def __init__(self, local_channels=128, num_classes=None):
super(Model, self).__init__()
self.base1 = resnet50(pretrained=True)
self.base2 = xception(num_classes=num_classes)
planes = 2048
bn_relu = []
bn_relu += [nn.BatchNorm2d(planes)]
bn_relu += [nn.ReLU(inplace=True)]
bn_relu = nn.Sequential(*bn_relu)
bn_relu.apply(weights_init_kaiming)
self.bn_relu = bn_relu
local_layer = []
local_layer += [
nn.MaxPool2d(kernel_size=(1, 7), stride=(1, 7), padding=0)
]
local_layer += [nn.Conv2d(planes, local_channels, 1)]
local_layer = nn.Sequential(*local_layer)
local_layer.apply(weights_init_kaiming)
self.local_layer = local_layer
global_layer = []
global_layer += [nn.AvgPool2d(kernel_size=(7, 7), stride=(7, 7))]
global_layer = nn.Sequential(*global_layer)
self.global_layer = global_layer
if num_classes is not None:
classifier = []
classifier += [nn.Linear(planes, num_classes)]
classifier = nn.Sequential(*classifier)
classifier.apply(weights_init_classifier)
self.classifier = classifier
def __init__(self,
num_class=1,
input_channel=3,
output_stride=32,
layer=101):
super(resnet_backbone, self).__init__()
if layer == 101:
self.resnet = resnet101(pretrained=True,
output_stride=output_stride)
elif layer == 152:
self.resnet = resnet152(pretrained=True,
output_stride=output_stride)
elif layer == 50:
self.resnet = resnet50(pretrained=True,
output_stride=output_stride)
else:
raise ValueError("only support ResNet101 and ResNet152 now")
if input_channel == 1:
self.conv1 = nn.Conv2d(1,
64,
kernel_size=7,
stride=2,
padding=3,
bias=False)
elif input_channel == 3:
self.conv1 = self.resnet.conv1
else:
raise ValueError("input channel should be 3 or 1")
def __init__(self, topN=4, n_class=200, use_gpu=False):
super(attention_net, self).__init__()
self.n_class = n_class
self.use_gpu = use_gpu
self.pretrained_model = resnet.resnet50(pretrained=False)
self.pretrained_model.avgpool = nn.AdaptiveAvgPool2d(1)
self.pretrained_model.fc = nn.Linear(512 * 4, self.n_class)
self.proposal_net = ProposalNet()
self.topN = topN
self.concat_net = nn.Linear(2048 * (CAT_NUM + 1 + 1), self.n_class)
self.partcls_net = nn.Linear(512 * 4, self.n_class)
self.pad_side = 224
_, edge_anchors_small, _ = generate_default_anchor_maps(
setting='small')
self.edge_anchors_small = (edge_anchors_small + 224).astype(np.int)
_, edge_anchors_large, _ = generate_default_anchor_maps(
setting='large')
self.edge_anchors_large = (edge_anchors_large + 224).astype(np.int)
train_loader_t = DataLoader(dataset=train_dataset_t,
batch_size=int(batch_size / K),
shuffle=True,
num_workers=4,
drop_last=True)
test_dataset_t = imgdataset(dataset_dir=tar_dir_test,
txt_path=tar_annfile_test,
transformer='test')
test_loader_t = DataLoader(dataset=test_dataset_t,
batch_size=4,
shuffle=False,
num_workers=0)
########### MODEL ###########
model, param = resnet50(pretrained=imageNet_pretrain, num_classes=702)
model.cuda()
model = nn.DataParallel(model) #, device_ids=[0,1])
losses = Losses(K=K,
batch_size=batch_size,
bank_size=len(train_dataset_t),
ann_file=tar_annfile,
cam_num=num_cam)
losses = losses.cuda()
optimizer = torch.optim.SGD(param,
lr=base_lr,
momentum=0.9,
weight_decay=5e-4,
nesterov=True)
train_loader = DataLoader(dataset=train_dataset,
batch_size=1,
shuffle=False,
num_workers=4)
img_dir = dataset_path + 'Market-1501/'
test_dataset = imgdataset(dataset_dir=img_dir,
txt_path=ann_file_test,
transformer='test')
test_loader = DataLoader(dataset=test_dataset,
batch_size=1,
