def main(config):
cudnn.benchmark = True
if not os.path.exists(config.log_dir):
os.makedirs(config.log_dir)
if not os.path.exists(config.model_save_dir):
os.makedirs(config.model_save_dir)
if not os.path.exists(config.sample_dir):
os.makedirs(config.sample_dir)
if not os.path.exists(config.result_dir):
os.makedirs(config.result_dir)
celeba_loader = None
rafd_loader = None
if config.dataset in ['CelebA', 'Both']:
celeba_loader = get_loader(config.celeba_image_dir, config.attr_path, config.selected_attrs,
config.celeba_crop_size, config.image_size, config.batch_size,
'CelebA', config.mode, config.num_workers)
if config.dataset in ['RaFD', 'Both']:
rafd_loader = get_loader(config.rafd_image_dir, None, None,
config.rafd_crop_size, config.image_size, config.batch_size,
'RaFD', config.mode, config.num_workers)
print('zz')
solver = Solver(celeba_loader, rafd_loader, config)
print(solver)
def __init__(self, config):
self.train_data_loader = get_loader(config.train_image_path,
config.train_label_path,
config.image_size,
config.batch_size)
self.val_data_loader = get_loader(config.val_image_path,
config.val_label_path,
config.image_size, config.batch_size)
self.image_size = config.image_size
self.lr = config.lr
self.beta1 = config.beta1
self.beta2 = config.beta2
self.epoches = config.epoches
self.yolov3_cfg = config.model_config_path
self.num_classes = config.num_classes
self.conf_thres = config.conf_thres
self.nms_thres = config.nms_thres
self.iou_thres = config.iou_thres
self.parser_classes_tag(config.classes_file)
self.log_path = config.log_save_path
self.model_save_path = config.model_save_path
commons.make_dirs([self.log_path, self.model_save_path])
self.log_step = config.log_step
self.model_save_step = config.model_save_step
self.build_model()
self.use_tensorboard = config.use_tensorboard
self.pretrained_model = config.pretrained_model
if config.use_tensorboard:
self.build_tensorboard()
if self.pretrained_model:
print("load pretrained model in", self.pretrained_model)
self.load_pretrain_model()
def main():
config = get_config(mode='train')
val_config = get_config(mode='valid')
with open(os.path.join(config.save_path, 'config.json'), 'w') as json_f:
config.to_json(json_f)
raw_data = load_json(config.all_path)
train_data_loader = get_loader(raw_data=raw_data,
max_len=config.max_len,
batch_size=config.batch_size,
shuffle=True,
user_map_dict=config.user_map_dict,
max_users=config.max_users)
raw_data = load_json(val_config.all_path)
eval_data_loader = get_loader(raw_data=raw_data,
max_len=val_config.max_len,
batch_size=val_config.eval_batch_size,
shuffle=False,
user_map_dict=config.user_map_dict,
max_users=config.max_users)
model_solver = getattr(solvers, "Solver{}".format(config.model))
solver = model_solver(config,
train_data_loader,
eval_data_loader,
is_train=True)
solver.build()
solver.train()
solver.writer.close()
return config
def build_dataloader(self, epoch):
print(f'make dataloder epoch {epoch}')
self.train_loader = get_loader(self.data_dir, self.data_name, self.data_type, self.crop_range, self.hu_range,
self.down_size, num_workers=2, shuffle=True,mode='train', epoch=epoch,
n_splits=self.n_splits)
self.val_loader = get_loader(self.data_dir, self.data_name, self.data_type, self.crop_range, self.hu_range,
self.down_size, num_workers=0, shuffle=False, mode='val', epoch=epoch,
n_splits=self.n_splits)
self.num_steps = len(self.train_loader)
def train_model(model, train_img_dir, train_csv_file, valid_img_dir, valid_csv_file,\
batch_size,num_workers, epoch, lr, transform=None, center=True):
train_loader = get_loader(image_dir=train_img_dir, csv_dir=train_csv_file,\
batch_size=batch_size, num_workers=num_workers, transform=transform, center=center)
valid_loader = get_loader(image_dir=valid_img_dir, csv_dir=valid_csv_file,\
batch_size=batch_size, num_workers=num_workers, transform=transform, center=center)
# tf_summary_write = tf and tf.summary.FileWriter('../log/')
model.cuda()
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=lr)
# log = Logger('../log/')
for e in range(epoch):
model.train(True)
print ('#'*20)
print ('epoch%d'%e)
for i, (img, target) in enumerate(train_loader):
