def main():
args = cfg.parse_args()
random.seed(args.random_seed)
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed(args.random_seed)
# set tf env
_init_inception()
inception_path = check_or_download_inception(None)
create_inception_graph(inception_path)
# weight init
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv2d') != -1:
if args.init_type == 'normal':
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif args.init_type == 'orth':
nn.init.orthogonal_(m.weight.data)
elif args.init_type == 'xavier_uniform':
nn.init.xavier_uniform(m.weight.data, 1.)
else:
raise NotImplementedError('{} unknown inital type'.format(
args.init_type))
elif classname.find('BatchNorm2d') != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0.0)
gen_net = Generator(bottom_width=args.bottom_width,
gf_dim=args.gf_dim,
latent_dim=args.latent_dim).cuda()
dis_net = eval('models.' + args.model + '.Discriminator')(args=args).cuda()
gen_net.apply(weights_init)
dis_net.apply(weights_init)
initial_gen_net_weight = torch.load(os.path.join(args.init_path,
'initial_gen_net.pth'),
map_location="cpu")
initial_dis_net_weight = torch.load(os.path.join(args.init_path,
'initial_dis_net.pth'),
map_location="cpu")
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
exp_str = args.dir
args.load_path = os.path.join('output', exp_str, 'pth',
'epoch{}.pth'.format(args.load_epoch))
# state dict:
assert os.path.exists(args.load_path)
checkpoint = torch.load(args.load_path)
print('=> loaded checkpoint %s' % args.load_path)
state_dict = checkpoint['generator']
gen_net = load_subnet(args, state_dict, initial_gen_net_weight).cuda()
avg_gen_net = deepcopy(gen_net)
# set optimizer
gen_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, gen_net.parameters()), args.g_lr,
(args.beta1, args.beta2))
dis_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, dis_net.parameters()), args.d_lr,
(args.beta1, args.beta2))
gen_scheduler = LinearLrDecay(gen_optimizer, args.g_lr, 0.0, 0,
args.max_iter * args.n_critic)
dis_scheduler = LinearLrDecay(dis_optimizer, args.d_lr, 0.0, 0,
args.max_iter * args.n_critic)
# set up data_loader
dataset = datasets.ImageDataset(args)
train_loader = dataset.train
# fid stat
if args.dataset.lower() == 'cifar10':
fid_stat = 'fid_stat/fid_stats_cifar10_train.npz'
else:
raise NotImplementedError('no fid stat for %s' % args.dataset.lower())
assert os.path.exists(fid_stat)
# epoch number for dis_net
args.max_epoch = args.max_epoch * args.n_critic
if args.max_iter:
args.max_epoch = np.ceil(args.max_iter * args.n_critic /
len(train_loader))
# initial
np.random.seed(args.random_seed)
fixed_z = torch.cuda.FloatTensor(
np.random.normal(0, 1, (25, args.latent_dim)))
start_epoch = 0
best_fid = 1e4
args.path_helper = set_log_dir('logs', args.exp_name)
logger = create_logger(args.path_helper['log_path'])
#logger.info('=> loaded checkpoint %s (epoch %d)' % (checkpoint_file, start_epoch))
logger.info(args)
writer_dict = {
'writer': SummaryWriter(args.path_helper['log_path']),
'train_global_steps': start_epoch * len(train_loader),
'valid_global_steps': start_epoch // args.val_freq,
}
gen_avg_param = copy_params(gen_net)
# train loop
for epoch in tqdm(range(int(start_epoch), int(args.max_epoch)),
desc='total progress'):
lr_schedulers = (gen_scheduler,
dis_scheduler) if args.lr_decay else None
train(args, gen_net, dis_net, gen_optimizer, dis_optimizer,
gen_avg_param, train_loader, epoch, writer_dict, lr_schedulers)
if epoch and epoch % args.val_freq == 0 or epoch == int(
args.max_epoch) - 1:
backup_param = copy_params(gen_net)
load_params(gen_net, gen_avg_param)
inception_score, fid_score = validate(args, fixed_z, fid_stat,
gen_net, writer_dict)
logger.info(
'Inception score: %.4f, FID score: %.4f || @ epoch %d.' %
(inception_score, fid_score, epoch))
load_params(gen_net, backup_param)
if fid_score < best_fid:
best_fid = fid_score
is_best = True
else:
is_best = False
else:
is_best = False
avg_gen_net.load_state_dict(gen_net.state_dict())
load_params(avg_gen_net, gen_avg_param)
save_checkpoint(
{
'epoch': epoch + 1,
'model': args.model,
'gen_state_dict': gen_net.state_dict(),
'dis_state_dict': dis_net.state_dict(),
'avg_gen_state_dict': avg_gen_net.state_dict(),
'gen_optimizer': gen_optimizer.state_dict(),
'dis_optimizer': dis_optimizer.state_dict(),
'best_fid': best_fid,
'path_helper': args.path_helper
}, is_best, args.path_helper['ckpt_path'])
''' Classifying '''
''' Getting results on testing set '''
print '\n\n ## Testing...'
y_pred = net.predict(dataset['x']['testing'])
c_accuracy = accuracy_score(y_true=np.array(dataset['y']['testing']), y_pred=y_pred)
c_report = classification_report(np.array(dataset['y']['testing']), y_pred)
cm = confusion_matrix(y_true=np.array(dataset['y']['testing']), y_pred=y_pred, labels=net.labels)
cm_normalized = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]
print '\n # Kitt net accuracy score on testing data:', colored(str(c_accuracy), 'green')
print '\n # Kitt net classification report on testing data:\n', colored(str(c_report), 'cyan')
print '\n # Kitt net confusion matrix on testing data:\n'
print_cm(cm=cm, labels=net.labels)
print '\n'
print_cm(cm=cm_normalized, labels=net.labels, normed=True)
terrain_ids = (1, 2, 3, 4, 5, 6, 7, 8, 10, 11, 12, 13, 14, 15)
terrain_types = load_params('terrain_types')[0]
terrains = [terrain_types[str(t_id)] for t_id in terrain_ids]
plt.matshow(cm_normalized, vmin=0, vmax=1)
plt.colorbar()
plt.xticks(range(14), terrains, rotation=45)
plt.yticks(range(14), terrains)
for t1_i, terrain1 in enumerate(terrains):
for t2_i, terrain2 in enumerate(terrains):
if cm_normalized[t1_i][t2_i] >= 0.01:
plt.text(t2_i, t1_i, round(cm_normalized[t1_i][t2_i], 2), va='center', ha='center', fontsize=12)
plt.show()
#plt.savefig('../../thesis/img/amter_classification_nn_cm.eps', bbox_inches='tight', pad_inches=0.1)
dataset.close()
def main():
args = cfg.parse_args()
torch.cuda.manual_seed(args.random_seed)
# set tf env
_init_inception()
inception_path = check_or_download_inception(None)
create_inception_graph(inception_path)
# import network
gen_net = eval('models.' + args.model + '.Generator')(args=args).cuda()
dis_net = eval('models.' + args.model + '.Discriminator')(args=args).cuda()
# weight init
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv2d') != -1:
if args.init_type == 'normal':
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif args.init_type == 'orth':
nn.init.orthogonal_(m.weight.data)
elif args.init_type == 'xavier_uniform':
nn.init.xavier_uniform(m.weight.data, 1.)
