def __init__(self, load=True, width=256, height=256):
self.writer = SummaryWriter('runs/SCL_1')
self.spatial_features_size = 20
self.load = load
self.mod = False
self.reverse_margins = False
self.dataset = SCL_dataset(width, height, self.mod,
self.reverse_margins)
self.model = SCL_model(self.spatial_features_size, width,
height).to(device)
self.descriptor_model = Descriptor_net(self.spatial_features_size,
width, height).to(device)
self.contrastive_criterion = TripletLoss(self.mod,
self.reverse_margins)
# self.pixel_criterion=PixelwiseLoss(width, height)
self.pixel_criterion = PixelTripletLoss(width, height, self.mod)
self.load_from = "models/SCL.pth"
self.save_to = "models/SCL.pth"
self.d_load_from = "models/d_SCL.pth"
self.d_save_to = "models/d_SCL.pth"
if self.load:
self.model.load_state_dict(
torch.load(self.load_from, map_location=device))
self.model.eval()
self.delayed_model = SCL_model(self.spatial_features_size, width,
height).to(device)
self.delayed_model.load_state_dict(
torch.load(self.load_from, map_location=device))
self.delayed_model.eval()
self.descriptor_model.load_state_dict(
torch.load(self.d_load_from, map_location=device))
self.descriptor_model.eval()
self.d_delayed_model = Descriptor_net(self.spatial_features_size,
width, height).to(device)
self.d_delayed_model.load_state_dict(
torch.load(self.d_load_from, map_location=device))
self.d_delayed_model.eval()
self.tau = 0.001
self.optimizer = optim.Adam(self.model.parameters(),
lr=1e-4,
weight_decay=1e-4)
self.descriptor_optimizer = optim.Adam(
self.descriptor_model.parameters(), lr=1e-4, weight_decay=1e-4)
self.max_iter = 3500
self.dataloader = torch.utils.data.DataLoader(self.dataset,
1,
shuffle=True,
pin_memory=device)
self.descriptor_dataloader = torch.utils.data.DataLoader(
self.dataset, 1, shuffle=True, pin_memory=device)
self.images = self.dataset.frames[0]
self.embedding = torch.empty((len(self.images), 32))
self.cost_counter = 0
def train_model(triplet_train_loader, triplet_val_loader, test_fun, margin,
net_output, model_directory):
target_embedding_net = TargetEmbeddingNet(300, net_output)
triple_model = TripletNet(target_embedding_net, target_embedding_net)
if cuda:
triple_model.cuda()
loss_fn = TripletLoss(margin)
lr = 1e-4
optimizer = optim.Adam(triple_model.parameters(), lr=lr)
scheduler = lr_scheduler.StepLR(optimizer, 1000, gamma=0.5, last_epoch=-1)
n_epochs = 20
log_interval = 100
train_losses, val_losses, metrices = fit(triplet_train_loader,
triplet_val_loader,
triple_model,
loss_fn,
optimizer,
scheduler,
n_epochs,
cuda,
log_interval,
callback_test=test_fun,
keep_checkpoint_max=10,
model_dir=model_directory)
return metrices
def net_config(image, label, model, args, is_train):
assert args.model in model_list, "{} is not in lists: {}".format(
args.model, model_list)
out = model.net(input=image, embedding_size=args.embedding_size)
if not is_train:
return None, out
if args.loss_name == "triplet":
metricloss = TripletLoss(margin=args.margin, )
elif args.loss_name == "quadruplet":
metricloss = QuadrupletLoss(
train_batch_size=args.train_batch_size,
samples_each_class=args.samples_each_class,
margin=args.margin,
)
elif args.loss_name == "eml":
metricloss = EmlLoss(
train_batch_size=args.train_batch_size,
samples_each_class=args.samples_each_class,
)
elif args.loss_name == "npairs":
metricloss = NpairsLoss(
train_batch_size=args.train_batch_size,
samples_each_class=args.samples_each_class,
reg_lambda=args.npairs_reg_lambda,
)
cost = metricloss.loss(out, label)
avg_cost = fluid.layers.mean(x=cost)
return avg_cost, out
def run(train_loader, test_pairs, options):
# Construct model
cuda = options["device_id"] != -1
if options["use_pair_feature"]:
point_net = PointNet(options)
pair_net = PairNetWithPairFeatures(options, point_net)
model = TripletNetWithPairFeatures(pair_net)
else:
point_net = PointNet(options)
pair_net = PairNet(options, point_net)
model = TripletNet(pair_net)
if cuda:
model.cuda()
print(model)
optimizer = torch.optim.Adam(model.parameters(), lr=options["lr"], weight_decay=options["weight_decay"])
loss_fn = TripletLoss(options["margin"])
best_overall_metric = 0 # a single value
best_metrics = None # a dictionary
best_epoch = 0
save_model(model, options["save_dir"], 'best', 0) # save the initial first model
for epoch in range(options["epochs"]):
epoch_loss, non_zero_triplet_ratio = train_triplet_epoch(train_loader, model, loss_fn, optimizer, cuda,
use_pair_feature=options["use_pair_feature"])
print("Epoch: {}, train-loss: {:06.4f}, non_zero_triplet_ratio: {:06.4f}, ".format(epoch, epoch_loss,
non_zero_triplet_ratio))
if epoch % options["eval_epoch"] == 0 and epoch != 0:
if options["use_pair_feature"]:
prediction_score = pair_prediction_with_pair_feature(model.pair_net, test_pairs, pair_features, cuda, options, batch_size=10000)
else:
prediction_score = pair_prediction(model.pair_net, test_pairs, cuda, options, batch_size=10000)
test_triplets = []
for term_pair, score in zip(test_pairs, prediction_score):
test_triplets.append((term_pair[0], term_pair[1], -1.0 * score))
metrics = evaluation_main(test_triplets)
if metrics["all"] >= best_overall_metric:
best_overall_metric =metrics["all"]
best_epoch = epoch
best_metrics = metrics
save_model(model, options["save_dir"], 'best', epoch) # save the initial first model
return best_overall_metric, best_epoch, best_metrics
def __init__(self, settings, input_shape=(1, 28, 28), embedding_size=10, margin_loss=0.2):
super(EmbeddingNet, self).__init__()
in_channels = input_shape[0]
sequence = [nn.Conv2d(in_channels, 128, kernel_size=7, padding=1), # (28, 28)
nn.ReLU(inplace=True),
nn.MaxPool2d(2),
nn.Conv2d(128, 128, kernel_size=3, padding=1), # (11, 11)
nn.ReLU(inplace=True),
nn.MaxPool2d(2),
nn.Conv2d(128, 256, kernel_size=3, padding=1), # (5, 5)
nn.ReLU(inplace=True),
Flatten(), # (4, 4)
nn.Linear(6 * 6 * 256, 4096),
nn.ReLU(inplace=True),
nn.Linear(4096, embedding_size),
nn.LayerNorm(embedding_size, elementwise_affine=False)
]
self.model = nn.Sequential(*sequence)
self.model.apply(weight_init)
self.loss_fun = TripletLoss(margin_loss).triplet_loss
self.cum_loss = 0.0
self.scheduler, self.optimizer = self.init_optimizer(settings)
self.device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
def parse_parameters(params, optimizer, lr, loss, schedule=None):
optimizer = optimizer.lower()
if 'adam' in optimizer:
optimizer = optim.Adam(lr=lr, params=params, weight_decay=1e-5)
elif 'sgd' in optimizer:
optimizer = optim.SGD(lr=lr, momentum=0.9, params=params)
else:
print('optimizer {} not yet supported'.format(optimizer))
scheduler = None
if schedule is not None:
scheduler = optim.lr_scheduler.ExponentialLR(optimizer,
math.exp(schedule))
loss = loss.lower()
if 'cont' in loss:
criterion = ContrastiveLoss()
elif 'trip' in loss:
criterion = TripletLoss()
else:
print('loss {} not supported'.format(loss))
return (optimizer, criterion, scheduler)
def train(train_path, labels, boxes, output_dir, num_epochs, hard, verbose):
df_train, df_val = _get_toy_dataset(labels, boxes)
if verbose:
logging.info("Train size: {}, validation size: {}".format(
len(df_train), len(df_val)))
sampler = None
if hard:
sampler = BalancedBatchSampler(df_train, n_classes=4, n_samples=4)
train_dl, single_train_dl, val_dl = get_dataloaders(
df_train, df_val, train_path, sampler)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
if verbose:
logging.info("Using device {}".format(device))
net = TripletNetwork(embedding_size=128).to(device)
criterion = TripletLoss()
selector = None
if hard:
selector = HardBatchTripletSelector()
optimizer = optim.Adam(net.parameters(), lr=1e-4)
net, history = _train(
model=net,
optimizer=optimizer,
criterion=criterion,
train_dataloader=train_dl,
single_train_dataloader=single_train_dl,
val_dataloader=val_dl,
num_epochs=num_epochs,
save_path=output_dir,
device=device,
selector=selector,
)
_plot_history(history)
negative, negative_labels = load_data(NEGATIVE_DIR,
n_class=NUM_CLASS,
cls=5,
img_size=IMAGE_SIZE)
print("-" * 30)
print("Data loaded!!")