img_v = Variable(img, requires_grad=True).cuda()
target = Variable(target, requires_grad=False).cuda()
optimizer.zero_grad()
pred = model(img_v)
loss = criterion(pred, target)
loss.backward()
optimizer.step()
if i%9 == 0:
print ("training loss is %f"%loss.data[0])
# log.scalar_summary('train_loss', loss.data[0], e*len(train_loader)+i)
#add_summary_value(tf_summary_write, 'train_loss', loss.data[0], e*len(train_loader)+i)
#tf_summary_write.flush()
model.train(False)
val_loss = 0.0
for img, target in valid_loader:
img_nv = Variable(img, volatile=True).cuda()
target = Variable(target, volatile=True).cuda()
pred = model(img_nv)
loss = criterion(pred, target)
val_loss+= loss.data[0]
print ('-'*20)
val_loss = val_loss / len(valid_loader)
# log.scalar_summary('val_loss', val_loss, e)
# add_summary_value(tf_summary_write, 'val_loss', val_loss, e)
print ('val_loss: %f'%val_loss)
print ('-'*20)
torch.save(model.state_dict(), '../model/model%d.pkl'%e)
def __init__(self, args):
self.args = args
self.device = torch.device('cuda: 0')
self.device_ids = list(map(lambda x: int(x), args.gpus.split(',')))
self.loader1 = get_loader(args.root1, args)
self.loader2 = get_loader(args.root2, args)
self.netG_A = generator(input_nc=3, output_nc=3, n_blocks=9)
self.netG_B = generator(input_nc=3, output_nc=3, n_blocks=9)
self.netD_A = discriminator(input_nc=3)
self.netD_B = discriminator(input_nc=3)
self.netG_A.to(self.device)
self.netG_B.to(self.device)
self.netD_A.to(self.device)
self.netD_B.to(self.device)
self.netG_A = nn.DataParallel(self.netG_A, device_ids=self.device_ids)
self.netG_B = nn.DataParallel(self.netG_B, device_ids=self.device_ids)
self.netD_A = nn.DataParallel(self.netD_A, device_ids=self.device_ids)
self.netD_B = nn.DataParallel(self.netD_B, device_ids=self.device_ids)
init_weights(self.netG_A)
init_weights(self.netG_B)
init_weights(self.netD_A)
init_weights(self.netD_B)
self.fake_A_pool = ImagePool(50)
self.fake_B_pool = ImagePool(50)
self.criterionCycle = nn.L1Loss()
self.criterionGAN = LSGANLoss()
self.start_epoch = 0
self.optimizer_G = torch.optim.Adam(itertools.chain(
self.netG_A.parameters(), self.netG_B.parameters()),
lr=self.args.lr,
betas=(0.5, 0.999))
self.optimizer_D = torch.optim.Adam(itertools.chain(
self.netD_A.parameters(), self.netD_B.parameters()),
lr=self.args.lr,
betas=(0.5, 0.999))
def lambda_rule(epoch):
lr_l = 1.0 - max(0, epoch + 1 - 100) / float(100 + 1)
return lr_l
self.scheduler_G = lr_scheduler.LambdaLR(self.optimizer_G,
lr_lambda=lambda_rule)
self.scheduler_D = lr_scheduler.LambdaLR(self.optimizer_D,
lr_lambda=lambda_rule)
self.create_experiment()
self.get_logger()
def main():
global args, best_err1, best_err5
# Get arguments using parser
args = parser.parse_args()
# Load a dataset and make data_loader
val_short_loader = DataLoader.get_loader(X_path=args.data_path + "test/",
y_path=args.data_path,
duration='short',
batch_size=int(args.batch_size /
2))
val_long_loader = DataLoader.get_loader(X_path=args.data_path + "test/",
y_path=args.data_path,
duration='long',
batch_size=int(args.batch_size /
2))
val_loader = [val_short_loader, val_long_loader]
print("===== Finished Loading Datasets =====")
# Create a CNN model and LSTM model
cnn_model = tn.timeConvNet()
lstm_model = VANILLA(args).cuda()
print(cnn_model)
print(lstm_model)
"""
Left as blank since we haven't specify the model to use.
"""
#lstm_model = torch.nn.DataParallel(lstm_model).cuda()
cnn_model = torch.nn.DataParallel(cnn_model).cuda()
"""
if load is true, then load the parameter
"""
print("=> loading checkpoint '{}'".format(args.pretrained))
cnn_checkpoint = torch.load(args.pretrained + '/cnn_model_best.pth.tar')
lstm_checkpoint = torch.load(args.pretrained + '/lstm_model_best.pth.tar')
print("=> loaded checkpoint '{}'".format(args.pretrained))
cnn_model.load_state_dict(cnn_checkpoint['state_dict'])
lstm_model.load_state_dict(lstm_checkpoint['state_dict'])
cudnn.benchmark = True
criterion = nn.CrossEntropyLoss().cuda()
print("===== Finished Constructing Models =====")
validate(val_loader, cnn_model, lstm_model, criterion)
def main():
"""Main Function."""