else:
raise NotImplementedError('{} unknown inital type'.format(
args.init_type))
elif classname.find('BatchNorm2d') != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0.0)
gen_net.apply(weights_init)
dis_net.apply(weights_init)
# set optimizer
gen_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, gen_net.parameters()), args.g_lr,
(args.beta1, args.beta2))
dis_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, dis_net.parameters()), args.d_lr,
(args.beta1, args.beta2))
gen_scheduler = LinearLrDecay(gen_optimizer, args.g_lr, 0.0, 0,
args.max_iter * args.n_critic)
dis_scheduler = LinearLrDecay(dis_optimizer, args.d_lr, 0.0, 0,
args.max_iter * args.n_critic)
# set up data_loader
dataset = datasets.ImageDataset(args)
train_loader = dataset.train
# fid stat
if args.dataset.lower() == 'cifar10':
fid_stat = 'fid_stat/fid_stats_cifar10_train.npz'
elif args.dataset.lower() == 'stl10':
fid_stat = 'fid_stat/stl10_train_unlabeled_fid_stats_48.npz'
else:
raise NotImplementedError(f'no fid stat for {args.dataset.lower()}')
assert os.path.exists(fid_stat)
# epoch number for dis_net
args.max_epoch = args.max_epoch * args.n_critic
if args.max_iter:
args.max_epoch = np.ceil(args.max_iter * args.n_critic /
len(train_loader))
# initial
fixed_z = torch.cuda.FloatTensor(
np.random.normal(0, 1, (25, args.latent_dim)))
gen_avg_param = copy_params(gen_net)
start_epoch = 0
best_fid = 1e4
# set writer
if args.load_path:
print(f'=> resuming from {args.load_path}')
assert os.path.exists(args.load_path)
checkpoint_file = os.path.join(args.load_path, 'Model',
'checkpoint.pth')
assert os.path.exists(checkpoint_file)
checkpoint = torch.load(checkpoint_file)
start_epoch = checkpoint['epoch']
best_fid = checkpoint['best_fid']
gen_net.load_state_dict(checkpoint['gen_state_dict'])
dis_net.load_state_dict(checkpoint['dis_state_dict'])
gen_optimizer.load_state_dict(checkpoint['gen_optimizer'])
dis_optimizer.load_state_dict(checkpoint['dis_optimizer'])
avg_gen_net = deepcopy(gen_net)
avg_gen_net.load_state_dict(checkpoint['avg_gen_state_dict'])
gen_avg_param = copy_params(avg_gen_net)
del avg_gen_net
args.path_helper = checkpoint['path_helper']
logger = create_logger(args.path_helper['log_path'])
logger.info(
f'=> loaded checkpoint {checkpoint_file} (epoch {start_epoch})')
else:
# create new log dir
assert args.exp_name
args.path_helper = set_log_dir('logs', args.exp_name)
logger = create_logger(args.path_helper['log_path'])
logger.info(args)
writer_dict = {
'writer': SummaryWriter(args.path_helper['log_path']),
'train_global_steps': start_epoch * len(train_loader),
'valid_global_steps': start_epoch // args.val_freq,
}
# train loop
for epoch in tqdm(range(int(start_epoch), int(args.max_epoch)),
desc='total progress'):
lr_schedulers = (gen_scheduler,
dis_scheduler) if args.lr_decay else None
train(args, gen_net, dis_net, gen_optimizer, dis_optimizer,
gen_avg_param, train_loader, epoch, writer_dict, lr_schedulers)
if epoch and epoch % args.val_freq == 0 or epoch == int(
args.max_epoch) - 1:
backup_param = copy_params(gen_net)
load_params(gen_net, gen_avg_param)
inception_score, fid_score = validate(args, fixed_z, fid_stat,
gen_net, writer_dict)
logger.info(
f'Inception score: {inception_score}, FID score: {fid_score} || @ epoch {epoch}.'
)
load_params(gen_net, backup_param)
if fid_score < best_fid:
best_fid = fid_score
is_best = True
else:
is_best = False
else:
is_best = False
avg_gen_net = deepcopy(gen_net)
load_params(avg_gen_net, gen_avg_param)
save_checkpoint(
{
'epoch': epoch + 1,
'model': args.model,
'gen_state_dict': gen_net.state_dict(),
'dis_state_dict': dis_net.state_dict(),
'avg_gen_state_dict': avg_gen_net.state_dict(),
'gen_optimizer': gen_optimizer.state_dict(),
'dis_optimizer': dis_optimizer.state_dict(),
'best_fid': best_fid,
'path_helper': args.path_helper
}, is_best, args.path_helper['ckpt_path'])
del avg_gen_net
def main():
args = cfg.parse_args()
torch.cuda.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
np.random.seed(args.random_seed)
random.seed(args.random_seed)
torch.backends.cudnn.deterministic = True
# set tf env
_init_inception()
inception_path = check_or_download_inception(None)
create_inception_graph(inception_path)
# epoch number for dis_net
dataset = datasets.ImageDataset(args, cur_img_size=8)
train_loader = dataset.train
if args.max_iter:
args.max_epoch = np.ceil(args.max_iter / len(train_loader))
else:
args.max_iter = args.max_epoch * len(train_loader)
args.max_epoch = args.max_epoch * args.n_critic
# import network
gen_net = eval('models.' + args.gen_model + '.Generator')(args=args).cuda()
dis_net = eval('models.' + args.dis_model +
'.Discriminator')(args=args).cuda()
gen_net.set_arch(args.arch, cur_stage=2)
# weight init
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv2d') != -1:
if args.init_type == 'normal':
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif args.init_type == 'orth':
nn.init.orthogonal_(m.weight.data)
elif args.init_type == 'xavier_uniform':
nn.init.xavier_uniform_(m.weight.data, 1.)
else:
raise NotImplementedError('{} unknown inital type'.format(
args.init_type))
elif classname.find('BatchNorm2d') != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0.0)
gen_net.apply(weights_init)
dis_net.apply(weights_init)
gpu_ids = [i for i in range(int(torch.cuda.device_count()))]
gen_net = torch.nn.DataParallel(gen_net.to("cuda:0"), device_ids=gpu_ids)
dis_net = torch.nn.DataParallel(dis_net.to("cuda:0"), device_ids=gpu_ids)
gen_net.module.cur_stage = 0
dis_net.module.cur_stage = 0
gen_net.module.alpha = 1.
dis_net.module.alpha = 1.
