def main():
# get dataset
queries_dataset, gallery_dataset = \
get_dataset_val(root=
"/media/fanyang/workspace/DataSet/MARS/bbox_test",
txt_file=os.path.join(Proj_Dir, 'data/id2folderVal.txt'), seq_len=SEQ_LEN)
# prepare model
model = MixModel(num_classes=625, batch_size=BATCH_SIZE, seq_len=SEQ_LEN,
transform_input=True)
model.load_state_dict(
torch.load("/home/fanyang/PycharmProjects/PersonReID_CL/ckptseq/model-v1.pkl"))
model.cuda()
queries_loader = get_loader(dataset=queries_dataset, batch_size=BATCH_SIZE,
shuffle=False, drop_last=False)
gallery_loader = get_loader(dataset=gallery_dataset, batch_size=BATCH_SIZE,
shuffle=False, drop_last=False)
query_test_count = gallery_dataset.get_query_test_count()
validation(queries_loader=queries_loader, gallery_loader=gallery_loader, model=model,
query_test_count=query_test_count)
def main():
# get dataset
queries_dataset, gallery_dataset = \
get_dataset_val(root=
"/media/fanyang/workspace/DataSet/MARS/bbox_test",
txt_file=os.path.join(Proj_Dir, 'data/id2folderVal.txt'), seq_len=SEQ_LEN)
# prepare model
model = Inceptionv2()
# resnet.criterion2 = CenterLoss(feature_len=2048, num_classes=625)
# resnet.criterion2 = ImprovedCenterLoss(feature_len=2048, num_classes=625)
model.load_state_dict(
torch.load(
"/home/fanyang/PycharmProjects/PersonReID_CL/ckpt/no-cl-model-inceptionv2/record-step-2109-model.pkl"
))
model.classifier = None
model.cuda()
queries_loader = get_loader(dataset=queries_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
drop_last=False)
gallery_loader = get_loader(dataset=gallery_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
drop_last=False)
query_test_count = gallery_dataset.get_query_test_count()
validation(queries_loader=queries_loader,
gallery_loader=gallery_loader,
model=model,
query_test_count=query_test_count)
def main():
# get dataset
train_dataset = get_dataset(
root="/media/fanyang/workspace/DataSet/MARS/bbox_train",
txt_file=os.path.join(Proj_Dir, 'data/train.txt'))
val_dataset = get_dataset(
root="/media/fanyang/workspace/DataSet/MARS/bbox_train",
txt_file=os.path.join(Proj_Dir, 'data/test.txt'))
# prepare model
resnet = resnet50(pretrained=False)
# two criterion
resnet.criterion = nn.CrossEntropyLoss()
resnet.criterion2 = SiameseLoss()
resnet.fc = nn.Linear(in_features=2048, out_features=625)
resnet.load_state_dict_diy(
torch.load(
"/home/fanyang/PycharmProjects/PersonReID_CL/ckpt/no-cl-model-resnet50/last-layer-finetuned-model.pkl"
))
# resnet.optimizer = optim.Adam(params=resnet.fc.parameters(), lr=1e-3)
resnet.optimizer = optim.SGD(params=resnet.parameters(), lr=5e-4)
resnet.cuda()
# for record the train process
writer_dir = "ckpt/siamese-model-resnet50"
saver_dir = writer_dir
writer = SummaryWriter(log_dir=os.path.join(Proj_Dir, writer_dir))
routine = Routine2Criteion(model=resnet,
saver_dir=saver_dir,
writer=writer)
while True:
train_loader = get_loader(dataset=train_dataset, batch_size=BATCH_SIZE)
val_loader = get_loader(dataset=val_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
drop_last=False)
routine.train_one_epoch(loader=train_loader,
record_n_times_per_epoch=400)
# adjust the learning per epoch
adjust_learning_rate(resnet.optimizer)
routine.validation(loader=val_loader)
def main():
# get dataset
train_dataset = get_dataset(
root="/media/fanyang/workspace/DataSet/MARS/bbox_train",
txt_file=os.path.join(Proj_Dir, 'data/train.txt'))
val_dataset = get_dataset(
root="/media/fanyang/workspace/DataSet/MARS/bbox_train",
txt_file=os.path.join(Proj_Dir, 'data/test.txt'))
# prepare model
model = Inceptionv2()
# two criterion
model.criterion = nn.CrossEntropyLoss()
model.criterion2 = CenterLoss(feature_len=1024, num_classes=625)
# model.load_state_dict(
# torch.load(
# "/home/fanyang/PycharmProjects/PersonReID_CL/ckpt/no-cl-model-resnet50/last-layer-finetuned-model.pkl"))
model.cuda()
model.optimizer = optim.Adam(params=model.parameters(), lr=1e-3)
# model.optimizer = optim.SGD(params=model.parameters(), lr=1e-3)
# for record the train process
writer_dir = "ckpt/cl-model-inceptionv2"
saver_dir = writer_dir
writer = SummaryWriter(log_dir=os.path.join(Proj_Dir, writer_dir))
routine = Routine2Criteion(model=model, saver_dir=saver_dir, writer=writer)