print("-" * 30)
# 3 models extracting the pattern feature
triplet_mdoel = TripletNet(CNNEmbeddingNetL2()).cuda()
# pattern_model = TripletNet(
# torch.hub.load('pytorch/vision:v0.9.0', 'mobilenet_v2', pretrained=True)
# ).cuda()
criterion = TripletLoss(margin=MARGIN)
# optimizer = optim.Adam(triplet_model.parameters(), lr=LR)
optimizer = optim.SGD(triplet_mdoel.embedding_net.parameters(),
lr=LR,
momentum=0.9)
"""Training Phase"""
""" Stage 1: Train the Embedding Network"""
triplet_model = train_triplet(triplet_mdoel,
criterion,
optimizer,
training,
training_labels,
negative,
n_epoch=N_EPOCH,
batch_size=TRAINING_BATCH_SIZE)
def main():
transforms_args = [
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
train_dataset = CoCoDataset(
args.coco_path,
"training",
target_size=args.target_size,
transform=transforms.Compose(
transforms_args +
[RandomErasing(probability=args.p, sh=args.sh, r1=args.r1)]))
test_dataset = CoCoDataset(args.coco_path,
"validation_wo_occlusion",
target_size=args.target_size,
transform=transforms.Compose(transforms_args))
train_batch_sampler = TrainBalancedBatchSampler(torch.from_numpy(
np.array(train_dataset.all_targets())),
K=args.K,
P=args.P,
n_batches=args.n_batches)
test_batch_sampler = TestBalancedBatchSampler(torch.from_numpy(
np.array(test_dataset.all_targets())),
K=args.K,
P=args.P,
n_batches=args.n_batches)
train_loader = DataLoader(train_dataset,
batch_sampler=train_batch_sampler,
**kwargs)
test_loader = DataLoader(test_dataset,
batch_sampler=test_batch_sampler,
**kwargs)
# init model
model, optim_state_dict, init_epoch = load_model(
args.backbone,
args.snapshot,
imagenet_weights=args.imagenet_weights,
freeze=args.freeze)
print("Resume training from epoch", init_epoch)
if cuda:
model.cuda()
# init optimizer
if args.optim == "Adam":
optimizer = optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=1e-4)
elif args.optim == "SGD":
optimizer = optim.SGD(model.parameters(),
momentum=0.9,
lr=args.lr,
weight_decay=1e-4)
else:
raise ValueError("Optimizer is not supported")
if optim_state_dict is not None:
optimizer.load_state_dict(optim_state_dict)
# define loss function
if args.triplet_selector == "hard":
selector = HardestNegativeTripletSelector(args.soft_margin)
elif args.triplet_selector == "semi":
selector = SemihardNegativeTripletSelector(args.soft_margin)
elif args.triplet_selector == "random":
selector = RandomNegativeTripletSelector(args.soft_margin)
else:
selector = AllTripletSelector()
train_loss_fn = TripletLoss(selector, soft_margin=args.soft_margin)
test_loss_fn = TripletLoss(AllTripletSelector(),
soft_margin=args.soft_margin)
# define learning rate scheduler
lr_scheduler = LrScheduler(args.epoch_decay_start, args.n_epoch, args.lr)
log_file = os.path.join(
args.logger_dir, '%s_%s.csv' % (args.backbone, args.triplet_selector))
for epoch in range(init_epoch + 1, args.n_epoch):
lr_scheduler.adjust_learning_rate(optimizer, epoch, args.optim)
for param_group in optimizer.param_groups:
print("LR: ", param_group['lr'])
train_loss = train_epoch(model, train_loader, train_loss_fn, optimizer,
cuda)
if epoch % args.eval_freq == 0:
test_loss = test_epoch(model, test_loader, test_loss_fn, cuda)
print('Epoch [%d/%d], Train loss: %.4f, Test loss: %.4f' %
(epoch, args.n_epoch, train_loss, test_loss))
log = [epoch, train_loss, test_loss]
if os.path.isfile(log_file):
with open(log_file, mode='a', newline='') as csv_f:
writer = csv.writer(csv_f)
writer.writerow(log)
else:
with open(log_file, mode='w', newline='') as csv_f:
writer = csv.writer(csv_f)
# write header
writer.writerow(["epoch", "train_loss", "test_loss"])
writer.writerow(log)
if epoch % args.save_freq == 0:
torch.save(
{
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'epoch': epoch
},
os.path.join(
args.snapshot_path, '%s_%s_%d.pth' %
(args.backbone, args.triplet_selector, epoch)))
# init the logger
init_logger(args)
# data loading
dataset, trainloader, queryloader, galleryloader = init_data_loaders(args)
num_train_pids = dataset.num_train_pids
# init model
model = init_model(args, num_train_pids)
if use_gpu:
model = nn.DataParallel(model).cuda()
vis = utils.get_visdom_for_current_run(args.save_dir, args.prefix + '_stage1_training')
# init objective functions
criterion_xent = CrossEntropyLabelSmooth(num_classes=num_train_pids, use_gpu=use_gpu)
criterion_htri = TripletLoss(margin=args.margin)
# init optimizer
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
if args.stepsize > 0:
scheduler = lr_scheduler.StepLR(optimizer, step_size=args.stepsize, gamma=args.gamma)
start_epoch = args.start_epoch
# if only evaluation was needed
if args.evaluate:
print("Evaluate only")
test(model, queryloader, galleryloader, use_gpu, args)
exit(0)
start_time = time.time()
def main():
import torch
from torch.optim import lr_scheduler
import torch.optim as optim
from torch.autograd import Variable
from trainer import fit
import numpy as np
cuda = torch.cuda.is_available()
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size',
type=int,
default=100,
metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size',
type=int,
default=100,
metavar='N',
help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs',
type=int,
default=100,
metavar='N',
help='number of epochs to train (default: 10)')
parser.add_argument('--lr',
type=float,
default=0.001,
metavar='LR',
help='learning rate (default: 0.01)')
parser.add_argument('--momentum',
type=float,
default=0.5,
metavar='M',
help='SGD momentum (default: 0.5)')
parser.add_argument('--no-cuda',
action='store_true',
default=False,
help='disables CUDA training')
parser.add_argument('--seed',
type=int,
default=1,
metavar='S',
help='random seed (default: 1)')
parser.add_argument(
'--log-interval',
type=int,
default=10,
metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--save-model',
action='store_true',
default=True,
help='For Saving the current Model')
args = parser.parse_args()
use_cuda = not args.no_cuda and torch.cuda.is_available()
torch.manual_seed(args.seed)
np.random.seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
from datetime import datetime
import os
loging = False
ismultitask = False
loso = False
if (args.save_model):
model_save_path = 'model/triplet/'
if not os.path.isdir(model_save_path):
os.makedirs(model_save_path)
if loging:
fname = model_save_path + datetime.today().strftime(
"%m_%d_%H_%M") + ".txt"
f = open(fname, 'w')
x_data, y_data = load_smt()
# nonbciilli = np.s_[0,1,2,4,5,8,16,17,18,20,21,27,28,29,30,32,35,36,38,42,43,44,51]
valtype = 'sess'
if valtype == 'loso':
for subj in range(0, 54):
model = Deep4CNN(ismult=ismultitask).to(device)
#model.load_state_dict(torch.load(model_save_path+ "J_" + str(subj) + 'basecnn.pt'))
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum)
optimizer_fine = optim.SGD(model.parameters(),
lr=0.005,
momentum=args.momentum)
dataset_train = GigaDataset(x=x_data,
y=y_data,
valtype=valtype,
istrain=True,
sess=1,
subj=subj)
train_loader = torch.utils.data.DataLoader(
dataset_train,
batch_size=args.batch_size,
shuffle=True,
**kwargs)
dataset_test = GigaDataset(x=x_data,
y=y_data,
valtype=valtype,
istrain=False,
sess=2,
subj=subj)
test_loader = torch.utils.data.DataLoader(
dataset_test,
batch_size=args.batch_size,
shuffle=False,
**kwargs)
# dataset_fine = GigaDataset_LOSO(x=x_data, y=y_data, fine=True, istrain=True, sess=2, subj=subj)
# fine_loader = torch.utils.data.DataLoader(dataset_fine, batch_size=args.batch_size, shuffle=True, **kwargs)
for epoch in range(1, args.epochs + 1):
train(args, model, device, train_loader, optimizer, epoch)
print("joint-train")
#LOSO joint training
j_loss, j_score = eval(args, model, device, test_loader)
if epoch > 30:
if (args.save_model):
torch.save(
model.state_dict(), model_save_path + "model_" +
str(subj) + "_" + str(epoch) + '.pt')
# #fine tuning
# for epoch in range(1, 10):
# train_mt(args, model, device, fine_loader, optimizer_fine, epoch)
#
# print("fine-tuning")
# f_loss, f_score = eval(args, model, device, test_loader)
if (args.save_model):
torch.save(model.state_dict(),
model_save_path + "F_" + str(subj) + 'basecnn.pt')
if loging:
f = open(fname, 'a')
f.write(
str(subj) + " " + "jl : " + str(j_loss) + " " +
str(j_score) + '\n')
f.close()
elif valtype == 'sess':
from networks import EmbeddingDeep4CNN, TripletNet, FineShallowCNN, EmbeddingDeepCNN
from losses import TripletLoss
margin = 1
embedding_net = EmbeddingDeep4CNN()
print(embedding_net)
model = TripletNet(embedding_net)
if cuda:
model.cuda()
loss_fn = TripletLoss(margin)
lr = 1e-3
#optimizer = optim.Adam(model.parameters(), lr=lr)
n_epochs = 5
#%%
log_interval = 10
if n_epochs == 0:
pass
#model.load_state_dict(torch.load('triplet_deep4_1000_2.pt'))
else: #트리플렛넷 학습
# For classification
dataset_train = GigaDataset(x=x_data,
y=y_data,
valtype=valtype,
istrain=True,
sess=1)
train_loader = torch.utils.data.DataLoader(
dataset_train,
batch_size=args.batch_size,
shuffle=True,
**kwargs)
dataset_test = GigaDataset(x=x_data,
y=y_data,
valtype=valtype,
istrain=False,
sess=2,
subj=-1)
test_loader = torch.utils.data.DataLoader(
dataset_test,
batch_size=args.batch_size,
shuffle=False,
**kwargs)
triplet_dataset_train = TripletGiga(x=x_data,
y=y_data,
valtype=valtype,
istrain=True,
sess=1)
triplet_train_loader = torch.utils.data.DataLoader(
triplet_dataset_train,
batch_size=args.batch_size,
shuffle=True,
**kwargs)
triplet_dataset_test = TripletGiga(x=x_data,
y=y_data,
valtype=valtype,
istrain=False,
sess=2,
subj=-1)
triplet_test_loader = torch.utils.data.DataLoader(
triplet_dataset_test,
batch_size=args.batch_size,
shuffle=False,
**kwargs)
optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)
scheduler = lr_scheduler.StepLR(optimizer,
8,
gamma=1,
last_epoch=-1)
from trainer import fit
fit(triplet_train_loader, triplet_test_loader, model, loss_fn,
optimizer, scheduler, n_epochs, cuda, log_interval)
#%%
train_embeddings_tl, train_labels_tl = extract_embeddings(
train_loader, embedding_net, 1000)
# plot_embeddings(train_embeddings_tl, train_labels_tl)
val_embeddings_tl, val_labels_tl = extract_embeddings(
test_loader, embedding_net, 1000)
# plot_embeddings(val_embeddings_tl, val_labels_tl)
# #
from sklearn.pipeline import Pipeline
from sklearn.discriminant_analysis import LinearDiscriminantAnalysis
from sklearn.model_selection import ShuffleSplit, cross_val_score
lda = LinearDiscriminantAnalysis()
lda.fit(train_embeddings_tl, train_labels_tl)
print(lda.score(val_embeddings_tl, val_labels_tl))