# set optimizer
if args.optimizer == "adam":
gen_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, gen_net.parameters()), args.g_lr,
(args.beta1, args.beta2))
dis_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, dis_net.parameters()), args.d_lr,
(args.beta1, args.beta2))
elif args.optimizer == "adamw":
gen_optimizer = AdamW(filter(lambda p: p.requires_grad,
gen_net.parameters()),
args.g_lr,
weight_decay=args.wd)
dis_optimizer = AdamW(filter(lambda p: p.requires_grad,
dis_net.parameters()),
args.g_lr,
weight_decay=args.wd)
gen_scheduler = LinearLrDecay(gen_optimizer, args.g_lr, 0.0, 0,
args.max_iter * args.n_critic)
dis_scheduler = LinearLrDecay(dis_optimizer, args.d_lr, 0.0, 0,
args.max_iter * args.n_critic)
# fid stat
if args.dataset.lower() == 'cifar10':
fid_stat = 'fid_stat/fid_stats_cifar10_train.npz'
elif args.dataset.lower() == 'stl10':
fid_stat = 'fid_stat/stl10_train_unlabeled_fid_stats_48.npz'
elif args.fid_stat is not None:
fid_stat = args.fid_stat
else:
raise NotImplementedError(f'no fid stat for {args.dataset.lower()}')
assert os.path.exists(fid_stat)
# initial
fixed_z = torch.cuda.FloatTensor(
np.random.normal(0, 1, (64, args.latent_dim)))
gen_avg_param = copy_params(gen_net)
start_epoch = 0
best_fid = 1e4
# set writer
if args.load_path:
print(f'=> resuming from {args.load_path}')
assert os.path.exists(args.load_path)
checkpoint_file = os.path.join(args.load_path)
assert os.path.exists(checkpoint_file)
checkpoint = torch.load(checkpoint_file)
start_epoch = checkpoint['epoch']
best_fid = checkpoint['best_fid']
gen_net.load_state_dict(checkpoint['gen_state_dict'])
dis_net.load_state_dict(checkpoint['dis_state_dict'])
gen_optimizer.load_state_dict(checkpoint['gen_optimizer'])
dis_optimizer.load_state_dict(checkpoint['dis_optimizer'])
# avg_gen_net = deepcopy(gen_net)
# avg_gen_net.load_state_dict(checkpoint['avg_gen_state_dict'])
gen_avg_param = checkpoint['gen_avg_param']
# del avg_gen_net
cur_stage = cur_stages(start_epoch, args)
gen_net.module.cur_stage = cur_stage
dis_net.module.cur_stage = cur_stage
gen_net.module.alpha = 1.
dis_net.module.alpha = 1.
args.path_helper = checkpoint['path_helper']
else:
# create new log dir
assert args.exp_name
args.path_helper = set_log_dir('logs', args.exp_name)
logger = create_logger(args.path_helper['log_path'])
logger.info(args)
writer_dict = {
'writer': SummaryWriter(args.path_helper['log_path']),
'train_global_steps': start_epoch * len(train_loader),
'valid_global_steps': start_epoch // args.val_freq,
}
def return_states():
states = {}
states['epoch'] = epoch
states['best_fid'] = best_fid_score
states['gen_state_dict'] = gen_net.state_dict()
states['dis_state_dict'] = dis_net.state_dict()
states['gen_optimizer'] = gen_optimizer.state_dict()
states['dis_optimizer'] = dis_optimizer.state_dict()
states['gen_avg_param'] = gen_avg_param
states['path_helper'] = args.path_helper
return states
# train loop
for epoch in range(start_epoch + 1, args.max_epoch):
train(
args,
gen_net,
dis_net,
gen_optimizer,
dis_optimizer,
gen_avg_param,
train_loader,
epoch,
writer_dict,
fixed_z,
)
backup_param = copy_params(gen_net)
load_params(gen_net, gen_avg_param)
fid_score = validate(
args,
fixed_z,
fid_stat,
epoch,
gen_net,
writer_dict,
)
logger.info(f'FID score: {fid_score} || @ epoch {epoch}.')
load_params(gen_net, backup_param)
is_best = False
if epoch == 1 or fid_score < best_fid_score:
best_fid_score = fid_score
is_best = True
if is_best or epoch % 1 == 0:
states = return_states()
save_checkpoint(states,
is_best,
args.path_helper['ckpt_path'],
filename=f'checkpoint_epoch_{epoch}.pth')
def main():
args = cfg.parse_args()
random.seed(args.random_seed)
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed(args.random_seed)
np.random.seed(args.random_seed)
# set tf env
_init_inception()
inception_path = check_or_download_inception(None)
create_inception_graph(inception_path)
# import netwo
# weight init
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv2d') != -1:
if args.init_type == 'normal':
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif args.init_type == 'orth':
nn.init.orthogonal_(m.weight.data)
elif args.init_type == 'xavier_uniform':
nn.init.xavier_uniform(m.weight.data, 1.)