# dataloader parameters
gpu = torch.cuda.is_available()
data_path = 'data/labels_done.txt'
batch_size = 32
num_workers = 2
# network parameters
architecture = 'VGGNet11'
pretrained = True
finetuned = True
# training parameters
learning_rate = 1e-4
max_epochs = 200
criterion = nn.CrossEntropyLoss()
# get dataloader
dataloader, dataset_size = get_loader(data_path, batch_size, num_workers)
print('Dataset Size:', dataset_size)
# create network object
net = SingleFrame(architecture, pretrained, finetuned)
print(net)
# create optimizer
if not finetuned:
optimizer = torch.optim.Adam(net.fc.parameters(), learning_rate)
else:
optimizer = torch.optim.Adam(net.parameters(), learning_rate)
# train the network
best_acc, losses, accuracies = train_network(net, dataloader, dataset_size,
batch_size, criterion, optimizer, max_epochs, gpu)
# plot statistics
print('Best Training Accuracy:', best_acc*100)
def main(config):
# Save config with experiment data.
target_dir = os.path.join(config.exp_dir, config.exp_id)
if not os.path.exists(target_dir):
os.makedirs(target_dir)
with open(os.path.join(target_dir, "config.json"), 'a') as f:
print(json.dumps(config.__dict__, sort_keys=True, indent=4), file=f)
cudnn.benchmark = True
# Data loader.
celeba_loader = get_loader(config.cluster_npz_path, config.dataset_path,
config.celeba_crop_size, config.image_size,
config.batch_size, config.dataset, config.mode,
config.num_workers,
config.selected_clusters_train)
# Solver for training and testing StarGAN.
solver = Solver(celeba_loader, config)
if config.mode == 'train':
solver.train()
elif config.mode == 'test':
solver.test()
def main():
logger.info(args)
model = Net(numAngle=CONFIGS["MODEL"]["NUMANGLE"], numRho=CONFIGS["MODEL"]["NUMRHO"], backbone=CONFIGS["MODEL"]["BACKBONE"])
model = model.cuda(device=CONFIGS["TRAIN"]["GPU_ID"])
# load the pretrained model (you are free to load your own models)
state_dict = torch.hub.load_state_dict_from_url("http://data.kaizhao.net/projects/deep-hough-transform/dht_r50_fpn_sel-c9a29d40.pth", check_hash=True)
model.load_state_dict(state_dict)
if args.model:
if isfile(args.model):
logger.info("=> loading pretrained model '{}'".format(args.model))
checkpoint = torch.load(args.model)
model.load_state_dict(checkpoint)
logger.info("=> loaded checkpoint '{}'"
.format(args.model))
else:
logger.info("=> no pretrained model found at '{}'".format(args.model))
exit()
# dataloader
test_loader = get_loader(CONFIGS["DATA"]["TEST_DIR"], CONFIGS["DATA"]["TEST_LABEL_FILE"],
batch_size=1, num_thread=CONFIGS["DATA"]["WORKERS"], test=True)
logger.info("Data loading done.")
logger.info("Start testing.")
total_time = test(test_loader, model, args)
logger.info("Test done! Total %d imgs at %.4f secs without image io, fps: %.3f" % (len(test_loader), total_time, len(test_loader) / total_time))
def set_rcam(self):
print("Creating CAM for {}".format(self.args.model))
if 'resnet' in str.lower(type(self.model).__name__):
last_conv = 'layer4'
else:
print("Model not implemented. Setting rcam=False by default.")