# from torchvision import datasets, models, transforms
# temp = model.embedding_net.children()
# newmodel = torch.nn.Sequential(*(list(model.embedding_net.children())[:]))
# for param in model.embedding_net.parameters():
# param.requires_grad = True
#newembedding_net = torch.nn.Sequential(*(list(model.embedding_net.children())[:]))
#
from sklearn.manifold import TSNE
tsne = TSNE(n_components=2, perplexity=30)
train_tsne = tsne.fit_transform(val_embeddings_tl)
plot_embeddings(train_tsne, val_labels_tl)
for param in model.embedding_net.parameters():
param.requires_grad = True
#embedding_net2 = EmbeddingDeep4CNN()
newmodel = nn.Sequential(model.embedding_net, nn.Linear(1000, 2),
nn.LogSoftmax(dim=1)).to(device)
print(newmodel)
#newmodel.fc_lr = nn.Linear(1000,2)
newmodel.to(device)
optimizer = optim.SGD(newmodel.parameters(), lr=0.01, momentum=0.9)
#optimizer = optim.Adam(newmodel.parameters())
for epoch in range(1, 20):
train(args, newmodel, device, train_loader, optimizer, epoch)
j_loss, j_score = eval(args, newmodel, device, test_loader)
if args.save_model:
torch.save(model.state_dict(), 'triplet_deep4_1000_2.pt')
def testseq(dataset_name, use_gpu):
dataset_root = './video2img/track1_sct_img_test_big/'
dataset = Graph_data_manager.AICityTrack2(root=dataset_root)
width = 224
height = 224
transform_train = T.Compose([
T.Random2DTranslation(height, width),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
transform_test = T.Compose([
T.Resize((height, width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = True if use_gpu else False
seq_len = 4
num_instance = 4
train_batch = 32
test_batch = 1
queryloader = DataLoader(
VideoDataset(dataset.query,
seq_len=seq_len,
sample='dense',
transform=transform_test),
batch_size=test_batch,
shuffle=False,
num_workers=4,
pin_memory=pin_memory,
drop_last=False,
)
arch = "resnet50ta"
pretrained_model = "./log/track12_ta224_checkpoint_ep500.pth.tar"
start_epoch = 0
print("Initializing model: {}".format(arch))
dataset.num_train_pids = 517
if arch == 'resnet503d':
model = resnet3d.resnet50(num_classes=dataset.num_train_pids,
sample_width=width,
sample_height=height,
sample_duration=seq_len)
if not os.path.exists(pretrained_model):
raise IOError(
"Can't find pretrained model: {}".format(pretrained_model))
print("Loading checkpoint from '{}'".format(pretrained_model))
checkpoint = torch.load(pretrained_model)
state_dict = {}
for key in checkpoint['state_dict']:
if 'fc' in key: continue
state_dict[key.partition("module.")
[2]] = checkpoint['state_dict'][key]
model.load_state_dict(state_dict, strict=False)
else:
if not os.path.exists(pretrained_model):
model = models.init_model(name=arch,
num_classes=dataset.num_train_pids,
loss={'xent', 'htri'})
else:
model = models.init_model(name=arch,
num_classes=dataset.num_train_pids,
loss={'xent', 'htri'})
checkpoint = torch.load(pretrained_model)
model.load_state_dict(checkpoint['state_dict'])
start_epoch = checkpoint['epoch'] + 1
print("Loaded checkpoint from '{}'".format(pretrained_model))
print("- start_epoch: {}\n- rank1: {}".format(
start_epoch, checkpoint['rank1']))
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
criterion_xent = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
criterion_htri = TripletLoss(margin=0.3)
lr = 0.0003
gamma = 0.1
stepsize = 200
weight_decay = 5e-04
optimizer = torch.optim.Adam(model.parameters(),
lr=lr,
weight_decay=weight_decay)
if stepsize > 0:
scheduler = lr_scheduler.StepLR(optimizer,
step_size=stepsize,
gamma=gamma)
start_epoch = start_epoch
if use_gpu:
model = nn.DataParallel(model).cuda()
test(model, queryloader, 'avg', use_gpu, dataset, -1, meta_data_tab=None)
def main():
args.save_dir = args.save_dir + '/' + args.arch
torch.manual_seed(args.seed)
# os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu:
use_gpu = False
# add data to save_dir
args.save_dir = args.save_dir + '_' + args.dataset + '_combined_multisteplr11'
if args.pretrained_model is not None:
args.save_dir = os.path.dirname(args.pretrained_model)
if not osp.exists(args.save_dir):
os.makedirs(args.save_dir)
log_name = 'test.log' if args.evaluate else 'train.log'
log_name += time.strftime('-%Y-%m-%d-%H-%M-%S')
sys.stdout = Logger(osp.join(args.save_dir, log_name))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_dataset(name=args.dataset)
print("Train Transforms: \n\
Random2DTranslation, \n\
RandomHorizontalFlip, \n\
ToTensor, \n\
normalize\
")
transform_train = T.Compose([
T.Random2DTranslation(args.height, args.width),
T.RandomHorizontalFlip(),
# T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
# T.RandomErasing(p=0.5, scale=(0.02, 0.4), ratio=(0.3, 3.3), value=[0.485, 0.456, 0.406])
])
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = True if use_gpu else False
trainloader = DataLoader(
VideoDataset(dataset.train, seq_len=args.seq_len,
sample=args.data_selection, transform=transform_train),
sampler=RandomIdentitySampler(
dataset.train, num_instances=args.num_instances),
batch_size=args.train_batch, num_workers=args.workers,
pin_memory=pin_memory, drop_last=True,
)
queryloader = DataLoader(
VideoDataset(dataset.query, seq_len=args.seq_len,
sample='dense', transform=transform_test),
batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
pin_memory=pin_memory, drop_last=False,
)
galleryloader = DataLoader(
VideoDataset(dataset.gallery, seq_len=args.seq_len,
sample='dense', transform=transform_test),
batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
pin_memory=pin_memory, drop_last=False,
)
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch, num_classes=dataset.num_train_pids, seq_len=args.seq_len)
# pretrained model loading
if args.pretrained_model is not None:
if not os.path.exists(args.pretrained_model):
raise IOError("Can't find pretrained model: {}".format(
args.pretrained_model))
print("Loading checkpoint from '{}'".format(args.pretrained_model))
pretrained_state = torch.load(args.pretrained_model)['state_dict']
print(len(pretrained_state), ' keys in pretrained model')
current_model_state = model.state_dict()
pretrained_state = {key: val
for key, val in pretrained_state.items()
if key in current_model_state and val.size() == current_model_state[key].size()}
print(len(pretrained_state),
' keys in pretrained model are available in current model')
current_model_state.update(pretrained_state)
model.load_state_dict(current_model_state)
print("Model size: {:.5f}M".format(sum(p.numel()
for p in model.parameters())/1000000.0))
if use_gpu:
model = nn.DataParallel(model).cuda()
criterion_xent = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
criterion_htri = TripletLoss(margin=args.margin)
optimizer = torch.optim.Adam(
model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
if args.stepsize > 0:
scheduler = lr_scheduler.StepLR(
optimizer, step_size=args.stepsize, gamma=args.gamma)
start_epoch = args.start_epoch
if args.evaluate:
print("Evaluate only")
test(model, queryloader, galleryloader, use_gpu)
return
start_time = time.time()
best_rank1 = -np.inf
is_first_time = True
for epoch in range(start_epoch, args.max_epoch):
eta_seconds = (time.time() - start_time) * (args.max_epoch - epoch) / max(epoch, 1)
eta_str = str(datetime.timedelta(seconds=int(eta_seconds)))
print("==> Epoch {}/{} \teta {}".format(epoch+1, args.max_epoch, eta_str))
train(model, criterion_xent, criterion_htri,
optimizer, trainloader, use_gpu)
if args.stepsize > 0:
scheduler.step()
rank1 = 'NA'
mAP = 'NA'
is_best = False
if args.eval_step > 0 and (epoch+1) % args.eval_step == 0 or (epoch+1) == args.max_epoch:
print("==> Test")
rank1, mAP = test(model, queryloader, galleryloader, use_gpu)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
# save the model as required
if (epoch+1) % args.save_step == 0:
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, is_best, osp.join(args.save_dir, args.save_prefix, 'model' + '.pth.tar-' + str(epoch+1)))
is_first_time = False
if not is_first_time:
utils.disable_all_print_once()
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
print("Finished. Total elapsed time (h:m:s): {}".format(elapsed))
def main():
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
else:
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_vid_dataset(root=args.root, name=args.dataset)
transform_train = T.Compose([
T.Random2DTranslation(args.height, args.width),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = True if use_gpu else False
# decompose tracklets into images for image-based training
new_train = []
for img_paths, pid, camid in dataset.train:
for img_path in img_paths:
new_train.append((img_path, pid, camid))
trainloader = DataLoader(
ImageDataset(new_train, transform=transform_train),
sampler=RandomIdentitySampler(new_train,
num_instances=args.num_instances),
batch_size=args.train_batch,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(
VideoDataset(dataset.query,
seq_len=args.seq_len,
sample='evenly',
transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
VideoDataset(dataset.gallery,
seq_len=args.seq_len,
sample='evenly',
transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dataset.num_train_pids,
loss={'xent', 'htri'})
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
criterion_xent = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
criterion_htri = TripletLoss(margin=args.margin)
optimizer = init_optim(args.optim, model.parameters(), args.lr,
args.weight_decay)
if args.stepsize > 0:
scheduler = lr_scheduler.StepLR(optimizer,
step_size=args.stepsize,
gamma=args.gamma)
start_epoch = args.start_epoch
if args.resume:
print("Loading checkpoint from '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
start_epoch = checkpoint['epoch']
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
test(model, queryloader, galleryloader, args.pool, use_gpu)
return
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
print("==> Start training")
for epoch in range(start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch, model, criterion_xent, criterion_htri, optimizer,
trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
if args.stepsize > 0: scheduler.step()
if args.eval_step > 0 and (epoch + 1) % args.eval_step == 0 or (
epoch + 1) == args.max_epoch:
print("==> Test")
rank1 = test(model, queryloader, galleryloader, args.pool, use_gpu)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint(
{
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, is_best,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
print("==> Best Rank-1 {:.1%}, achieved at epoch {}".format(
best_rank1, best_epoch))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
train_time = str(datetime.timedelta(seconds=train_time))
print(
"Finished. Total elapsed time (h:m:s): {}. Training time (h:m:s): {}.".