else:
raise NotImplementedError('{} unknown inital type'.format(
args.init_type))
elif classname.find('BatchNorm2d') != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0.0)
gen_net = eval('models.' + args.model + '.Generator')(args=args).cuda()
dis_net = eval('models.' + args.model + '.Discriminator')(args=args).cuda()
gen_net.apply(weights_init)
dis_net.apply(weights_init)
avg_gen_net = deepcopy(gen_net)
initial_gen_net_weight = torch.load(os.path.join(args.init_path,
'initial_gen_net.pth'),
map_location="cpu")
initial_dis_net_weight = torch.load(os.path.join(args.init_path,
'initial_dis_net.pth'),
map_location="cpu")
assert id(initial_dis_net_weight) != id(dis_net.state_dict())
assert id(initial_gen_net_weight) != id(gen_net.state_dict())
# set optimizer
gen_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, gen_net.parameters()), args.g_lr,
(args.beta1, args.beta2))
dis_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, dis_net.parameters()), args.d_lr,
(args.beta1, args.beta2))
gen_scheduler = LinearLrDecay(gen_optimizer, args.g_lr, 0.0, 0,
args.max_iter * args.n_critic)
dis_scheduler = LinearLrDecay(dis_optimizer, args.d_lr, 0.0, 0,
args.max_iter * args.n_critic)
# set up data_loader
dataset = datasets.ImageDataset(args)
train_loader = dataset.train
# fid stat
if args.dataset.lower() == 'cifar10':
fid_stat = 'fid_stat/fid_stats_cifar10_train.npz'
elif args.dataset.lower() == 'stl10':
fid_stat = 'fid_stat/fid_stats_stl10_train.npz'
else:
raise NotImplementedError('no fid stat for %s' % args.dataset.lower())
assert os.path.exists(fid_stat)
# epoch number for dis_net
args.max_epoch = args.max_epoch * args.n_critic
if args.max_iter:
args.max_epoch = np.ceil(args.max_iter * args.n_critic /
len(train_loader))
# initial
fixed_z = torch.cuda.FloatTensor(
np.random.normal(0, 1, (25, args.latent_dim)))
start_epoch = 0
best_fid = 1e4
print('=> resuming from %s' % args.load_path)
assert os.path.exists(args.load_path)
checkpoint_file = args.load_path
assert os.path.exists(checkpoint_file)
checkpoint = torch.load(checkpoint_file)
pruning_generate(gen_net, checkpoint['gen_state_dict'])
dis_net.load_state_dict(checkpoint['dis_state_dict'])
total = 0
total_nonzero = 0
for m in dis_net.modules():
if isinstance(m, nn.Conv2d):
total += m.weight_orig.data.numel()
mask = m.weight_orig.data.abs().clone().gt(0).float().cuda()
total_nonzero += torch.sum(mask)
conv_weights = torch.zeros(total)
index = 0
for m in dis_net.modules():
if isinstance(m, nn.Conv2d):
size = m.weight_orig.data.numel()
conv_weights[index:(
index + size)] = m.weight_orig.data.view(-1).abs().clone()
index += size
y, i = torch.sort(conv_weights)
# thre_index = int(total * args.percent)
# only care about the non zero weights
# e.g: total = 100, total_nonzero = 80, percent = 0.2, thre_index = 36, that means keep 64
thre_index = total - total_nonzero
thre = y[int(thre_index)]
pruned = 0
print('Pruning threshold: {}'.format(thre))
zero_flag = False
masks = OrderedDict()
for k, m in enumerate(dis_net.modules()):
if isinstance(m, nn.Conv2d):
weight_copy = m.weight_orig.data.abs().clone()
mask = weight_copy.gt(thre).float()
masks[k] = mask
pruned = pruned + mask.numel() - torch.sum(mask)
m.weight_orig.data.mul_(mask)
if int(torch.sum(mask)) == 0:
zero_flag = True
print(
'layer index: {:d} \t total params: {:d} \t remaining params: {:d}'
.format(k, mask.numel(), int(torch.sum(mask))))
print('Total conv params: {}, Pruned conv params: {}, Pruned ratio: {}'.
format(total, pruned, pruned / total))
pruning_generate(avg_gen_net, checkpoint['gen_state_dict'])
see_remain_rate(gen_net)
if not args.finetune_G:
gen_weight = gen_net.state_dict()
gen_orig_weight = rewind_weight(initial_gen_net_weight,
gen_weight.keys())
gen_weight.update(gen_orig_weight)
gen_net.load_state_dict(gen_weight)
gen_avg_param = copy_params(gen_net)
if args.finetune_D:
dis_net.load_state_dict(checkpoint['dis_state_dict'])
else:
dis_net.load_state_dict(initial_dis_net_weight)
for k, m in enumerate(dis_net.modules()):
if isinstance(m, nn.Conv2d):
m.weight_orig.data.mul_(masks[k])
orig_dis_net = eval('models.' + args.model +
'.Discriminator')(args=args).cuda()
orig_dis_net.load_state_dict(checkpoint['dis_state_dict'])
orig_dis_net.eval()
args.path_helper = set_log_dir('logs',
args.exp_name + "_{}".format(args.percent))
logger = create_logger(args.path_helper['log_path'])
#logger.info('=> loaded checkpoint %s (epoch %d)' % (checkpoint_file, start_epoch))
logger.info(args)
writer_dict = {
'writer': SummaryWriter(args.path_helper['log_path']),
'train_global_steps': start_epoch * len(train_loader),
'valid_global_steps': start_epoch // args.val_freq,
}
# train loop
for epoch in tqdm(range(int(start_epoch), int(args.max_epoch)),
desc='total progress'):
lr_schedulers = (gen_scheduler,
dis_scheduler) if args.lr_decay else None
see_remain_rate(gen_net)
see_remain_rate_orig(dis_net)
if not args.use_kd_D:
train_with_mask(args, gen_net, dis_net, gen_optimizer,
dis_optimizer, gen_avg_param, train_loader, epoch,
writer_dict, masks, lr_schedulers)
else:
train_with_mask_kd(args, gen_net, dis_net, orig_dis_net,
gen_optimizer, dis_optimizer, gen_avg_param,
train_loader, epoch, writer_dict, masks,
lr_schedulers)
if epoch and epoch % args.val_freq == 0 or epoch == int(
args.max_epoch) - 1:
backup_param = copy_params(gen_net)
load_params(gen_net, gen_avg_param)
inception_score, fid_score = validate(args, fixed_z, fid_stat,
gen_net, writer_dict, epoch)
logger.info(
'Inception score: %.4f, FID score: %.4f || @ epoch %d.' %
(inception_score, fid_score, epoch))
load_params(gen_net, backup_param)
if fid_score < best_fid:
best_fid = fid_score
is_best = True
else:
is_best = False
else:
is_best = False
avg_gen_net.load_state_dict(gen_net.state_dict())
load_params(avg_gen_net, gen_avg_param)
save_checkpoint(
{
'epoch': epoch + 1,
'model': args.model,
'gen_state_dict': gen_net.state_dict(),
'dis_state_dict': dis_net.state_dict(),
'avg_gen_state_dict': avg_gen_net.state_dict(),
'gen_optimizer': gen_optimizer.state_dict(),
'dis_optimizer': dis_optimizer.state_dict(),
'best_fid': best_fid,
'path_helper': args.path_helper
}, is_best, args.path_helper['ckpt_path'])
def main():
args = cfg.parse_args()
random.seed(args.random_seed)
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed(args.random_seed)
# set tf env
_init_inception()
inception_path = check_or_download_inception(None)
create_inception_graph(inception_path)
# weight init
gen_net = eval('models.' + args.model + '.Generator')(args=args)
dis_net = eval('models.' + args.model + '.Discriminator')(args=args)
# weight init
def weights_init(m):
if isinstance(m, nn.Conv2d):
if args.init_type == 'normal':
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif args.init_type == 'orth':
nn.init.orthogonal_(m.weight.data)
elif args.init_type == 'xavier_uniform':
nn.init.xavier_uniform(m.weight.data, 1.)