return
self.weights = EvaluationMetrics(list(range(self.args.num_classes)))
def hook_weights(module, input, output):
weights.append(F.adaptive_max_pool2d(output, (1,1)))
handle = self.model._modules.get(last_conv).register_forward_hook(hook_weights)
train_loader, _ = get_loader(self.args.dataset,
batch_size=1,
num_workers=self.args.workers
)
for i, (image, label) in enumerate(train_loader):
weights = []
_ = self.model(to_var(image, volatile=True))
weights = weights[0].squeeze()
label = label.squeeze()[0]
self.weights.update(label, weights)
if (i+1)%1000 == 0:
print("{:5.1f}% ({}/{})".format((i+1)/len(train_loader)*100, i+1, len(train_loader)))
handle.remove()
def main():
config = get_config(mode='test')
with open(os.path.join(config.save_path, 'config.json'), 'r') as json_f:
temp_config_str = json_f.read()
config.max_users = int(
re.findall(r"'max_users': ([0-9]+?),", temp_config_str)[0])
config.max_len = int(
re.findall(r"'max_len': ([0-9]+?),", temp_config_str)[0])
config.rnn_hidden_size = int(
re.findall(r"'rnn_hidden_size': ([0-9]+?),", temp_config_str)[0])
raw_data = load_json(config.all_path)
test_data_loader = get_loader(raw_data=raw_data,
max_len=config.max_len,
batch_size=config.batch_size,
shuffle=False,
user_map_dict=config.user_map_dict,
max_users=config.max_users)
model_solver = getattr(solvers, "Solver{}".format(config.model))
solver = model_solver(config, None, test_data_loader, is_train=False)
solver.build()
solver.test()
return config
def main(args):
device = torch.device('cuda:{}'.format(args.device)) \
if args.cuda else torch.device('cpu')
model = EfficientDet.from_pretrained(args.model).to(device) \
if args.pretrained else EfficientDet.from_name(args.model).to(device)
if args.mode == 'trainval':
logger("Model's trainable parameters: {}".format(
count_parameters(model)))
loader = get_loader(path=cfg.TRAIN_SET,
annotations=cfg.TRAIN_ANNOTATIONS,
batch_size=cfg.BATCH_SIZE)
optimizer, scheduler, criterion, ema_decay = build_tools(model)
writer = setup_writer(args.experiment, args)
best_score = -1
for epoch in range(cfg.NUM_EPOCHS):
model, optimizer, scheduler, writer = \
train(model, optimizer, loader, scheduler,
criterion, ema_decay, device, writer)
if epoch > cfg.VAL_DELAY and \
(epoch + 1) % cfg.VAL_INTERVAL == 0:
ema_decay.assign(model)
model, writer, best_score = \
validate(model, device, writer,
cfg.MODEL.SAVE_PATH, best_score=best_score)
ema_decay.resume(model)
elif args.mode == 'eval':
validate(model, device)
def main():
logger.info(args)
model = Net(numAngle=CONFIGS["MODEL"]["NUMANGLE"],
numRho=CONFIGS["MODEL"]["NUMRHO"],
backbone=CONFIGS["MODEL"]["BACKBONE"])
model = model.cuda(device=CONFIGS["TRAIN"]["GPU_ID"])
if args.model:
if isfile(args.model):
logger.info("=> loading pretrained model '{}'".format(args.model))
checkpoint = torch.load(args.model)
model.load_state_dict(checkpoint['state_dict'])
logger.info("=> loaded checkpoint '{}' (epoch {})".format(
args.model, checkpoint['epoch']))
else:
logger.info("=> no pretrained model found at '{}'".format(
args.model))
# dataloader
test_loader = get_loader(CONFIGS["DATA"]["TEST_DIR"],
CONFIGS["DATA"]["TEST_LABEL_FILE"],
batch_size=1,
num_thread=CONFIGS["DATA"]["WORKERS"],
test=True)
logger.info("Data loading done.")
logger.info("Start testing.")
total_time = test(test_loader, model, args)
logger.info(
"Test done! Total %d imgs at %.4f secs without image io, fps: %.3f" %
(len(test_loader), total_time, len(test_loader) / total_time))
def main(config):
if config.out_folder is None:
config.out_folder = 'new_samples'
os.system('mkdir {0}'.format(config.out_folder))
config.manual_seed = random.randint(1, 10000)
print("Random Seed: ", config.manual_seed)
random.seed(config.manual_seed)
torch.manual_seed(config.manual_seed)
torch.cuda.manual_seed_all(config.manual_seed)
cudnn.benchmark = True
dataloader = get_loader(_dataset=config.dataset,
dataroot=config.dataroot,
batch_size=config.batch_size,
num_workers=int(config.workers),
image_size=config.image_size)
print("Prepare dataloader complete!")
# for data, _ in dataloader:
# print(data.size())
# break
trainer = Trainer(config, dataloader)
trainer.train()
def main(config):
if config.out_folder is None:
config.out_folder = 'samples'
os.system('mkdir {0}'.format(config.out_folder))
config.manual_seed = random.randint(1, 10000)
print("Random Seed: ", config.manual_seed)
random.seed(config.manual_seed)
torch.manual_seed(config.manual_seed)
torch.cuda.manual_seed_all(config.manual_seed)
cudnn.benchmark = True
print("Used dataset:", config.dataset)
dataloader = get_loader(_dataset=config.dataset,
dataroot=config.dataroot,
batch_size=config.batch_size,
num_workers=int(config.workers),
image_size=config.image_size)
print("Prepare dataloader complete!")