format(elapsed, train_time))
# determine the epoch of pretrained model
epoch = int((opt.pretrained_model.split("/")[-1]).split("-")[-1])
# if only evaluation is required
if opt.evaluate:
print("-- evaluate only")
test(epoch, model)
exit(0)
# Define optimizer and loss function
person_id_criterion = CrossEntropyLabelSmooth(
train_dataset.number_classes(), use_gpu=opt.cuda)
attribute_criterion = AttributeCriterion(attribute_choices,
CrossEntropyLabelSmooth)
triplet_criterion = TripletLoss(opt.margin)
optimizer = optim.Adam(model.parameters(), lr=opt.lr,
weight_decay=5e-4) # Default lr = 3e-4
print("Using triplet loss = ", triplet_criterion)
print("Using person_id = ", person_id_criterion)
print("Using Attribute loss = ", attribute_criterion)
print("Optimizer = ", optimizer)
# scheduler creation
lr_scheduler, num_epochs = create_scheduler(opt, optimizer)
if epoch > 0:
lr_scheduler.step(epoch)
print("Scheduled epochs: ", num_epochs)
print("learning rates ",
def main():
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
else:
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'))
# tensorboardX
# writer = SummaryWriter(log_dir=osp.join(args.save_dir,'summary'))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_img_dataset(
root=args.root, name=args.dataset, split_id=args.split_id,
cuhk03_labeled=args.cuhk03_labeled, cuhk03_classic_split=args.cuhk03_classic_split,
)
transform_train = T.Compose([
T.Random2DTranslation(args.height, args.width),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
if args.random_erasing:
transform_train = T.Compose([
T.Random2DTranslation(args.height, args.width),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
RandomErasing(probability=args.probability, mean=[0.0, 0.0, 0.0]),
])
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = True if use_gpu else False
if args.loss == 'xent,htri':
trainloader = DataLoader(
ImageDataset(dataset.train, transform=transform_train),
sampler=RandomIdentitySampler(dataset.train, num_instances=args.num_instances),
batch_size=args.train_batch, num_workers=args.workers,
pin_memory=pin_memory, drop_last=True,
)
elif args.loss == 'xent':
trainloader = DataLoader(
ImageDataset(dataset.train, transform=transform_train),
batch_size=args.train_batch, shuffle=True, num_workers=args.workers,
pin_memory=pin_memory, drop_last=True,
)
queryloader = DataLoader(
ImageDataset(dataset.query, transform=transform_test),
batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
pin_memory=pin_memory, drop_last=False,
)
galleryloader = DataLoader(
ImageDataset(dataset.gallery, transform=transform_test),
batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
pin_memory=pin_memory, drop_last=False,
)
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch, num_classes=dataset.num_train_pids, loss=args.loss)
print("Model size: {:.5f}M".format(sum(p.numel() for p in model.parameters())/1000000.0))
criterion_xent = CrossEntropyLabelSmooth(num_classes=dataset.num_train_pids, use_gpu=use_gpu)
criterion_htri = TripletLoss(margin=args.margin)
optimizer = init_optim(args.optim, model.parameters(), args.lr, args.weight_decay)
if args.stepsize > 0:
if not args.warmup:
scheduler = lr_scheduler.StepLR(optimizer, step_size=args.stepsize, gamma=args.gamma)
start_epoch = args.start_epoch
if args.resume:
print("Loading checkpoint from '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
start_epoch = checkpoint['epoch']
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
test(model, queryloader, galleryloader, use_gpu)
return
def adjust_lr(optimizer, ep):
if ep < 20:
lr = 1e-4 * (ep + 1) / 2
elif ep < 80:
#lr = 1e-3 * len(args.gpu_devices)
lr = 1e-3
elif ep < 180:
#lr = 1e-4 * len(args.gpu_devices)
lr = 1e-4
elif ep < 300:
#lr = 1e-5 * len(args.gpu_devices)
lr = 1e-5
elif ep < 320:
#lr = 1e-5 * 0.1 ** ((ep - 320) / 80) * len(args.gpu_devices)
lr = 1e-5 * 0.1 ** ((ep - 320) / 80)
elif ep < 400:
lr = 1e-6
elif ep < 480:
#lr = 1e-4 * len(args.gpu_devices)
lr = 1e-4
else:
#lr = 1e-5 * len(args.gpu_devices)
lr = 1e-5
for p in optimizer.param_groups:
p['lr'] = lr
length = len(trainloader)
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
#best_rerank1 = -np.inf
#best_rerankepoch = 0
print("==> Start training")
for epoch in range(start_epoch, args.max_epoch):
start_train_time = time.time()
if args.stepsize > 0:
if args.warmup:
adjust_lr(optimizer, epoch + 1)
else:
scheduler.step()
train(epoch, model, criterion_xent, criterion_htri, optimizer, trainloader, use_gpu=use_gpu, summary=None, length=length)
train_time += round(time.time() - start_train_time)
if (epoch+1) > args.start_eval and args.eval_step > 0 and (epoch+1) % args.eval_step == 0 or (epoch+1) == args.max_epoch:
print("==> Test")
rank1 = test(epoch, model, queryloader, galleryloader, use_gpu=True, summary=None)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
####### Best Rerank
#is_rerankbest = rerank1 > best_rerank1
#if is_rerankbest:
# best_rerank1 = rerank1
# best_rerankepoch = epoch + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, is_best, osp.join(args.save_dir, 'checkpoint_ep' + str(epoch+1) + '.pth.tar'))
writer.close()
print("==> Best Rank-1 {:.1%}, achieved at epoch {}".format(best_rank1, best_epoch))
#print("==> Best Rerank-1 {:.1%}, achieved at epoch {}".format(best_rerank1, best_rerankepoch))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
train_time = str(datetime.timedelta(seconds=train_time))
print("Finished. Total elapsed time (h:m:s): {}. Training time (h:m:s): {}.".format(elapsed, train_time))
def main():
# 为了看看repo提供的model.pth.tar-9在validation集的mAp和rank-1
# 我自己训练的tar-9只有mAP: 15.1%； Rank-1: 23.3% ，不知道为什么
# 更改args.load_weights = '/model/caohw9/track3_model/model.pth.tar-9'
global args
print(args)
set_random_seed(args.seed)
if not args.use_avai_gpus:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu:
use_gpu = False
sys.stdout = Logger(osp.join(args.save_dir, "log.txt"))
if use_gpu:
print('Currently using GPU {}'.format(args.gpu_devices))
cudnn.benchmark = True
else:
warnings.warn(
'Currently using CPU, however, GPU is highly recommended')
# 初始化loader
print('Initializing image data manager')
dm = ImageDataManager(use_gpu, **trainset_kwargs(args))
trainloader, testloader_dict = dm.return_dataloaders(
) #trainloader用于训练，testloader_dict包含['query']和['gallery']2个loader
print('suffessfully initialize loaders!')
# 初始化模型
print('Initializing model: {}'.format(
args.arch)) #args.arch default='resnet101'
model = models.init_model(name=args.arch,
num_classes=dm.num_train_pids,
loss={'xent', 'htri'},
pretrained=not args.no_pretrained,
use_gpu=use_gpu)
print('Model size: {:.3f} M'.format(count_num_param(model)))
# 加载预训练参数
if args.load_weights and check_isfile(args.load_weights):
load_pretrained_weights(model, args.load_weights)
#加载训练过的模型后，先看看validation
print('=> Validation')
print('Evaluating {} ...'.format(
args.test_set)) #args.test_set应该是指的validation set？
queryloader = testloader_dict['query']
galleryloader = testloader_dict['test']
model = nn.DataParallel(model).cuda() if use_gpu else model
rank1 = test(model, queryloader, galleryloader, use_gpu) #validation！
# 多GPU训练
else:
model = nn.DataParallel(model).cuda() if use_gpu else model
# 定义loss,optimizer, lr_scheduler
criterion_xent = CrossEntropyLoss(num_classes=dm.num_train_pids,
use_gpu=use_gpu,
label_smooth=args.label_smooth)
criterion_htri = TripletLoss(margin=args.margin)
optimizer = init_optimizer(model, **optimizer_kwargs(args))
scheduler = init_lr_scheduler(optimizer, **lr_scheduler_kwargs(args))
# 是否是resume训练
if args.resume and check_isfile(args.resume):
args.start_epoch = resume_from_checkpoint(
args.resume, model, optimizer=optimizer) #获取中断时刻的epoch数
# 开始训练！
time_start = time.time()
print('=> Start training')
for epoch in range(args.start_epoch, args.max_epoch):
train(epoch, model, criterion_xent, criterion_htri, optimizer,
trainloader, use_gpu) #训练
scheduler.step() #更新lr
# 当epoch数超过args.start_eval，每隔一定频率args.eval_freq，或者达到最后一个epoch，进行validation+存储checkpoint
if (epoch + 1) > args.start_eval and args.eval_freq > 0 and (
epoch + 1) % args.eval_freq == 0 or (epoch +
1) == args.max_epoch:
print('=> Validation')
print('Evaluating {} ...'.format(
args.test_set)) #args.test_set应该是指的validation set？
queryloader = testloader_dict['query']
galleryloader = testloader_dict['test']
rank1 = test(model, queryloader, galleryloader,
use_gpu) #validation！
save_checkpoint(
{
'state_dict': model.state_dict(), #模型的状态字典
'rank1': rank1,
'epoch': epoch + 1,
'arch': args.arch, #default='resnet101'
'optimizer': optimizer.state_dict(
), #优化器对象的状态字典，包含优化器的状态和超参数(如lr, momentum,weight_decay等)
},
args.save_dir) #validation同时保存checkpoint
# 训练结束！
elapsed = round(time.time() - time_start) #持续时间
elapsed = str(datetime.timedelta(seconds=elapsed))
print('Elapsed {}'.format(elapsed))
def run_experiment(sim_type,
network_type,
dataset,
mus,
latent_dims,
batch_size=64,
keys=None,
composite_labels=None,
margin=1,
dropout_val=0.2,
num_iters=20000,
normalize=True,
feature_extractor='hog'):
"""Overarching function that runs an experiment. Calls _experiment for each
combination of mu and latent dim.