else:
raise NotImplementedError('{} unknown inital type'.format(
args.init_type))
elif isinstance(m, nn.BatchNorm2d):
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0.0)
gen_net.apply(weights_init)
dis_net.apply(weights_init)
gen_net = gen_net.cuda()
dis_net = dis_net.cuda()
avg_gen_net = deepcopy(gen_net)
initial_gen_net_weight = deepcopy(gen_net.state_dict())
initial_dis_net_weight = deepcopy(dis_net.state_dict())
assert id(initial_dis_net_weight) != id(dis_net.state_dict())
assert id(initial_gen_net_weight) != id(gen_net.state_dict())
# set optimizer
gen_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, gen_net.parameters()), args.g_lr,
(args.beta1, args.beta2))
dis_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, dis_net.parameters()), args.d_lr,
(args.beta1, args.beta2))
gen_scheduler = LinearLrDecay(gen_optimizer, args.g_lr, 0.0, 0,
args.max_iter * args.n_critic)
dis_scheduler = LinearLrDecay(dis_optimizer, args.d_lr, 0.0, 0,
args.max_iter * args.n_critic)
# set up data_loader
dataset = datasets.ImageDataset(args)
train_loader = dataset.train
# fid stat
if args.dataset.lower() == 'cifar10':
fid_stat = 'fid_stat/fid_stats_cifar10_train.npz'
elif args.dataset.lower() == 'stl10':
fid_stat = 'fid_stat/fid_stats_stl10_train.npz'
else:
raise NotImplementedError('no fid stat for %s' % args.dataset.lower())
assert os.path.exists(fid_stat)
# epoch number for dis_net
args.max_epoch = args.max_epoch * args.n_critic
if args.max_iter:
args.max_epoch = np.ceil(args.max_iter * args.n_critic /
len(train_loader))
# initial
np.random.seed(args.random_seed)
fixed_z = torch.cuda.FloatTensor(
np.random.normal(0, 1, (25, args.latent_dim)))
start_epoch = 0
best_fid = 1e4
args.path_helper = set_log_dir('logs',
args.exp_name + "_{}".format(args.percent))
logger = create_logger(args.path_helper['log_path'])
# logger.info('=> loaded checkpoint %s (epoch %d)' % (checkpoint_file, start_epoch))
logger.info(args)
writer_dict = {
'writer': SummaryWriter(args.path_helper['log_path']),
'train_global_steps': start_epoch * len(train_loader),
'valid_global_steps': start_epoch // args.val_freq,
}
print('=> resuming from %s' % args.load_path)
assert os.path.exists(args.load_path)
checkpoint_file = os.path.join(args.load_path, 'Model', 'checkpoint.pth')
assert os.path.exists(checkpoint_file)
checkpoint = torch.load(checkpoint_file)
gen_net.load_state_dict(checkpoint['gen_state_dict'])
torch.manual_seed(args.random_seed)
pruning_generate(gen_net, (1 - args.percent), args.pruning_method)
torch.manual_seed(args.random_seed)
pruning_generate(avg_gen_net, (1 - args.percent), args.pruning_method)
see_remain_rate(gen_net)
if args.second_seed:
dis_net.apply(weights_init)
if args.finetune_D:
dis_net.load_state_dict(checkpoint['dis_state_dict'])
else:
dis_net.load_state_dict(initial_dis_net_weight)
gen_weight = gen_net.state_dict()
gen_orig_weight = rewind_weight(initial_gen_net_weight, gen_weight.keys())
assert id(gen_weight) != id(gen_orig_weight)
gen_weight.update(gen_orig_weight)
gen_net.load_state_dict(gen_weight)
gen_avg_param = copy_params(gen_net)
if args.use_kd_D:
orig_dis_net = eval('models.' + args.model +
'.Discriminator')(args=args).cuda()
orig_dis_net.load(checkpoint['dis_state_dict'])
orig_dis_net.eval()
# train loop
for epoch in tqdm(range(int(start_epoch), int(args.max_epoch)),
desc='total progress'):
lr_schedulers = (gen_scheduler,
dis_scheduler) if args.lr_decay else None
see_remain_rate(gen_net)
if not args.use_kd_D:
train(args, gen_net, dis_net, gen_optimizer, dis_optimizer,
gen_avg_param, train_loader, epoch, writer_dict,
lr_schedulers)
else:
train_kd(args, gen_net, dis_net, orig_dis_net, gen_optimizer,
dis_optimizer, gen_avg_param, train_loader, epoch,
writer_dict, lr_schedulers)
if epoch and epoch % args.val_freq == 0 or epoch == int(
args.max_epoch) - 1:
backup_param = copy_params(gen_net)
load_params(gen_net, gen_avg_param)
inception_score, fid_score = validate(args, fixed_z, fid_stat,
gen_net, writer_dict, epoch)
logger.info(
'Inception score: %.4f, FID score: %.4f || @ epoch %d.' %
(inception_score, fid_score, epoch))
load_params(gen_net, backup_param)
if fid_score < best_fid:
best_fid = fid_score
is_best = True
else:
is_best = False
else:
is_best = False
avg_gen_net.load_state_dict(gen_net.state_dict())
load_params(avg_gen_net, gen_avg_param)
save_checkpoint(
{
'epoch': epoch + 1,
'model': args.model,
'gen_state_dict': gen_net.state_dict(),
'dis_state_dict': dis_net.state_dict(),
'avg_gen_state_dict': avg_gen_net.state_dict(),
'gen_optimizer': gen_optimizer.state_dict(),
'dis_optimizer': dis_optimizer.state_dict(),
'best_fid': best_fid,
'path_helper': args.path_helper
}, is_best, args.path_helper['ckpt_path'])
:return: normalized signal with a signal noise
"""
global signal_noise_std
''' First, normalize the signal '''
normed_signal = norm_signal(signal=signal, the_min=sensors_ranges[sen][0], the_max=sensors_ranges[sen][1])
