.BatchNorm2d(feats),
nn.ReLU())
self._init_reduction(self.reduction)
self.fc_id_2048_0 = nn.Linear(feats, num_classes)
self.fc_id_2048_1 = nn.Linear(feats, num_classes)
self.fc_id_2048_2 = nn.Linear(feats, num_classes)
self.fc_id_256_1_0 = nn.Linear(feats, num_classes)
self.fc_id_256_1_1 = nn.Linear(feats, num_classes)
self.fc_id_256_2_0 = nn.Linear(feats, num_classes)
self.fc_id_256_2_1 = nn.Linear(feats, num_classes)
self.fc_id_256_2_2 = nn.Linear(feats, num_classes)
self._init_fc(self.fc_id_2048_0)
self._init_fc(self.fc_id_2048_1)
self._init_fc(self.fc_id_2048_2)
self._init_fc(self.fc_id_256_1_0)
self._init_fc(self.fc_id_256_1_1)
self._init_fc(self.fc_id_256_2_0)
self._init_fc(self.fc_id_256_2_1)
self._init_fc(self.fc_id_256_2_2)
def __init__(self):
super(ResNet, self).__init__()
cnn = resnet101(pretrained=True)
self.cnn = cnn
self.features = nn.Sequential(cnn.conv1, cnn.bn1, cnn.relu,
cnn.maxpool, cnn.layer1, cnn.layer2,
cnn.layer3, cnn.layer4)
def resnet101(pretrained=False, **kwargs):
"""Constructs a ResNet-101 model.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
"""
model = ResNet_C5(Bottleneck, [3, 4, 23, 3], **kwargs)
model_full = resnet.resnet101(pretrained=True)
if pretrained:
model.load_pretrained(model_full)
return model
def _get_backbone(self, backbone):
if backbone == 'resnet-101':
resnet = resnet101(pretrained=True)
extractor = nn.Sequential(resnet.conv1, resnet.bn1, resnet.relu,
resnet.maxpool, resnet.layer1,
resnet.layer2, resnet.layer3,
resnet.layer4)
return extractor
else:
raise NotImplementedError
def test_gflops():
from fblib.util.model_resources.flops import compute_gflops
from torchvision.models import resnet
x50 = resnext101_32x4d(pretrained=False)
print('GFLOPS for ResNeXt: {}'.format(compute_gflops(x50)))
res50 = resnet.resnet101(pretrained=False)
print('GFLOPS for ResNet: {}'.format(compute_gflops(res50)))
def __init__(self, return_layers=None):
model = resnet101(pretrained=True,
replace_stride_with_dilation=[False, True, True])
return_layers = {
'layer1': 'layer1',
'layer2': 'layer2',
'layer3': 'layer3',
'layer4': 'layer4',
} or return_layers
super(ResNet101, self).__init__(model, return_layers)
def _resnet101(msda='fmix', pretrained=False, *args, **kwargs):
from torchvision.models.resnet import resnet101
model = resnet101(*args, **kwargs)
if pretrained:
state = load_state_dict_from_url(
f'http://marc.ecs.soton.ac.uk/pytorch-models/imagenet/resnet101/imagenet_resnet101_{msda}.pt',
progress=True)
model.load_state_dict(state)
return model
def get_backbone(self, name, pretrained):
if name == "resne34":
return resnet34(pretrained=pretrained)
elif name == "resnet50":
return resnet50(pretrained=pretrained)
elif name == "resnet101":
return resnet101(pretrained=pretrained)
elif name == "resnet152":
return resnet152(pretrained=pretrained)
else:
return resnet18(pretrained=pretrained) #default
def get_base_model(name):
if name == "resnet18":
return resnet.resnet18()
if name == "resnet34":
return resnet.resnet34()
if name == "resnet50":
return resnet.resnet50()
if name == "resnet101":
return resnet.resnet101()
if name == "resnet152":
return resnet.resnet152()
def resnet101(pretrained=False, fixed_feature=True):
""" "ResNet-101 model from torchvision's resnet model.
:param pretrained: if true, return a model pretrained on ImageNet
:param fixed_feature: if true and pretrained is true, model features are fixed while training.
"""
from torchvision.models.resnet import resnet101
model = resnet101(pretrained)
ff = True if pretrained and fixed_feature else False
return _ResNet(model, ff)
def create_model(model_name, num_classes=1000, pretrained=False, **kwargs):
if 'test_time_pool' in kwargs:
test_time_pool = kwargs.pop('test_time_pool')
else:
test_time_pool = True
if 'extra' in kwargs:
extra = kwargs.pop('extra')
else:
extra = True
if model_name == 'dpn68':
model = dpn68(
num_classes=num_classes, pretrained=pretrained, test_time_pool=test_time_pool)
elif model_name == 'dpn68b':
model = dpn68b(
num_classes=num_classes, pretrained=pretrained, test_time_pool=test_time_pool)
elif model_name == 'dpn92':
model = dpn92(
num_classes=num_classes, pretrained=pretrained, test_time_pool=test_time_pool, extra=extra)
elif model_name == 'dpn98':
model = dpn98(
num_classes=num_classes, pretrained=pretrained, test_time_pool=test_time_pool)
elif model_name == 'dpn131':
model = dpn131(
num_classes=num_classes, pretrained=pretrained, test_time_pool=test_time_pool)
elif model_name == 'dpn107':
model = dpn107(
num_classes=num_classes, pretrained=pretrained, test_time_pool=test_time_pool)
elif model_name == 'resnet18':
model = resnet18(num_classes=num_classes, pretrained=pretrained, **kwargs)
elif model_name == 'resnet34':
model = resnet34(num_classes=num_classes, pretrained=pretrained, **kwargs)
elif model_name == 'resnet50':
model = resnet50(num_classes=num_classes, pretrained=pretrained, **kwargs)
elif model_name == 'resnet101':
model = resnet101(num_classes=num_classes, pretrained=pretrained, **kwargs)
elif model_name == 'resnet152':
model = resnet152(num_classes=num_classes, pretrained=pretrained, **kwargs)
elif model_name == 'densenet121':
model = densenet121(num_classes=num_classes, pretrained=pretrained, **kwargs)
elif model_name == 'densenet161':
model = densenet161(num_classes=num_classes, pretrained=pretrained, **kwargs)
elif model_name == 'densenet169':
model = densenet169(num_classes=num_classes, pretrained=pretrained, **kwargs)
elif model_name == 'densenet201':
model = densenet201(num_classes=num_classes, pretrained=pretrained, **kwargs)
elif model_name == 'inception_v3':
model = inception_v3(
num_classes=num_classes, pretrained=pretrained, transform_input=False, **kwargs)
else:
assert False, "Unknown model architecture (%s)" % model_name
return model