Args:
sim_type (str): Simulation type. Accepts either "pairwise" or "triplet"
network_type (str): Network type. Accepts either "embenc" or "vae"
dataset (str): Dataset name. Accepts "mnist", "cifar10" and "imagenet"
mus (list): list of mu values for which experiment is to be run.
latent_dims (list): list of latent dimensions for which experiment is to be run.
batch_size (int, optional): batch size from which pairs/triples are generated. Defaults to 64.
keys (dict, optional): Mapping from numeric labels to string label names. Defaults to None.
composite_labels (dict, optional): Maps current labels to composite labels. Defaults to None.
margin (int, optional): Margin to be used for metric loss. Defaults to 1.
dropout_val (float, optional): dropout value for networks. Defaults to 0.2.
num_iters (int, optional): number of iterations. Defaults to 20000.
Returns:
dict: mapping from (mu, latent_dim) to results for _experiment on those params.
"""
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print("loading data...", end='')
trainX, trainY, testX, testY, trainLabels, testLabels = load_data(
dataset,
composite_labels=composite_labels,
normalize=normalize,
feature_extractor=feature_extractor)
print("done.")
# instantiate the generator and the networks based on type.
print("initializing data generators and losses...", end='')
input_dim = trainX.shape[-1]
if sim_type == 'pairwise':
datagen = PairGenerator(trainX, trainY, batch_size)
metric_loss = ContrastiveLoss(
margin) if network_type != 'vae' else PairwiseVAELoss(margin)
elif sim_type == 'triplet':
datagen = TripletGenerator(trainX, trainY, batch_size)
metric_loss = TripletLoss(
margin) if network_type != 'vae' else TripletVAELoss(margin)
reconstruction_loss = reconstruction_error
print("done.")
all_results = {}
for mu in mus:
for latent_dim in latent_dims:
network = initialize_network(sim_type,
network_type,
input_dim,
latent_dim,
dropout_val,
device=device)
results = _experiment(network,
latent_dim,
datagen,
metric_loss,
reconstruction_loss,
trainX,
trainY,
testX,
testY,
keys,
mu,
dataset,
trainLabels,
testLabels,
num_iters=num_iters,
batch_size=batch_size,
device=device)
all_results[(mu, latent_dim)] = results
return all_results
class SCL_trainer:
def __init__(self, load=True, width=256, height=256):
self.writer = SummaryWriter('runs/SCL_1')
self.spatial_features_size = 20
self.load = load
self.mod = False
self.reverse_margins = False
self.dataset = SCL_dataset(width, height, self.mod,
self.reverse_margins)
self.model = SCL_model(self.spatial_features_size, width,
height).to(device)
self.descriptor_model = Descriptor_net(self.spatial_features_size,
width, height).to(device)
self.contrastive_criterion = TripletLoss(self.mod,
self.reverse_margins)
# self.pixel_criterion=PixelwiseLoss(width, height)
self.pixel_criterion = PixelTripletLoss(width, height, self.mod)
self.load_from = "models/SCL.pth"
self.save_to = "models/SCL.pth"
self.d_load_from = "models/d_SCL.pth"
self.d_save_to = "models/d_SCL.pth"
if self.load:
self.model.load_state_dict(
torch.load(self.load_from, map_location=device))
self.model.eval()
self.delayed_model = SCL_model(self.spatial_features_size, width,
height).to(device)
self.delayed_model.load_state_dict(
torch.load(self.load_from, map_location=device))
self.delayed_model.eval()
self.descriptor_model.load_state_dict(
torch.load(self.d_load_from, map_location=device))
self.descriptor_model.eval()
self.d_delayed_model = Descriptor_net(self.spatial_features_size,
width, height).to(device)
self.d_delayed_model.load_state_dict(
torch.load(self.d_load_from, map_location=device))
self.d_delayed_model.eval()
self.tau = 0.001
self.optimizer = optim.Adam(self.model.parameters(),
lr=1e-4,
weight_decay=1e-4)
self.descriptor_optimizer = optim.Adam(
self.descriptor_model.parameters(), lr=1e-4, weight_decay=1e-4)
self.max_iter = 3500
self.dataloader = torch.utils.data.DataLoader(self.dataset,
1,
shuffle=True,
pin_memory=device)
self.descriptor_dataloader = torch.utils.data.DataLoader(
self.dataset, 1, shuffle=True, pin_memory=device)
self.images = self.dataset.frames[0]
self.embedding = torch.empty((len(self.images), 32))
self.cost_counter = 0
def run_training(self, iterations=100):
self.dataset.count = 0
i = 0
for epoch in tqdm(range(iterations)):
for data in self.dataloader:
step = i
# inputs
data = data.permute(1, 0, 2, 3, 4).to(device)
anchor, pos, neg = data
# outputs
anchor_embedding, spatial_features_anchor = self.model(anchor)
pos_embedding, spatial_features_pos = self.model(pos)
neg_embedding, spatial_features_neg = self.model(neg)
loss = self.contrastive_criterion(anchor_embedding,
pos_embedding, neg_embedding)
# optimization
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# plot loss
# normalized_loss=loss.item()/self.contrastive_criterion.margins[self.contrastive_criterion.margin_step]
self.writer.add_scalar("training_loss", loss.item(), step)
# update margin - step the scheduler
if (i + 1) % 1000 == 0 and i > 2000:
self.dataset.update_margins()
self.contrastive_criterion.update_margin()
if (i + 1) % 1000 == 0 or i == 0:
for j in range(len(self.images)):
img = self.dataset.transform(
self.images[j]).unsqueeze(0).to(device)
with torch.no_grad():
self.embedding[j], _ = self.model(img)
for j in range(len(self.images)):
dis = torch.abs(self.embedding[-1] -
self.embedding[j]).pow(2).sum(-1)
cost = -dis
self.writer.add_scalar("cost_%d" % self.cost_counter,
cost, j)
self.writer.add_embedding(self.embedding, global_step=step)
self.cost_counter += 1
i += 1
# soft update
if self.load:
for param1, param2 in zip(self.model.parameters(),
self.delayed_model.parameters()):
param2.data.copy_(self.tau * param1.data +
(1 - self.tau) * param2.data)
param1.data.copy_(param2)
# save the model and visualize the embeddings
torch.save(self.model.state_dict(), self.save_to)
def train_descriptors(self, iterations=150):
self.dataset.count = 0
i = 0
for epoch in tqdm(range(iterations)):
for data in self.descriptor_dataloader:
step = i
# inputs
data = data.permute(1, 0, 2, 3, 4).to(device)
img1, img2, _ = data
# outputs
embedding1, spatial_features1 = self.model(img1)
embedding2, spatial_features2 = self.model(img2)
# to be used in pixel loss
img1_out = self.descriptor_model(spatial_features1)
vis1 = self.dataset.unormalize(img1.squeeze())
vis2 = self.dataset.unormalize(img1_out.squeeze())
img1_un = vis1.permute(1, 2, 0)
img2_out = self.descriptor_model(spatial_features2)
img2_un = self.dataset.unormalize(img2.squeeze()).permute(
1, 2, 0)
if (i + 1) % 1000 == 0:
self.writer.add_image(
"original vs. prediction",
torchvision.utils.make_grid([vis1, vis2]), step)
if i > 5000:
self.dataset.update_margins()
self.pixel_criterion.update_margins()
# loss computation
loss_p = self.pixel_criterion(img1_un, img1_out, img2_un,
img2_out)
# optimization
self.descriptor_optimizer.zero_grad()
loss_p.backward()
self.descriptor_optimizer.step()
# plot loss
# normalized_loss=loss_p.item()/self.pixel_criterion.margins[self.pixel_criterion.margin_step]
self.writer.add_scalar("training_descriptor_loss",
loss_p.item(), step)
torch.save(self.descriptor_model.state_dict(), self.d_save_to)
i += 1
# soft update
if self.load:
for param1, param2 in zip(self.descriptor_model.parameters(),
self.d_delayed_model.parameters()):
param2.data.copy_(self.tau * param1.data +
(1 - self.tau) * param2.data)
param1.data.copy_(param2)
def main():
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
else:
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_dataset(name=args.dataset)
transform_train = T.Compose([
T.Random2DTranslation(args.height, args.width),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = False
trainloader = DataLoader(
VideoDataset(dataset.train,
seq_len=args.seq_len,
sample='random',
transform=transform_train),
sampler=RandomIdentitySampler(dataset.train,
num_instances=args.num_instances),
batch_size=args.train_batch,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(
VideoDataset(dataset.query,
seq_len=args.seq_len,
sample='dense',
transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
VideoDataset(dataset.gallery,
seq_len=args.seq_len,
sample='dense',
transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dataset.num_train_pids,
loss={'xent', 'htri'})
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
criterion_xent = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
criterion_htri = TripletLoss(margin=args.margin)
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
if args.stepsize > 0:
scheduler = lr_scheduler.StepLR(optimizer,