''' Adding signal noise of defined std '''
noised_signal = add_signal_noise(signal=normed_signal, std=signal_noise_std)
return noised_signal
if __name__ == '__main__':
terrain_types, all_sensors, sensors_ranges, noise_types, noise_params = \
load_params('terrain_types', 'sensors', 'sensors_ranges', 'noise_types', 'noise_params')
args = parse_arguments()
terrains_to_use = [terrain_types[str(i)] for i in [1, 2, 3, 4, 5, 6, 7, 8, 10, 11, 12, 13, 14, 15] if i not in args.rem_terrains]
if not args.rem_terrains:
terrains_flag = 'allt'
else:
terrains_flag = 'rt'
for t_i in sorted(args.rem_terrains):
terrains_flag += str(t_i)
sensors_to_use = all_sensors
if args.sensors == 'angle':
sensors_to_use = all_sensors[:18]
elif args.sensors == 'foot':
sensors_to_use = all_sensors[18:]
help='Number of simulation runs')
parser.add_argument('-t', '--terrains', type=int, default=range(1, 16), nargs='+', choices=range(1, 16),
help='Terrains to be generated (integers)')
parser.add_argument('-n', '--noise', type=str, default='no_noise', choices=noise_types,
help='Terrain noise type')
parser.add_argument('-nt', '--n_timesteps', type=int, default=100,
help='Number of simulation steps')
parser.add_argument('-g', '--gait', type=str, default='tripod', choices=['tripod'],
help='Amos II gait')
parser.add_argument('-sn', '--sim_noise', type=float, default=0.0,
help='Amos II simulation noise')
return parser.parse_args()
if __name__ == '__main__':
terrain_types, noise_types, noise_params = load_params('terrain_types', 'noise_types', 'noise_params')
args = parse_arguments()
noise_type, (noise_prefix, noise_param) = args.noise, noise_params[args.noise]
gait = args.gait
n_jobs = args.n_jobs
n_timesteps = args.n_timesteps
sim_noise = args.sim_noise
terrains_i = sorted(args.terrains)
terrains_to_use = [terrain_types[str(i)] for i in terrains_i]
os.chdir('../../simulation/mbulinai22015-gorobots_edu-fork/practices/amosii')
for i_terrain, terrain_type in zip(terrains_i, terrains_to_use):
destination_dir = '../../../../data/'+noise_type+'/'+noise_prefix+terrain_type+'/'
if not os.path.exists(destination_dir):
def main():
args = cfg_train.parse_args()
torch.cuda.manual_seed(args.random_seed)
# set tf env
_init_inception()
inception_path = check_or_download_inception(None)
create_inception_graph(inception_path)
# import network
# gen_net = eval('models.' + args.gen_model + '.' + args.gen)(args=args).cuda()
genotype_gen = eval('genotypes.%s' % args.arch_gen)
gen_net = eval('models.' + args.gen_model + '.' + args.gen)(
args, genotype_gen).cuda()
# gen_net = eval('models.' + args.gen_model + '.' + args.gen)(args = args).cuda()
if 'Discriminator' not in args.dis:
genotype_dis = eval('genotypes.%s' % args.arch_dis)
dis_net = eval('models.' + args.dis_model + '.' + args.dis)(
args, genotype_dis).cuda()
else:
dis_net = eval('models.' + args.dis_model + '.' +
args.dis)(args=args).cuda()
# weight init
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv2d') != -1:
if args.init_type == 'normal':
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif args.init_type == 'orth':
nn.init.orthogonal_(m.weight.data)
elif args.init_type == 'xavier_uniform':
nn.init.xavier_uniform(m.weight.data, 1.)
else:
raise NotImplementedError('{} unknown inital type'.format(
args.init_type))
elif classname.find('BatchNorm2d') != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0.0)
gen_net.apply(weights_init)
dis_net.apply(weights_init)
# set up data_loader
dataset = datasets.ImageDataset(args)
train_loader = dataset.train
val_loader = dataset.valid
# set optimizer
gen_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, gen_net.parameters()), args.g_lr,
(args.beta1, args.beta2))
dis_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, dis_net.parameters()), args.d_lr,
(args.beta1, args.beta2))
gen_scheduler = LinearLrDecay(gen_optimizer, args.g_lr, args.g_lr * 0.01,
260 * len(train_loader),
args.max_iter * args.n_critic)
dis_scheduler = LinearLrDecay(dis_optimizer, args.d_lr, args.d_lr * 0.01,
260 * len(train_loader),
args.max_iter * args.n_critic)
# fid stat
if args.dataset.lower() == 'cifar10':
fid_stat = 'fid_stat/fid_stats_cifar10_train.npz'
elif args.dataset.lower() == 'stl10':
fid_stat = 'fid_stat/stl10_train_unlabeled_fid_stats_48.npz'
elif args.dataset.lower() == 'mnist':
fid_stat = 'fid_stat/stl10_train_unlabeled_fid_stats_48.npz'
else:
raise NotImplementedError(f'no fid stat for {args.dataset.lower()}')
assert os.path.exists(fid_stat)
# epoch number for dis_net
args.max_epoch = args.max_epoch * args.n_critic
if args.max_iter:
args.max_epoch = np.ceil(args.max_iter * args.n_critic /
len(train_loader))
# initial
fixed_z = torch.cuda.FloatTensor(
np.random.normal(0, 1, (25, args.latent_dim)))
fixed_z_sample = torch.cuda.FloatTensor(
np.random.normal(0, 1, (args.eval_batch_size, args.latent_dim)))
gen_avg_param = copy_params(gen_net)
start_epoch = 0
best_fid = 1e4
best_fid_epoch = 0
is_with_fid = 0
std_with_fid = 0.