# gen = Generator(config.nz, config.ngf, config.nch, config.n_classes)
# dis = Discriminator(config.ndf, config.nch, config.n_classes)
#
# gen = gen.to(device)
# dis = dis.to(device)
#
# # prepare label
# onehot = torch.zeros(10, 10)
# onehot = onehot.scatter_(1, torch.LongTensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]).view(10, 1), 1).view(10, 10, 1, 1)
# fill = torch.zeros([10, 10, config.image_size, config.image_size])
# for i in range(10):
# fill[i, i, :, :] = 1 # shape: 10, 10, image_size, image_size
#
# for x_, y_ in dataloader:
# step_batch = x_.size(0)
#
# x_ = x_.to(device) # shape: batch, 1, image_size, image_size
# y_ = y_.to(device) # shape: batch
# y_fill = fill[y_] # shape: batch, 10, image_size, image_size
#
# dis_out = dis(x_, y_fill) # dis input: x_, y_fill
# # output: batch
#
# z_ = torch.randn((step_batch, 100)).view(-1, 100, 1, 1).to(device)
# y_ = (torch.rand(step_batch, 1) * 10).long().to(device).squeeze() # batch
#
# y_label = onehot[y_] # batch, 10, 1, 1
# y_fill = fill[y_] # batch, 10, 32, 32
#
# gen_out = gen(z_, y_label) # gen input: z_, y_label
# # output: batch, 1, image_size, image_size
#
# break
trainer = Trainer(config, dataloader)
trainer.train()
def main(config):
# make directory
if config.ckpt_folder is None:
config.ckpt_folder = 'checkpoints'
os.system('mkdir {0}'.format(config.ckpt_folder))
print("[*] Make checkpoints folder!")
if config.sample_folder is None:
config.sample_folder = 'samples'
os.system('mkdir {0}'.format(config.sample_folder))
print("[*] Make samples folder!")
config.manual_seed = random.randint(1, 10000)
print("Random Seed: ", config.manual_seed)
random.seed(config.manual_seed)
torch.manual_seed(config.manual_seed)
torch.cuda.manual_seed_all(config.manual_seed)
# for faster training
cudnn.benchmark = True
# define train/test dataloader
train_loader, test_loader = get_loader(data_folder=config.dataroot,
batch_size=config.batch_size,
image_size=256)
if config.mode == 'train':
# define trainer class
trainer = Trainer(train_loader, test_loader, config)
trainer.train()
elif config.mode == 'test':
finder = Finder(config)
finder.find_topn()
def test(model_path, test_path, bs):
model = MusicModel()
model = torch.load(model_path)
test_loader = get_loader(test_path, bs)
test_loader = iter(test_loader)
TP = 0
FN = 0
FP = 0
TN = 0
desc = ' - (Testing) - '
for (data, label) in tqdm(test_loader, desc=desc, ncols=80):
result = float(model(data).squeeze(-1).squeeze(-1))
label = int(label[0])
if label == 1:
if result >= 0.5:
TP += 1
else:
FN += 1
else:
if result >= 0.5:
FP += 1
else:
TN += 1
acc = float(TP + TN) / float(TP + FN + FP + TN)
acc = round(acc * 100, 2)
print('ACC:' + str(acc))
def main(config):
if config.outf is None:
config.outf = 'samples'
os.system('mkdir {0}'.format(config.outf))
config.manual_seed = random.randint(1, 10000)
print("Random Seed: ", config.manual_seed)
random.seed(config.manual_seed)
torch.manual_seed(config.manual_seed)
if config.cuda:
torch.cuda.manual_seed_all(config.manual_seed)
cudnn.benchmark = True
dataroot = config.dataroot
h_datapath = os.path.join(dataroot, "HV")
r_datapath = os.path.join(dataroot, "RV")
t_datapath = os.path.join(dataroot, 'testRV')
# dataroot, cache, image_size, n_channels, image_batch, video_batch, video_length):
h_loader, r_loader, t_loader = get_loader(h_datapath, r_datapath, t_datapath)
config.n_steps = min(len(h_loader), len(r_loader))
trainer = Trainer(config, h_loader, r_loader, t_loader)
trainer.train()
def get_accuracy(self):
# load dataloader
_, _, t_l = get_loader('../Dataset/HV', '../Dataset/RV',
'../Dataset/testRV', 1)
# build network
self.c2d = CNN().cuda()
self.c2d.load_state_dict(
torch.load('cnn.pkl')) # load pre-trained cnn extractor
for l, p in self.c2d.named_parameters():
p.requires_grad = False
self.gru = GRU(self.c2d).cuda()
self.gru.load_state_dict(torch.load(self.ckpt))
print(self.gru)
self.gru.eval()
avg_acc = 0
for idx, (video, label) in enumerate(t_l):
acc = 0.