step_size=args.stepsize,
gamma=args.gamma)
start_epoch = args.start_epoch
start_time = time.time()
print(start_time)
for batch_idx, (imgs, pids, _) in enumerate(trainloader):
print(batch_idx)
print('x')
if use_gpu:
imgs, pids = imgs.cuda(), pids.cuda()
imgs, pids = Variable(imgs), Variable(pids)
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
print("Finished. Total elapsed time (h:m:s): {}".format(elapsed))
def main():
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
else:
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_imgreid_dataset(
root=args.root,
name=args.dataset,
split_id=args.split_id,
cuhk03_labeled=args.cuhk03_labeled,
cuhk03_classic_split=args.cuhk03_classic_split,
use_lmdb=args.use_lmdb,
)
transform_train = T.Compose([
T.Random2DTranslation(args.height, args.width),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = True if use_gpu else False
trainloader = DataLoader(
ImageDataset(dataset.train,
transform=transform_train,
use_lmdb=args.use_lmdb,
lmdb_path=dataset.train_lmdb_path),
sampler=RandomIdentitySampler(dataset.train,
num_instances=args.num_instances),
batch_size=args.train_batch,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(
ImageDataset(dataset.query,
transform=transform_test,
use_lmdb=args.use_lmdb,
lmdb_path=dataset.train_lmdb_path),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
ImageDataset(dataset.gallery,
transform=transform_test,
use_lmdb=args.use_lmdb,
lmdb_path=dataset.train_lmdb_path),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dataset.num_train_pids,
loss={'xent', 'htri'})
print("Model size: {:.3f} M".format(count_num_param(model)))
criterion_xent = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
criterion_htri = TripletLoss(margin=args.margin)
optimizer = init_optim(args.optim, model.parameters(), args.lr,
args.weight_decay)
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=args.stepsize,
gamma=args.gamma)
if args.load_weights:
# load pretrained weights but ignore layers that don't match in size
if check_isfile(args.load_weights):
checkpoint = torch.load(args.load_weights)
pretrain_dict = checkpoint['state_dict']
model_dict = model.state_dict()
pretrain_dict = {
k: v
for k, v in pretrain_dict.items()
if k in model_dict and model_dict[k].size() == v.size()
}
model_dict.update(pretrain_dict)
model.load_state_dict(model_dict)
print("Loaded pretrained weights from '{}'".format(
args.load_weights))
if args.resume:
if check_isfile(args.resume):
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
args.start_epoch = checkpoint['epoch']
rank1 = checkpoint['rank1']
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(
args.start_epoch, rank1))
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
distmat = test(model,
queryloader,
galleryloader,
use_gpu,
return_distmat=True)
if args.vis_ranked_res:
visualize_ranked_results(
distmat,
dataset,
save_dir=osp.join(args.save_dir, 'ranked_results'),
topk=20,
)
return
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
print("==> Start training")
for epoch in range(args.start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch, model, criterion_xent, criterion_htri, optimizer,
trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
scheduler.step()
if (epoch + 1) > args.start_eval and args.eval_step > 0 and (
epoch + 1) % args.eval_step == 0 or (epoch +
1) == args.max_epoch:
print("==> Test")
rank1 = test(model, queryloader, galleryloader, use_gpu)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint(
{
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, is_best,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
print("==> Best Rank-1 {:.1%}, achieved at epoch {}".format(
best_rank1, best_epoch))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
train_time = str(datetime.timedelta(seconds=train_time))
print(
"Finished. Total elapsed time (h:m:s): {}. Training time (h:m:s): {}.".
format(elapsed, train_time))
def main():
runId = datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
cfg.OUTPUT_DIR = os.path.join(cfg.OUTPUT_DIR, runId)
if not os.path.exists(cfg.OUTPUT_DIR):
os.mkdir(cfg.OUTPUT_DIR)
print(cfg.OUTPUT_DIR)
torch.manual_seed(cfg.RANDOM_SEED)
random.seed(cfg.RANDOM_SEED)
np.random.seed(cfg.RANDOM_SEED)
os.environ['CUDA_VISIBLE_DEVICES'] = cfg.MODEL.DEVICE_ID
use_gpu = torch.cuda.is_available() and cfg.MODEL.DEVICE == "cuda"
if not cfg.EVALUATE_ONLY:
sys.stdout = Logger(osp.join(cfg.OUTPUT_DIR, 'log_train.txt'))
else:
sys.stdout = Logger(osp.join(cfg.OUTPUT_DIR, 'log_test.txt'))
print("==========\nConfigs:{}\n==========".format(cfg))
if use_gpu:
print("Currently using GPU {}".format(cfg.MODEL.DEVICE_ID))
cudnn.benchmark = True
torch.cuda.manual_seed_all(cfg.RANDOM_SEED)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(cfg.DATASETS.NAME))
dataset = data_manager.init_dataset(root=cfg.DATASETS.ROOT_DIR,
name=cfg.DATASETS.NAME)
print("Initializing model: {}".format(cfg.MODEL.NAME))
if cfg.MODEL.ARCH == 'video_baseline':
torch.backends.cudnn.benchmark = False
model = models.init_model(name=cfg.MODEL.ARCH,
num_classes=625,
pretrain_choice=cfg.MODEL.PRETRAIN_CHOICE,
last_stride=cfg.MODEL.LAST_STRIDE,
neck=cfg.MODEL.NECK,
model_name=cfg.MODEL.NAME,
neck_feat=cfg.TEST.NECK_FEAT,
model_path=cfg.MODEL.PRETRAIN_PATH)
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
transform_train = T.Compose([
T.Resize(cfg.INPUT.SIZE_TRAIN),
T.RandomHorizontalFlip(p=cfg.INPUT.PROB),
T.Pad(cfg.INPUT.PADDING),
T.RandomCrop(cfg.INPUT.SIZE_TRAIN),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
T.RandomErasing(probability=cfg.INPUT.RE_PROB,
mean=cfg.INPUT.PIXEL_MEAN)
])
transform_test = T.Compose([
T.Resize(cfg.INPUT.SIZE_TEST),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
pin_memory = True if use_gpu else False
cfg.DATALOADER.NUM_WORKERS = 0
trainloader = DataLoader(VideoDataset(
dataset.train,
seq_len=cfg.DATASETS.SEQ_LEN,
sample=cfg.DATASETS.TRAIN_SAMPLE_METHOD,
transform=transform_train,
dataset_name=cfg.DATASETS.NAME),
sampler=RandomIdentitySampler(
dataset.train,
num_instances=cfg.DATALOADER.NUM_INSTANCE),
batch_size=cfg.SOLVER.SEQS_PER_BATCH,
num_workers=cfg.DATALOADER.NUM_WORKERS,
pin_memory=pin_memory,
drop_last=True)
queryloader = DataLoader(VideoDataset(
dataset.query,
seq_len=cfg.DATASETS.SEQ_LEN,
sample=cfg.DATASETS.TEST_SAMPLE_METHOD,
transform=transform_test,
max_seq_len=cfg.DATASETS.TEST_MAX_SEQ_NUM,
dataset_name=cfg.DATASETS.NAME),
batch_size=cfg.TEST.SEQS_PER_BATCH,
shuffle=False,
num_workers=cfg.DATALOADER.NUM_WORKERS,
pin_memory=pin_memory,
drop_last=False)
galleryloader = DataLoader(
VideoDataset(dataset.gallery,
seq_len=cfg.DATASETS.SEQ_LEN,
sample=cfg.DATASETS.TEST_SAMPLE_METHOD,
transform=transform_test,
max_seq_len=cfg.DATASETS.TEST_MAX_SEQ_NUM,
dataset_name=cfg.DATASETS.NAME),
batch_size=cfg.TEST.SEQS_PER_BATCH,
shuffle=False,
num_workers=cfg.DATALOADER.NUM_WORKERS,
pin_memory=pin_memory,
drop_last=False,
)
if cfg.MODEL.SYN_BN:
if use_gpu:
model = nn.DataParallel(model)
if cfg.SOLVER.FP_16:
model = apex.parallel.convert_syncbn_model(model)
model.cuda()
start_time = time.time()
xent = CrossEntropyLabelSmooth(num_classes=dataset.num_train_pids)
tent = TripletLoss(cfg.SOLVER.MARGIN)
optimizer = make_optimizer(cfg, model)
scheduler = WarmupMultiStepLR(optimizer, cfg.SOLVER.STEPS,
cfg.SOLVER.GAMMA, cfg.SOLVER.WARMUP_FACTOR,
cfg.SOLVER.WARMUP_ITERS,
cfg.SOLVER.WARMUP_METHOD)
# metrics = test(model, queryloader, galleryloader, cfg.TEST.TEMPORAL_POOL_METHOD, use_gpu)
no_rise = 0
best_rank1 = 0
start_epoch = 0
for epoch in range(start_epoch, cfg.SOLVER.MAX_EPOCHS):
# if no_rise == 10:
# break
scheduler.step()
print("noriase:", no_rise)
print("==> Epoch {}/{}".format(epoch + 1, cfg.SOLVER.MAX_EPOCHS))
print("current lr:", scheduler.get_lr()[0])
train(model, trainloader, xent, tent, optimizer, use_gpu)
if cfg.SOLVER.EVAL_PERIOD > 0 and (
(epoch + 1) % cfg.SOLVER.EVAL_PERIOD == 0 or
(epoch + 1) == cfg.SOLVER.MAX_EPOCHS):
print("==> Test")
metrics = test(model, queryloader, galleryloader,
cfg.TEST.TEMPORAL_POOL_METHOD, use_gpu)
rank1 = metrics[0]
if rank1 > best_rank1:
best_rank1 = rank1
no_rise = 0
else:
no_rise += 1
continue
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
torch.save(
state_dict,
osp.join(
cfg.OUTPUT_DIR, "rank1_" + str(rank1) + '_checkpoint_ep' +
str(epoch + 1) + '.pth'))
# best_p = osp.join(cfg.OUTPUT_DIR, "rank1_" + str(rank1) + '_checkpoint_ep' + str(epoch + 1) + '.pth')
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
print("Finished. Total elapsed time (h:m:s): {}".format(elapsed))