best_is = 0
best_is_epoch = 0
fid_with_is = 0
best_dts = 0
# set writer
if args.load_path:
print(f'=> resuming from {args.load_path}')
assert os.path.exists(args.load_path)
checkpoint_file = os.path.join(args.load_path, 'Model',
'checkpoint.pth')
assert os.path.exists(checkpoint_file)
checkpoint = torch.load(checkpoint_file)
start_epoch = checkpoint['epoch']
best_fid = checkpoint['best_fid']
gen_net.load_state_dict(checkpoint['gen_state_dict'])
dis_net.load_state_dict(checkpoint['dis_state_dict'])
gen_optimizer.load_state_dict(checkpoint['gen_optimizer'])
dis_optimizer.load_state_dict(checkpoint['dis_optimizer'])
avg_gen_net = deepcopy(gen_net)
avg_gen_net.load_state_dict(checkpoint['avg_gen_state_dict'])
gen_avg_param = copy_params(avg_gen_net)
del avg_gen_net
args.path_helper = checkpoint['path_helper']
logger = create_logger(args.path_helper['log_path'])
logger.info(
f'=> loaded checkpoint {checkpoint_file} (epoch {start_epoch})')
else:
# create new log dir
assert args.exp_name
args.path_helper = set_log_dir('logs', args.exp_name)
logger = create_logger(args.path_helper['log_path'])
logger.info(args)
writer_dict = {
'writer': SummaryWriter(args.path_helper['log_path']),
'train_global_steps': start_epoch * len(train_loader),
'valid_global_steps': start_epoch // args.val_freq,
}
# calculate the FLOPs and param count of G
input = torch.randn(args.gen_batch_size, args.latent_dim).cuda()
flops, params = profile(gen_net, inputs=(input, ))
flops, params = clever_format([flops, params], "%.3f")
logger.info('FLOPs is {}, param count is {}'.format(flops, params))
# train loop
dg_list = []
worst_lr = 1e-5
for epoch in tqdm(range(int(start_epoch), int(args.max_epoch)),
desc='total progress'):
lr_schedulers = (gen_scheduler,
dis_scheduler) if args.lr_decay else None
train(args, gen_net, dis_net, gen_optimizer, dis_optimizer,
gen_avg_param, train_loader, epoch, writer_dict, args.consistent,
lr_schedulers)
if epoch and epoch % args.val_freq == 0 or epoch == int(
args.max_epoch) - 1:
backup_param = copy_params(gen_net)
load_params(gen_net, gen_avg_param)
inception_score, std, fid_score = validate(args,
fixed_z,
fid_stat,
gen_net,
writer_dict,
args.path_helper,
search=False)
logger.info(
f'Inception score: {inception_score}, FID score: {fid_score}+-{std} || @ epoch {epoch}.'
)
load_params(gen_net, backup_param)
if fid_score < best_fid:
best_fid = fid_score
best_fid_epoch = epoch
is_with_fid = inception_score
std_with_fid = std
is_best = True
else:
is_best = False
if inception_score > best_is:
best_is = inception_score
best_std = std
fid_with_is = fid_score
best_is_epoch = epoch
else:
is_best = False
# save generated images
if epoch % args.image_every == 0:
gen_noise = torch.cuda.FloatTensor(
np.random.normal(0, 1,
(args.eval_batch_size, args.latent_dim)))
# gen_images = gen_net(fixed_z_sample)
# gen_images = gen_images.reshape(args.eval_batch_size, 32, 32, 3)
# gen_images = gen_images.cpu().detach()
gen_images = gen_net(fixed_z_sample).mul_(127.5).add_(
127.5).clamp_(0.0, 255.0).permute(0, 2, 3,
1).to('cpu',
torch.uint8).numpy()
fig = plt.figure()
grid = ImageGrid(fig, 111, nrows_ncols=(10, 10), axes_pad=0)
for x in range(args.eval_batch_size):
grid[x].imshow(gen_images[x]) # cmap="gray")
grid[x].set_xticks([])
grid[x].set_yticks([])
plt.savefig(
os.path.join(args.path_helper['sample_path'],
"epoch_{}.png".format(epoch)))
plt.close()
avg_gen_net = deepcopy(gen_net)
# avg_gen_net = eval('models.'+args.gen_model+'.' + args.gen)(args, genotype_gen).cuda()
# avg_gen_net = eval('models.' + args.gen_model + '.' + args.gen)(args=args).cuda()
load_params(avg_gen_net, gen_avg_param)
save_checkpoint(
{
'epoch': epoch + 1,
'gen_model': args.gen_model,
'dis_model': args.dis_model,
'gen_state_dict': gen_net.state_dict(),
'dis_state_dict': dis_net.state_dict(),
'avg_gen_state_dict': avg_gen_net.state_dict(),
'gen_optimizer': gen_optimizer.state_dict(),
'dis_optimizer': dis_optimizer.state_dict(),
'best_fid': best_fid,
'path_helper': args.path_helper
}, is_best, args.path_helper['ckpt_path'])
del avg_gen_net
logger.info(
'best_is is {}+-{}@{} epoch, fid is {}, best_fid is {}@{}, is is {}+-{}'
.format(best_is, best_std, best_is_epoch, fid_with_is, best_fid,
best_fid_epoch, is_with_fid, std_with_fid))
def main():
args = cfg.parse_args()
random.seed(args.random_seed)
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed(args.random_seed)
np.random.seed(args.random_seed)
torch.backends.cudnn.deterministic = False
torch.backends.cudnn.benchmark = True
os.environ['PYTHONHASHSEED'] = str(args.random_seed)
# set tf env
_init_inception()
inception_path = check_or_download_inception(None)
create_inception_graph(inception_path)
# import network
gen_net = eval('models.'+args.model+'.Generator')(args=args)
dis_net = eval('models.'+args.model+'.Discriminator')(args=args)
initial_gen_net_weight = torch.load(os.path.join(args.init_path, 'initial_gen_net.pth'), map_location="cpu")
initial_dis_net_weight = torch.load(os.path.join(args.init_path, 'initial_dis_net.pth'), map_location="cpu")
gen_net = gen_net.cuda()
dis_net = dis_net.cuda()
gen_net.load_state_dict(initial_gen_net_weight)
dis_net.load_state_dict(initial_dis_net_weight)
# set optimizer
gen_optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, gen_net.parameters()),
args.g_lr, (args.beta1, args.beta2))
dis_optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, dis_net.parameters()),
args.d_lr, (args.beta1, args.beta2))