shuffle=False,
num_workers=4)
########### TEST ###########
model, _ = resnet50(pretrained=snapshot, num_classes=702)
model.cuda()
model.eval()
print('extract feature for training set')
train_feas = extract_fea_camtrans(model, train_loader)
_, cam_ids, frames = get_info(ann_file_train)
print('generate spatial-temporal distribution')
dist = cdist(train_feas, train_feas)
dist = np.power(dist, 2)
#dist = re_ranking(original_dist=dist)
labels = cluster(dist)
num_ids = len(set(labels))
print('cluster id num:', num_ids)
distribution = get_st_distribution(cam_ids,
def main(
architecture,
batch_size,
length_scale,
centroid_size,
learning_rate,
l_gradient_penalty,
gamma,
weight_decay,
final_model,
output_dir,
):
writer = SummaryWriter(log_dir=f"runs/{output_dir}")
ds = all_datasets["CIFAR10"]()
input_size, num_classes, dataset, test_dataset = ds
# Split up training set
idx = list(range(len(dataset)))
random.shuffle(idx)
if final_model:
train_dataset = dataset
val_dataset = test_dataset
else:
val_size = int(len(dataset) * 0.8)
train_dataset = torch.utils.data.Subset(dataset, idx[:val_size])
val_dataset = torch.utils.data.Subset(dataset, idx[val_size:])
val_dataset.transform = (test_dataset.transform
) # Test time preprocessing for validation
if architecture == "WRN":
model_output_size = 640
epochs = 200
milestones = [60, 120, 160]
feature_extractor = WideResNet()
elif architecture == "ResNet18":
model_output_size = 512
epochs = 200
milestones = [60, 120, 160]
feature_extractor = resnet18()
elif architecture == "ResNet50":
model_output_size = 2048
epochs = 200
milestones = [60, 120, 160]
feature_extractor = resnet50()
elif architecture == "ResNet110":
model_output_size = 2048
epochs = 200
milestones = [60, 120, 160]
feature_extractor = resnet110()
elif architecture == "DenseNet121":
model_output_size = 1024
epochs = 200
milestones = [60, 120, 160]
feature_extractor = densenet121()
# Adapted resnet from:
# https://github.com/kuangliu/pytorch-cifar/blob/master/models/resnet.py
feature_extractor.conv1 = torch.nn.Conv2d(3,
64,
kernel_size=3,
stride=1,
padding=1,
bias=False)
feature_extractor.maxpool = torch.nn.Identity()
feature_extractor.fc = torch.nn.Identity()
if centroid_size is None:
centroid_size = model_output_size
model = ResNet_DUQ(
feature_extractor,
num_classes,
centroid_size,
model_output_size,
length_scale,
gamma,
)
model = model.cuda()
optimizer = torch.optim.SGD(model.parameters(),
lr=learning_rate,
momentum=0.9,
weight_decay=weight_decay)
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
milestones=milestones,
gamma=0.2)
def calc_gradients_input(x, y_pred):
gradients = torch.autograd.grad(
outputs=y_pred,
inputs=x,
grad_outputs=torch.ones_like(y_pred),
create_graph=True,
)[0]
gradients = gradients.flatten(start_dim=1)
return gradients
def calc_gradient_penalty(x, y_pred):
gradients = calc_gradients_input(x, y_pred)
# L2 norm
grad_norm = gradients.norm(2, dim=1)
# Two sided penalty
gradient_penalty = ((grad_norm - 1)**2).mean()
return gradient_penalty
def step(engine, batch):
model.train()
optimizer.zero_grad()
x, y = batch
x, y = x.cuda(), y.cuda()
x.requires_grad_(True)
y_pred = model(x)
y = F.one_hot(y, num_classes).float()
loss = F.binary_cross_entropy(y_pred, y, reduction="mean")
if l_gradient_penalty > 0:
gp = calc_gradient_penalty(x, y_pred)
loss += l_gradient_penalty * gp
loss.backward()
optimizer.step()
x.requires_grad_(False)
with torch.no_grad():
model.eval()
model.update_embeddings(x, y)