# strategy = tf.distribute.MirroredStrategy()
# print("Number of devices: {}".format(strategy.num_replicas_in_sync))
# TODO add lambda regularization
# TODO other regularizer (batch norm, dropout etc)
# TODO add activation layers ?
print(f"image data format is {tf.keras.backend.image_data_format()}")
latest_ckpt = tf.train.latest_checkpoint(CHECKPOINT_DIR)
datasets = data.get_hard_images()
model = DeepRankingNetwork(IMG_SHAPE, EMBEDDINGS_DIM)
base_lr = 0.001
max_lr = 0.06
model.compile(
optimizer=tf.keras.optimizers.Adam(base_lr), # TODO choose optimizer
loss=TripletLoss(GAP_PARAMETER, EMBEDDINGS_DIM, BATCH_SIZE),
metrics=[],
)
if latest_ckpt is not None:
model.load_weights(latest_ckpt)
model.fit(
datasets["train"][1],
validation_data=datasets["validation"][1],
validation_steps=datasets["validation"][0] // BATCH_SIZE,
steps_per_epoch=datasets["train"][0] // BATCH_SIZE,
epochs=EPOCHS,
callbacks=[
tf.keras.callbacks.LearningRateScheduler(
schedule=lambda epoch_index: base_lr + (max_lr - base_lr) *
def main():
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
else:
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
#print("Initializing dataset {}".format(args.dataset))
# dataset = data_manager.init_dataset(name=args.dataset)
transform_train = T.Compose([
T.Random2DTranslation(args.height, args.width),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
transform_train_p = T.Compose([
T.Random2DTranslation(256, 128),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
transform_test_p = T.Compose([
T.Resize((256, 128)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
train_file = 'data/cuhk_train.pkl'
test_file = 'data/cuhk_test.pkl'
gallery_file = 'data/cuhk_gallery.pkl'
data_root = args.data_root
dataset_train = CUHKGroup(train_file, data_root, True, transform_train,
transform_train_p)
dataset_test = CUHKGroup(test_file, data_root, False, transform_test,
transform_test_p)
dataset_query = CUHKGroup(test_file, data_root, False, transform_test,
transform_test_p)
dataset_gallery = CUHKGroup(gallery_file, data_root, False, transform_test,
transform_test_p)
pin_memory = True if use_gpu else False
if args.xent_only:
trainloader = DataLoader(
dataset_train,
batch_size=args.train_batch,
shuffle=True,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
else:
trainloader = DataLoader(
dataset_train,
sampler=RandomIdentitySampler(dataset_train,
num_instances=args.num_instances),
batch_size=args.train_batch,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(
dataset_test,
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
querygalleryloader = DataLoader(
dataset_query,
batch_size=args.gallery_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
galleryloader = DataLoader(
dataset_gallery,
batch_size=args.gallery_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
print("Initializing model: {}".format(args.arch))
if args.xent_only:
# model = models.init_model(name=args.arch, num_classes=dataset_train.num_train_gids, loss={'xent'})
model = models.init_model(name=args.arch,
num_classes=dataset_train.num_train_gids,
loss={'xent'})
else:
# model = models.init_model(name=args.arch, num_classes=dataset_train.num_train_gids, loss={'xent', 'htri'})
model = models.init_model(
name=args.arch,
num_classes=dataset_train.num_train_gids,
num_person_classes=dataset_train.num_train_pids,
loss={'xent', 'htri'})
#criterion_xent = CrossEntropyLabelSmooth(num_classes=dataset_train.num_train_gids, use_gpu=use_gpu)
#criterion_xent_person = CrossEntropyLabelSmooth(num_classes=dataset_train.num_train_pids, use_gpu=use_gpu)
if os.path.exists(args.pretrained_model):
print("Loading checkpoint from '{}'".format(args.pretrained_model))
checkpoint = torch.load(args.pretrained_model)
model_dict = model.state_dict()
pretrain_dict = checkpoint['state_dict']
pretrain_dict = {
k: v
for k, v in pretrain_dict.items() if k in model_dict
}
model_dict.update(pretrain_dict)
model.load_state_dict(model_dict)
criterion_xent = nn.CrossEntropyLoss(ignore_index=-1)
criterion_xent_person = nn.CrossEntropyLoss(ignore_index=-1)
criterion_htri = TripletLoss(margin=args.margin)
criterion_pair = ContrastiveLoss(margin=args.margin)
criterion_htri_filter = TripletLossFilter(margin=args.margin)
criterion_permutation = PermutationLoss()
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=args.lr,
weight_decay=args.weight_decay)
if args.stepsize > 0:
if args.warmup:
scheduler = WarmupMultiStepLR(optimizer, [200, 400, 600])
else:
scheduler = lr_scheduler.StepLR(optimizer,
step_size=args.stepsize,
gamma=args.gamma)
start_epoch = args.start_epoch
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
test_gcn_person_batch(model, queryloader, querygalleryloader,
galleryloader, args.pool, use_gpu)
#test_gcn_batch(model, queryloader, querygalleryloader, galleryloader, args.pool, use_gpu)
#test_gcn(model, queryloader, galleryloader, args.pool, use_gpu)
#test(model, queryloader, galleryloader, args.pool, use_gpu)
return
start_time = time.time()
best_rank1 = -np.inf
for epoch in range(start_epoch, args.max_epoch):
#print("==> Epoch {}/{} lr:{}".format(epoch + 1, args.max_epoch, scheduler.get_lr()[0]))
train_gcn(model, criterion_xent, criterion_xent_person, criterion_pair,
criterion_htri_filter, criterion_htri, criterion_permutation,
optimizer, trainloader, use_gpu)
#train(model, criterion_xent, criterion_htri, optimizer, trainloader, use_gpu)
if args.stepsize > 0: scheduler.step()
if args.eval_step > 0 and (epoch + 1) % args.eval_step == 0 or (
epoch + 1) == args.max_epoch:
print("==> Test")
rank1 = test_gcn_person_batch(model, queryloader,
querygalleryloader, galleryloader,
args.pool, use_gpu)
#rank1 = test_gcn(model, queryloader, galleryloader, args.pool, use_gpu=False)
#rank1 = test(model, queryloader, galleryloader, args.pool, use_gpu)
is_best = rank1 > best_rank1
if is_best: best_rank1 = rank1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint(
{
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, is_best,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
print("Finished. Total elapsed time (h:m:s): {}".format(elapsed))
def main():
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
else:
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_dataset(name=args.dataset)
transform_train = T.Compose([
T.Random2DTranslation(args.height, args.width),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = True if use_gpu else False
trainloader = DataLoader(
VideoDataset(dataset.train,
seq_len=args.seq_len,
sample='random',
transform=transform_train),
sampler=RandomIdentitySampler(dataset.train,
num_instances=args.num_instances),
batch_size=args.train_batch,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(
VideoDataset(dataset.query,
seq_len=args.seq_len,
sample='dense',
transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
VideoDataset(dataset.gallery,
seq_len=args.seq_len,
sample='dense',
transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing model: {}".format(args.arch))
if args.arch == 'resnet503d':
model = resnet3d.resnet50(num_classes=dataset.num_train_pids,
sample_width=args.width,
sample_height=args.height,
sample_duration=args.seq_len)
if not os.path.exists(args.pretrained_model):
raise IOError("Can't find pretrained model: {}".format(
args.pretrained_model))
print("Loading checkpoint from '{}'".format(args.pretrained_model))
checkpoint = torch.load(args.pretrained_model)
state_dict = {}
for key in checkpoint['state_dict']:
if 'fc' in key: continue
state_dict[key.partition("module.")