# forwarding
test_video = Variable(video).cuda()
predicted = self.gru(test_video)
predicted = predicted.cpu().numpy()
print('Predicted output:',
predicted) # [forwarding score ....., backwarding score]
print('Predicted output length:', len(predicted))
print('Actual label:', label)
print('Actual label length:', len(label))
def train():
with open(vocab_path, 'rb') as f:
vocab = pickle.load(f)
vocab_size = len(vocab)
print('vocab_size:', vocab_size)
dataloader = get_loader(image_dir,
caption_path,
vocab,
batch_size,
crop_size,
shuffle=True,
num_workers=num_workers)
encoder = Encoder(embedding_size).to(device)
decoder = Decoder(vocab_size, embedding_size, lstm_size).to(device)
if os.path.exists(encoder_path):
encoder.load_state_dict(torch.load(encoder_path))
if os.path.exists(decoder_path):
decoder.load_state_dict(torch.load(decoder_path))
loss_fn = torch.nn.CrossEntropyLoss()
parameters = list(encoder.fc.parameters()) + list(
encoder.bn.parameters()) + list(decoder.parameters())
optimizer = torch.optim.Adam(parameters,
lr=learning_rate,
betas=(0.9, 0.99))
num_steps = len(dataloader)
for epoch in range(num_epochs):
for index, (imgs, captions, lengths) in enumerate(dataloader):
imgs = imgs.to(device)
captions = captions.to(device)
targets = pack_padded_sequence(
captions, lengths,
batch_first=True)[0] # the tailing [0] is necessary
features = encoder(imgs)
y_predicted = decoder(features, captions, lengths)
loss = loss_fn(y_predicted, targets)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if index % log_every == 0:
print(
'Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, Perplexity: {:5.4f}'
.format(epoch, num_epochs, index, num_steps, loss.item(),
np.exp(loss.item())))
if index % save_every == 0 and index != 0:
print('Start saving encoder')
torch.save(encoder.state_dict(), encoder_path)
print('Start saving decoder')
torch.save(decoder.state_dict(), decoder_path)
def test(opts):
# Dirs
log_dir = os.path.join("experiments", opts.experiment_name)
checkpoint_dir = os.path.join(log_dir, "checkpoint")
results_dir = os.path.join(log_dir, "results")
# Path to data
image_dir = os.path.join(opts.data_root, opts.dataset_name, "image")
attribute_path = os.path.join(opts.data_root, opts.dataset_name,
"attributes.txt")
test_dataloader = get_loader(image_dir,
attribute_path,
dataset_name=opts.dataset_name,
image_size=opts.img_size,
n_style=opts.n_style,
batch_size=52,
mode='test',
binary=False)
# Model
criterion_pixel = torch.nn.L1Loss().to(device)
generator = GeneratorStyle(n_style=opts.n_style,
attr_channel=opts.attr_channel,
style_out_channel=opts.style_out_channel,
n_res_blocks=opts.n_res_blocks,
attention=opts.attention)
# Attrbute embedding
# attribute: N x 37 -> N x 37 x 64
attribute_embed = nn.Embedding(opts.attr_channel, opts.attr_embed)
# unsupervise font num + 1 dummy id (for supervise)
attr_unsuper_tolearn = nn.Embedding(
opts.unsuper_num + 1, opts.attr_channel) # attribute intensity
if opts.multi_gpu:
generator = nn.DataParallel(generator)
attribute_embed = nn.DataParallel(attribute_embed)
attr_unsuper_tolearn = nn.DataParallel(attr_unsuper_tolearn)
generator = generator.to(device)
attribute_embed = attribute_embed.to(device)
attr_unsuper_tolearn = attr_unsuper_tolearn.to(device)
test_logfile = open(
os.path.join(log_dir, f"test_loss_log_{opts.test_epoch}.txt"), 'w')
if opts.test_epoch == 0:
for test_epoch in range(opts.check_freq, opts.n_epochs + 1,
opts.check_freq):
test_one_epoch(opts, test_logfile, test_epoch, checkpoint_dir,
results_dir, generator, attribute_embed,
attr_unsuper_tolearn, test_dataloader,
criterion_pixel)
else:
test_one_epoch(opts, test_logfile, opts.test_epoch, checkpoint_dir,
results_dir, generator, attribute_embed,
attr_unsuper_tolearn, test_dataloader, criterion_pixel)
def simulation_epoch(self, mode, epoch, simulate_epoch_number):
if mode == 'train':
dataset = get_loader(self.slot_set, self.test_set, self.disease_set, self.goal_test_path, batch_size=self.batch_size, mode = 'train')