return loss.item()
def eval_step(engine, batch):
model.eval()
x, y = batch
x, y = x.cuda(), y.cuda()
x.requires_grad_(True)
y_pred = model(x)
return {"x": x, "y": y, "y_pred": y_pred}
trainer = Engine(step)
evaluator = Engine(eval_step)
metric = Average()
metric.attach(trainer, "loss")
metric = Accuracy(output_transform=lambda out: (out["y_pred"], out["y"]))
metric.attach(evaluator, "accuracy")
def bce_output_transform(out):
return (out["y_pred"], F.one_hot(out["y"], num_classes).float())
metric = Loss(F.binary_cross_entropy,
output_transform=bce_output_transform)
metric.attach(evaluator, "bce")
metric = Loss(calc_gradient_penalty,
output_transform=lambda out: (out["x"], out["y_pred"]))
metric.attach(evaluator, "gradient_penalty")
pbar = ProgressBar(dynamic_ncols=True)
pbar.attach(trainer)
kwargs = {"num_workers": 4, "pin_memory": True}
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
drop_last=True,
**kwargs)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=batch_size,
shuffle=False,
**kwargs)
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=batch_size,
shuffle=False,
**kwargs)
@trainer.on(Events.EPOCH_COMPLETED)
def log_results(trainer):
metrics = trainer.state.metrics
loss = metrics["loss"]
print(f"Train - Epoch: {trainer.state.epoch} Loss: {loss:.2f}")
writer.add_scalar("Loss/train", loss, trainer.state.epoch)
if trainer.state.epoch > (epochs - 5):
accuracy, auroc = get_cifar_svhn_ood(model)
print(f"Test Accuracy: {accuracy}, AUROC: {auroc}")
writer.add_scalar("OoD/test_accuracy", accuracy,
trainer.state.epoch)
writer.add_scalar("OoD/roc_auc", auroc, trainer.state.epoch)
accuracy, auroc = get_auroc_classification(val_dataset, model)
print(f"AUROC - uncertainty: {auroc}")
writer.add_scalar("OoD/val_accuracy", accuracy,
trainer.state.epoch)
writer.add_scalar("OoD/roc_auc_classification", auroc,
trainer.state.epoch)
evaluator.run(val_loader)
metrics = evaluator.state.metrics
acc = metrics["accuracy"]
bce = metrics["bce"]
GP = metrics["gradient_penalty"]
loss = bce + l_gradient_penalty * GP
print((f"Valid - Epoch: {trainer.state.epoch} "
f"Acc: {acc:.4f} "
f"Loss: {loss:.2f} "
f"BCE: {bce:.2f} "
f"GP: {GP:.2f} "))
writer.add_scalar("Loss/valid", loss, trainer.state.epoch)
writer.add_scalar("BCE/valid", bce, trainer.state.epoch)
writer.add_scalar("GP/valid", GP, trainer.state.epoch)
writer.add_scalar("Accuracy/valid", acc, trainer.state.epoch)
scheduler.step()
trainer.run(train_loader, max_epochs=epochs)
evaluator.run(test_loader)
acc = evaluator.state.metrics["accuracy"]
print(f"Test - Accuracy {acc:.4f}")
torch.save(model.state_dict(), f"runs/{output_dir}/model.pt")
writer.close()
tar_annfile_test = 'list_{}/list_{}_test.txt'.format(opt.tar, opt.tar)
train_dataset_t = imgdataset_camtrans(dataset_dir=tar_dir,
txt_path=tar_annfile,
transformer='train',
num_cam=num_cam,
K=K)
train_loader_t = DataLoader(dataset=train_dataset_t,
batch_size=int(batch_size / K),
shuffle=True,
num_workers=4,
drop_last=True)
########### MODEL ###########
imageNet_pretrain = 'resnet50-19c8e357.pth'
model, param = resnet50(pretrained=imageNet_pretrain,
num_classes=numpids + len(train_dataset2.pids))
model.cuda()
model = CamDataParallel(model) #, device_ids=[0,1])
losses = Losses(K=K,
batch_size=batch_size,
bank_size=len(train_dataset_t),
ann_file=tar_annfile,
cam_num=num_cam)
losses = losses.cuda()
optimizer = torch.optim.SGD(param,
lr=base_lr,
momentum=0.9,
weight_decay=5e-4,
nesterov=True)