[2]] = checkpoint['state_dict'][key]
model.load_state_dict(state_dict, strict=False)
else:
model = models.init_model(name=args.arch,
num_classes=dataset.num_train_pids,
loss={'xent', 'htri'})
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
criterion_xent = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
criterion_htri = TripletLoss(margin=args.margin)
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
if args.stepsize > 0:
scheduler = lr_scheduler.StepLR(optimizer,
step_size=args.stepsize,
gamma=args.gamma)
start_epoch = args.start_epoch
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
test(model, queryloader, galleryloader, args.pool, use_gpu)
return
start_time = time.time()
best_rank1 = -np.inf
if args.arch == 'resnet503d':
torch.backends.cudnn.benchmark = False
for epoch in range(start_epoch, args.max_epoch):
print("==> Epoch {}/{}".format(epoch + 1, args.max_epoch))
train(model, criterion_xent, criterion_htri, optimizer, trainloader,
use_gpu)
if args.stepsize > 0: scheduler.step()
if args.eval_step > 0 and (epoch + 1) % args.eval_step == 0 or (
epoch + 1) == args.max_epoch:
print("==> Test")
rank1 = test(model, queryloader, galleryloader, args.pool, use_gpu)
is_best = rank1 > best_rank1
if is_best: best_rank1 = rank1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint(
{
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, is_best,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
print("Finished. Total elapsed time (h:m:s): {}".format(elapsed))
# {"criterion": CrossEntropyLabelSmooth(num_classes=class_num), "weight": 1.0,},
# {"criterion": TripletLoss(margin="0.3", metric="euclidean"), "weight": 1.0,},
{
"criterion": TripletLoss_WRT(),
"weight": 1.0,
},
])
else:
criterion = Criterion([
{
"criterion": nn.CrossEntropyLoss(),
"weight": 1.0,
},
# {"criterion": CrossEntropyLabelSmooth(num_classes=class_num), "weight": 1.0,},
{
"criterion": TripletLoss(margin="0.3", metric="euclidean"),
"weight": 1.0,
},
])
# build optimizer
if args.optim == "sgd":
args.lr = 0.1
ignored_params = list(map(id, net.bottleneck.parameters())) + list(
map(id, net.classifier.parameters()))
base_params = filter(lambda p: id(p) not in ignored_params,
net.parameters())
optimizer = torch.optim.SGD(
[
{
"params": base_params,
def main():
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
# set a learning rate
if args.lr_factor == -1:
args.lr_factor = random()
args.lr = args.lr_factor * 10**-args.lr_base
#print(f"Choose learning rate {args.lr}")
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'), mode='a')
print("==========\nArgs:{}\n==========".format(args))
#assert torch.distributed.is_available()
#print("Initializing DDP by nccl-tcp({}) rank({}) world_size({})".format(args.init_method, args.rank, args.world_size))
#dist.init_process_group(backend='nccl', init_method=args.init_method, rank=args.rank, world_size=args.world_size)
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_dataset(name=args.dataset, root=args.root)
# Data augmentation
spatial_transform_train = [
ST.Scale((args.height, args.width), interpolation=3),
ST.RandomHorizontalFlip(),
ST.ToTensor(),
ST.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
spatial_transform_train = ST.Compose(spatial_transform_train)
temporal_transform_train = TT.TemporalRandomCrop(size=args.seq_len, stride=args.sample_stride)
#temporal_transform_train = TT.TemporalRandomCropPick(size=args.seq_len, stride=args.sample_stride)
spatial_transform_test = ST.Compose([
ST.Scale((args.height, args.width), interpolation=3),
ST.ToTensor(),
ST.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
temporal_transform_test = TT.TemporalBeginCrop(size=args.test_frames)
pin_memory = True if use_gpu else False
dataset_train = dataset.train
if args.dataset == 'duke':
dataset_train = dataset.train_dense
print('process duke dataset')
#sampler = RandomIdentitySampler(dataset_train, num_instances=args.num_instances)
if args.dataset == 'lsvid':
sampler = RandomIdentitySampler(dataset_train, num_instances=args.num_instances)
elif args.dataset == 'mars':
sampler = RandomIdentitySampler(dataset_train, num_instances=args.num_instances)
trainloader = DataLoader(
VideoDataset(dataset_train, spatial_transform=spatial_transform_train, temporal_transform=temporal_transform_train),
sampler=sampler,
batch_size=args.train_batch, num_workers=args.workers,
pin_memory=pin_memory, drop_last=True,
)
'''
for batch_idx, (vids, pids, camids, img_paths) in enumerate(trainloader):
print(batch_idx, pids, camids, img_paths)
break
return
'''
dataset_query = dataset.query
dataset_gallery = dataset.gallery
if args.dataset == 'lsvid':
dataset_query = dataset.val_query
dataset_gallery = dataset.val_gallery
print('process lsvid dataset')
queryloader = DataLoader(
VideoDataset(dataset_query, spatial_transform=spatial_transform_test, temporal_transform=temporal_transform_test),
batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
pin_memory=pin_memory, drop_last=False
)
galleryloader = DataLoader(
VideoDataset(dataset_gallery, spatial_transform=spatial_transform_test, temporal_transform=temporal_transform_test),
batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
pin_memory=pin_memory, drop_last=False
)
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch, use_gpu=use_gpu, num_classes=dataset.num_train_pids, loss={'xent', 'htri'})
#print(model)
if args.resume:
print("Loading checkpoint from '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
criterion_xent = nn.CrossEntropyLoss()
criterion_htri = TripletLoss(margin=args.margin, distance=args.distance, use_gpu=use_gpu)
criterion_htri_c = TripletInterCamLoss(margin=args.margin, distance=args.distance, use_gpu=use_gpu)
#criterion_htri_c = TripletWeightedInterCamLoss(margin=args.margin, distance=args.distance, use_gpu=use_gpu, alpha=args.cam_alpha)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=args.stepsize, gamma=args.gamma)
start_epoch = args.start_epoch
if use_gpu:
model = nn.DataParallel(model).cuda()
#model = model.cuda()
#model = nn.parallel.DistributedDataParallel(model)
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
print("==> Start training")
for epoch in range(start_epoch, args.max_epoch):
#print("Set sampler seed to {}".format(args.seed*epoch))
#sampler.set_seed(args.seed*epoch)
start_train_time = time.time()
train(epoch, model, criterion_xent, criterion_htri, criterion_htri_c, optimizer, trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
scheduler.step()
if (epoch+1) >= args.start_eval and (epoch+1) % args.eval_step == 0 or epoch == 0:
print("==> Test")
with torch.no_grad():
rank1 = test(model, queryloader, galleryloader, use_gpu)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, is_best, osp.join(args.save_dir, 'checkpoint_ep' + str(epoch+1) + '.pth.tar'))
print("==> Best Rank-1 {:.1%}, achieved at epoch {}".format(best_rank1, best_epoch))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
train_time = str(datetime.timedelta(seconds=train_time))
print("Finished. Total elapsed time (h:m:s): {}. Training time (h:m:s): {}.".format(elapsed, train_time))
def main():
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'), mode='a')
else:
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'), mode='a')
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_imgreid_dataset(name=args.dataset,
dataset_dir=args.root,
fore_dir=args.fore_dir)
transform_train = ST.Compose([
ST.Scale((args.height, args.width), interpolation=3),
ST.RandomHorizontalFlip(),
ST.ToTensor(),
ST.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
ST.RandomErasing(0.5)
])
transform_test = ST.Compose([
ST.Scale((args.height, args.width), interpolation=3),
ST.ToTensor(),
ST.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = True if use_gpu else False
trainloader = DataLoader(
ImageDataset_hardSplit_seg(dataset.train, transform=transform_train),
sampler=RandomIdentitySampler(dataset.train,
num_instances=args.num_instances),
batch_size=args.train_batch,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(
ImageDataset(dataset.query, transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
ImageDataset(dataset.gallery, transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dataset.num_train_pids)
print(model)
criterion_xent = CrossEntropyLabelSmooth(use_gpu=use_gpu)
criterion_htri = TripletLoss()
criterion_mask = MaskLoss()
criterion_split = HardSplitLoss()
criterion_cluster = ClusterLoss()
optimizer = init_optim(args.optim, model.parameters(), args.lr,
args.weight_decay)
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=args.stepsize,
gamma=args.gamma)
if args.resume:
if check_isfile(args.resume):
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
args.start_epoch = checkpoint['epoch']
rank1 = checkpoint['rank1']
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(
args.start_epoch, rank1))
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
test(model, queryloader, galleryloader, use_gpu)
return
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
print("==> Start training")
for epoch in range(args.start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch, model, criterion_xent, criterion_htri, criterion_mask,
criterion_split, criterion_cluster, optimizer, trainloader,
use_gpu)
train_time += round(time.time() - start_train_time)
scheduler.step()
if (epoch + 1) > args.start_eval and (
epoch + 1) % args.eval_step == 0 or epoch == 0:
print("==> Test")
rank1 = test(model, queryloader, galleryloader, use_gpu)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint(
{
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, is_best,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
print("==> Best Rank-1 {:.1%}, achieved at epoch {}".format(
best_rank1, best_epoch))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
train_time = str(datetime.timedelta(seconds=train_time))
print(
"Finished. Total elapsed time (h:m:s): {}. Training time (h:m:s): {}.".
format(elapsed, train_time))
print("==========\nArgs:{}\n==========".format(args))
print('datasets loaded')
train_loader = torch.utils.data.DataLoader(train_set,
batch_size=cfg.data.imgs_per_gpu,
shuffle=True)
query_loader = torch.utils.data.DataLoader(query_set,
batch_size=cfg.data.imgs_per_gpu,
shuffle=True)
print('dataloader created')
# Build model
backbone_model = Vgg16L2(num_dim=128)
model = TripletNet(backbone_model)
model.cuda()
print('model built')
margin = 1.
loss_fn = TripletLoss(margin)
lr = 1e-4
optimizer = optim.Adam(model.parameters(), lr=lr)
scheduler = lr_scheduler.StepLR(optimizer, 8, gamma=0.1, last_epoch=-1)
n_epochs = 20
log_interval = 500
print('start training')
fit(train_loader, query_loader, model, loss_fn, optimizer, scheduler, n_epochs,
cuda, log_interval)
print('done training')
torch.save(model.state_dict(),
'./checkpoint/triplet_vanilla_20_epochs_1e-4.pth')
def main():
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'), mode='a')
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_dataset(name=args.dataset, root=args.root)
# Data augmentation
spatial_transform_train = [
ST.Scale((args.height, args.width), interpolation=3),
ST.RandomHorizontalFlip(),
ST.ToTensor(),
ST.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
spatial_transform_train = ST.Compose(spatial_transform_train)
temporal_transform_train = TT.TemporalRandomCrop(size=args.seq_len,
stride=args.sample_stride)
spatial_transform_test = ST.Compose([
ST.Scale((args.height, args.width), interpolation=3),
ST.ToTensor(),
ST.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
temporal_transform_test = TT.TemporalBeginCrop(size=args.test_frames)
pin_memory = True if use_gpu else False
dataset_train = dataset.train
if args.dataset == 'duke':
dataset_train = dataset.train_dense
print('process duke dataset')
trainloader = DataLoader(
VideoDataset(dataset_train,
spatial_transform=spatial_transform_train,
temporal_transform=temporal_transform_train),
sampler=RandomIdentitySampler(dataset_train,
num_instances=args.num_instances),
batch_size=args.train_batch,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(VideoDataset(
dataset.query,
spatial_transform=spatial_transform_test,
temporal_transform=temporal_transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False)
galleryloader = DataLoader(VideoDataset(
dataset.gallery,
spatial_transform=spatial_transform_test,
temporal_transform=temporal_transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False)
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch,
use_gpu=use_gpu,
num_classes=dataset.num_train_pids,
loss={'xent', 'htri'})
print(model)
criterion_xent = nn.CrossEntropyLoss()
criterion_htri = TripletLoss(margin=args.margin,
distance=args.distance,
use_gpu=use_gpu)
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=args.stepsize,
gamma=args.gamma)
start_epoch = args.start_epoch
if use_gpu:
model = nn.DataParallel(model).cuda()
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
print("==> Start training")
for epoch in range(start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch, model, criterion_xent, criterion_htri, optimizer,
trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
scheduler.step()
if (epoch + 1) >= args.start_eval and (
epoch + 1) % args.eval_step == 0 or epoch == 0:
print("==> Test")
with torch.no_grad():
rank1 = test(model, queryloader, galleryloader, use_gpu)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint(
{
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, is_best,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
print("==> Best Rank-1 {:.1%}, achieved at epoch {}".format(
best_rank1, best_epoch))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
train_time = str(datetime.timedelta(seconds=train_time))
print(
"Finished. Total elapsed time (h:m:s): {}. Training time (h:m:s): {}.".
format(elapsed, train_time))