while True:
train_loader = get_loader(dataset=train_dataset, batch_size=BATCH_SIZE)
val_loader = get_loader(dataset=val_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
drop_last=False)
# print(type(batch_data))
# print(type(batch_label))
# exit()
routine.train_one_epoch(loader=train_loader,
record_n_times_per_epoch=400)
# adjust the learning per epoch
adjust_learning_rate(model.optimizer)
routine.validation(loader=val_loader)
def main():
val_dataset = get_dataset(
root="/media/fanyang/workspace/DataSet/MARS/bbox_train",
txt_file=os.path.join(Proj_Dir, 'data/test.txt'))
# prepare model
inception = inception_v3(pretrained=False,
aux_logits=False,
transform_input=True)
inception.fc = nn.Linear(in_features=2048, out_features=625)
inception.load_state_dict(
torch.load(
"/home/fanyang/PycharmProjects/PersonReID_CL/ckpt/model-inceptionv3-transform-input.pkl"
))
# inception.optimizer = optim.Adam(params=inception.fc.parameters(), lr=5e-4)
# set the last layer to None for using the penultimate layer's feature
inception.fc = None
inception.cuda()
# writer = SummaryWriter(log_dir=os.path.join(Proj_Dir, "ckpt/sgd-5e-5"))
writer_for_embedding = SummaryWriter(
log_dir=os.path.join(Proj_Dir, "ckpt/cnn_embedding"))
val_loader = get_loader(dataset=val_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
drop_last=False)
validation(loader=val_loader, model=inception, writer=writer_for_embedding)
def main():
# get dataset
train_dataset = get_dataset(
root="/media/fanyang/workspace/DataSet/MARS/bbox_train",
txt_file=os.path.join(Proj_Dir, 'data/train.txt'))
val_dataset = get_dataset(
root="/media/fanyang/workspace/DataSet/MARS/bbox_train",
txt_file=os.path.join(Proj_Dir, 'data/test.txt'))
# prepare model
inceptionv2 = Inceptionv2()
inceptionv2.criterion = nn.CrossEntropyLoss()
inceptionv2.cuda()
# inceptionv2.load_state_dict(
# torch.load(
# "/home/fanyang/PycharmProjects/PersonReID_CL/ckpt/model-inceptionv3-transform-input.pkl"))
inceptionv2.optimizer = optim.Adam(params=inceptionv2.parameters(),
lr=1e-3)
# inception.optimizer = optim.SGD(params=inception.parameters(), lr=5e-4)
# for record the train process
writer_dir = "ckpt/no-cl-model-inceptionv2"
saver_dir = writer_dir
writer = SummaryWriter(log_dir=os.path.join(Proj_Dir, writer_dir))
routine = Routine(model=inceptionv2, saver_dir=saver_dir, writer=writer)
while True:
train_loader = get_loader(dataset=train_dataset, batch_size=BATCH_SIZE)
val_loader = get_loader(dataset=val_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
drop_last=False)
routine.train_one_epoch(loader=train_loader,
record_n_times_per_epoch=400)
# adjust the learning per epoch
adjust_learning_rate(inceptionv2.optimizer)
routine.validation(loader=val_loader)
def main():
# get dataset
# train_dataset = get_dataset(root="/media/fanyang/workspace/DataSet/MARS/bbox_train",
# txt_file=os.path.join(Proj_Dir, 'data/train.txt'))
train_dataset = get_dataset_seq(
root="/media/fanyang/workspace/DataSet/MARS/bbox_train",
txt_file=os.path.join(Proj_Dir, 'data/id2folder.txt'),
seq_len=SEQ_LEN)
# prepare model
model = MixModel(num_classes=625,
batch_size=BATCH_SIZE,
seq_len=SEQ_LEN,
transform_input=True)
# model.load_inception_weights(
# "/home/fanyang/PycharmProjects/PersonReID_CL/ckpt/model-inceptionv3-transform-input.pkl")
model.load_state_dict(
torch.load(
"/home/fanyang/PycharmProjects/PersonReID_CL/ckptseq/model-v1.pkl")
)
for param in model.inception.parameters():
param.requires_grad = False
# get the parameters to train
parameters_need_to_train = []
for param in model.inception_lstm.parameters():
parameters_need_to_train.append(param)
for param in model.fc.parameters():
parameters_need_to_train.append(param)
model.criterion = nn.CrossEntropyLoss()
# model.optimizer = optim.Adam(params=parameters_need_to_train, lr=5e-4)
model.optimizer = optim.SGD(params=parameters_need_to_train, lr=1e-4)
model.cuda()
writer = SummaryWriter(log_dir=os.path.join(Proj_Dir, "ckpt/seq-sgd-1e-4"))
while True:
train_loader = get_loader(dataset=train_dataset, batch_size=BATCH_SIZE)
# print(type(batch_data))
# print(type(batch_label))
# exit()
train_one_epoch(loader=train_loader, model=model, writer=writer)
adjust_learning_rate(model.optimizer, decay_rate=.9)