else:
dataset = get_loader(self.slot_set, self.test_set, self.disease_set, self.goal_test_path, batch_size=self.batch_size, mode = 'test')
success_count = 0
total_object = 0
total_rewards = 0
total_env = 0
length = 0
total_simulate = 0
# 多核运算
#pool = mp.Pool(self.num_cores)
for i, (origin_state, goal, goal_disease) in enumerate(dataset):
temp_object = 0
origin_state = origin_state.to(self.device)
goal = goal.to(self.device)
goal_disease = goal_disease.to(self.device)
objective_list, flag, turn, reward_total, env_total = self.simulate(origin_state, goal, goal_disease, mode)
# objective_list, flag, status, reward_record.sum(), env_record.sum()
num_simulate = (reward_total != 0).sum().item()
total_simulate = total_simulate + num_simulate
length += origin_state.shape[0]
total_object = total_object + objective_list.sum().item()
total_rewards = total_rewards + reward_total.sum().item()
total_env = total_env + env_total.sum().item()
temp_object = objective_list.sum()
success_count += flag.sum().item()
if mode == 'train':
self.train_network(temp_object)
progress_bar(10, self.success_rate, self.avg_turns, self.avg_object, self.avg_reward, self.avg_envs, self.success_rate_test, self.avg_turns_test, self.avg_reward_test, self.best_success_rate_test, i + epoch * len(dataset), simulate_epoch_number* len(dataset))
self.save(self.parameter['model_savepath'] + '/newest/')
#print(i)
return success_count/length, float(total_simulate)/length, total_object/total_simulate, total_rewards/total_simulate, total_env/total_simulate
def main(config):
os.makedirs(config.sample_dir, exist_ok=True)
os.makedirs(config.checkpoint_dir, exist_ok=True)
config.manual_seed = random.randint(1, 10000)
print("Random Seed: ", config.manual_seed)
random.seed(config.manual_seed)
torch.manual_seed(config.manual_seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(config.manual_seed)
cudnn.benchmark = True
train_loader = get_loader(config.train_dir, config.image_size, config.batch_size)
validation_loader = get_loader(config.valid_dir, config.image_size, config.batch_size)
trainer = Trainer(config, train_loader, validation_loader)
trainer.train()
def get_inference_loader(self, data):
if self.vocabs is None:
print(
"No vocabs for categorical features provided...this is equivalent to randomly embed categorical features"
)
train_loader = dataloader.get_loader(copy.deepcopy(data),
batch_size=self.batch_size,
pad_after=self.pad_after,
vocabs=self.vocabs,
inference=True)
return train_loader
def main(config):
# Data Generation
train_loader, test_loader = get_loader(train_batch_size=config.train_batch,
test_batch_size=config.test_batch,
num_workers=config.workers)
# Create a model, an optimizer, a loss function
solver = Solver(config, train_loader, test_loader)
if config.mode == 'train':
solver.train()
elif config.mode == 'test':
solver.test()
def main(args, scope):
SEED = args.seed
random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
train_loader, val_loader = get_loader(args.dataset,
batch_size=args.batch_size,
num_workers=args.workers)
if args.mode == 'train':
logger.log("Training start!")
args.autoencoder = False
runner = Trainer(train_loader, val_loader, args)
runner.show_current_model()
runner.train()
logger.log("Training end!")
elif args.mode == 'train_adv':
logger.log("Adversarial Training with {}....".format(args.attack))
args.autoencoder = False
runner = AdvTrainer(train_loader, val_loader, args)
runner.show_current_model()
runner.train()
logger.log("Training end!")
elif args.mode == 'train_ae':
logger.log("Training start!")
args.autoencoder = True
args.ckpt_name = args.ckpt_ae
runner = AETrainer(train_loader, val_loader, args)
runner.show_current_model()
runner.train()
logger.log("Training end!")
elif args.mode == 'defense':
logger.log("Defense start!")
args.autoencoder = False
runner = Defender(val_loader, args)
runner.show_current_model()
runner.defend()
logger.log("Defense end!")
arg_file = os.path.join(str(runner.log_path), 'args.json')
with open(arg_file, 'w') as outfile:
json.dump(vars(args), outfile)
def main():
#定义网络
net = models.LeNetWithAngle(classes_num)
if use_gpu:
net = net.cuda()
#定义优化器
optimizer = torch.optim.SGD(net.parameters(),
lr=model_lr,
weight_decay=1e-5,
nesterov=True,
momentum=0.9)
print("net and optimzer load succeed")
#定义数据加载
trainloader, testloader = dataloader.get_loader(batch_size=batch_size,
root_path="./data/MNIST")
print("data load succeed")
#定义logger
logger = utils.Logger(tb_path="./logs/tblog/")
#定义学习率调整器
scheduler = lr_sche.StepLR(optimizer, 30, 0.1)
#定义损失函数
criterion = a_softmax.AngleSoftmaxLoss(gamma=0)
best_acc = 0
#开始训练
for i in range(1, epochs + 1):
scheduler.step(epoch=i)
net.train()
train_acc,train_loss,all_feature,all_labels=\
train(net,optimizer,criterion,trainloader,i)
utils.plot_features(all_feature, all_labels, classes_num, i,
"./logs/images/train/train_{}.png")
net.eval()
test_acc, test_loss, all_feature, all_labels = test(
net, criterion, testloader, i)
utils.plot_features(all_feature, all_labels, classes_num, i,
"./logs/images/test/test_{}.png")
print("{} epoch end, train acc is {:.4f}, test acc is {:.4f}".format(
i, train_acc, test_acc))
content = {
"Train/acc": train_acc,
"Test/acc": test_acc,
"Train/loss": train_loss,
"Test/loss": test_loss
}
logger.log(step=i, content=content)
if best_acc < test_acc:
best_acc = test_acc
utils.save_checkpoints("./logs/weights/net_{}.pth",i,\
net.state_dict(),(best_acc==test_acc))
utils.make_gif("./logs/images/train/", "./logs/train.gif")
utils.make_gif("./logs/images/test/", "./logs/test.gif")
print("Traing finished...")
def main(config):
if config.out_folder is None:
config.out_folder = 'samples'
os.system('mkdir {0}'.format(config.out_folder))
config.manual_seed = random.randint(1, 10000)
print("Random Seed: ", config.manual_seed)
random.seed(config.manual_seed)
torch.manual_seed(config.manual_seed)
torch.cuda.manual_seed_all(config.manual_seed)
cudnn.benchmark = True
dataloader = get_loader(_dataset=config.dataset,
dataroot=config.dataroot,
batch_size=config.batch_size,
num_workers=int(config.workers),
image_size=config.image_size)
print("Prepare dataloader complete!")
# gNet = Generator(config.nz, config.ngf, config.n_classes, config.image_size).to(device)
# dNet = Discriminator(config.ngf, config.n_classes, config.image_size).to(device)
# dNet.weight_init(mean=0, std=0.02)
# for input, label in dataloader:
# print("Input tensor size:", input.size())
# step_batch = input.size(0)
#
# input = input.view(-1, config.image_size**2).to(device)
# label = label.to(device).long()
#
# onehot = torch.zeros(step_batch, config.n_classes)
# onehot.scatter_(1, label.view(step_batch, 1), 1)
#
# random_z = torch.rand((step_batch, config.nz)).to(device)
# random_label = (torch.rand(step_batch, 1) * config.n_classes).long().to(device)
# random_onehot = torch.zeros(step_batch, config.n_classes)
# random_onehot.scatter_(1, random_label.view(step_batch, 1), 1).to(device)
#
# gNet_out = gNet(random_z, random_onehot)
# print("G out size:", gNet_out.size())
#
# dNet_out = dNet(input, onehot)
# print("D out size:", dNet_out.size())
#
# break
trainer = Trainer(config, dataloader)
trainer.train()
def main():
logger.info(args)
model = Net(numAngle=CONFIGS["MODEL"]["NUMANGLE"],
numRho=CONFIGS["MODEL"]["NUMRHO"],
backbone=CONFIGS["MODEL"]["BACKBONE"])
if args.model:
if isfile(args.model):
import torch
m = torch.load(args.model)
if 'state_dict' in m.keys():
m = m['state_dict']
torch.save(m, '_temp_model.pth')
del m
logger.info("=> loading pretrained model '{}'".format(args.model))
#model.load('_temp_model.pth')
logger.info("=> loaded checkpoint '{}'".format(args.model))
else:
logger.info("=> no pretrained model found at '{}'".format(
args.model))
# dataloader
test_loader = get_loader(CONFIGS["DATA"]["TEST_DIR"],
CONFIGS["DATA"]["TEST_LABEL_FILE"],
batch_size=int(os.environ.get("BS", "1")),
num_thread=CONFIGS["DATA"]["WORKERS"],
test=True)
logger.info("Data loading done.")
weights_nodes = {}
data_nodes = {}
def named_dump_func(name):
def dump_func(self, inputs, outputs):
input_name = name + '_input'
output_name = name + '_output'
if isinstance(self, nn.Conv2d):
weights_nodes[name] = self.weight.numpy()
data_nodes[input_name] = inputs[0].numpy()
data_nodes[output_name] = outputs[0].numpy()
return dump_func
if args.dump:
logger.info('Add hooks to dump data.')
for name, module in model.named_modules():
print(name)
module.register_forward_hook(named_dump_func(name))
test(test_loader, model, args)