gen_scheduler = LinearLrDecay(gen_optimizer, args.g_lr, 0.0, 0, args.max_iter * args.n_critic)
dis_scheduler = LinearLrDecay(dis_optimizer, args.d_lr, 0.0, 0, args.max_iter * args.n_critic)
# set up data_loader
dataset = datasets.ImageDataset(args)
train_loader = dataset.train
# fid stat
if args.dataset.lower() == 'cifar10':
fid_stat = 'fid_stat/fid_stats_cifar10_train.npz'
elif args.dataset.lower() == 'stl10':
fid_stat = 'fid_stat/fid_stats_stl10_train.npz'
else:
raise NotImplementedError('no fid stat for %s' % args.dataset.lower())
assert os.path.exists(fid_stat)
# epoch number for dis_net
args.max_epoch = args.max_epoch * args.n_critic
if args.max_iter:
args.max_epoch = np.ceil(args.max_iter * args.n_critic / len(train_loader))
# initial
fixed_z = torch.cuda.FloatTensor(np.random.normal(0, 1, (25, args.latent_dim)))
gen_avg_param = copy_params(gen_net)
start_epoch = 0
best_fid = 1e4
# set writer
if args.load_path:
print('=> resuming from %s' % args.load_path)
assert os.path.exists(args.load_path)
checkpoint_file = os.path.join(args.load_path, 'Model', 'checkpoint.pth')
assert os.path.exists(checkpoint_file)
checkpoint = torch.load(checkpoint_file)
start_epoch = checkpoint['epoch']
best_fid = checkpoint['best_fid']
gen_net.load_state_dict(checkpoint['gen_state_dict'])
dis_net.load_state_dict(checkpoint['dis_state_dict'])
gen_optimizer.load_state_dict(checkpoint['gen_optimizer'])
dis_optimizer.load_state_dict(checkpoint['dis_optimizer'])
avg_gen_net = deepcopy(gen_net)
avg_gen_net.load_state_dict(checkpoint['avg_gen_state_dict'])
gen_avg_param = copy_params(avg_gen_net)
del avg_gen_net
args.path_helper = checkpoint['path_helper']
logger = create_logger(args.path_helper['log_path'])
logger.info('=> loaded checkpoint %s (epoch %d)' % (checkpoint_file, start_epoch))
else:
# create new log dir
assert args.exp_name
args.path_helper = set_log_dir('logs', args.exp_name)
logger = create_logger(args.path_helper['log_path'])
logger.info(args)
writer_dict = {
'writer': SummaryWriter(args.path_helper['log_path']),
'train_global_steps': start_epoch * len(train_loader),
'valid_global_steps': start_epoch // args.val_freq,
}
# train loop
switch = False
for epoch in range(int(start_epoch), int(args.max_epoch)):
lr_schedulers = (gen_scheduler, dis_scheduler) if args.lr_decay else None
train(args, gen_net, dis_net, gen_optimizer, dis_optimizer, gen_avg_param, train_loader, epoch, writer_dict,
lr_schedulers)
if epoch and epoch % args.val_freq == 0 or epoch == int(args.max_epoch)-1:
backup_param = copy_params(gen_net)
load_params(gen_net, gen_avg_param)
inception_score, fid_score = validate(args, fixed_z, fid_stat, gen_net, writer_dict, epoch)
logger.info('Inception score: %.4f, FID score: %.4f || @ epoch %d.' % (inception_score, fid_score, epoch))
load_params(gen_net, backup_param)
if fid_score < best_fid:
best_fid = fid_score
is_best = True
else:
is_best = False
else:
is_best = False
avg_gen_net = deepcopy(gen_net)
load_params(avg_gen_net, gen_avg_param)
save_checkpoint({
'epoch': epoch + 1,
'model': args.model,
'gen_state_dict': gen_net.state_dict(),
'dis_state_dict': dis_net.state_dict(),
'avg_gen_state_dict': avg_gen_net.state_dict(),
'gen_optimizer': gen_optimizer.state_dict(),
'dis_optimizer': dis_optimizer.state_dict(),
'best_fid': best_fid,
'path_helper': args.path_helper,
'seed': args.random_seed
}, is_best, args.path_helper['ckpt_path'])
del avg_gen_net
def main():
args = cfg.parse_args()
torch.cuda.manual_seed(args.random_seed)
# set tf env
_init_inception()
inception_path = check_or_download_inception(None)
create_inception_graph(inception_path)
# import network
gen_net = eval("models_search." + args.gen_model +
".Generator")(args=args).cuda()
dis_net = eval("models_search." + args.dis_model +
".Discriminator")(args=args).cuda()
gen_net.set_arch(args.arch, cur_stage=2)
dis_net.cur_stage = 2
# weight init
def weights_init(m):
classname = m.__class__.__name__
if classname.find("Conv2d") != -1:
if args.init_type == "normal":
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif args.init_type == "orth":
nn.init.orthogonal_(m.weight.data)
elif args.init_type == "xavier_uniform":
nn.init.xavier_uniform(m.weight.data, 1.0)
else:
raise NotImplementedError("{} unknown inital type".format(
args.init_type))
elif classname.find("BatchNorm2d") != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0.0)
gen_net.apply(weights_init)
dis_net.apply(weights_init)
# set optimizer
gen_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, gen_net.parameters()),
args.g_lr,
(args.beta1, args.beta2),
)
dis_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, dis_net.parameters()),
args.d_lr,
(args.beta1, args.beta2),
)
gen_scheduler = LinearLrDecay(gen_optimizer, args.g_lr, 0.0, 0,
args.max_iter * args.n_critic)
dis_scheduler = LinearLrDecay(dis_optimizer, args.d_lr, 0.0, 0,
args.max_iter * args.n_critic)
# set up data_loader
dataset = datasets.ImageDataset(args)
train_loader = dataset.train
# fid stat
if args.dataset.lower() == "cifar10":
fid_stat = "fid_stat/fid_stats_cifar10_train.npz"
elif args.dataset.lower() == "stl10":
fid_stat = "fid_stat/stl10_train_unlabeled_fid_stats_48.npz"
else:
raise NotImplementedError(f"no fid stat for {args.dataset.lower()}")
assert os.path.exists(fid_stat)
# epoch number for dis_net
args.max_epoch = args.max_epoch * args.n_critic
if args.max_iter:
args.max_epoch = np.ceil(args.max_iter * args.n_critic /
len(train_loader))
# initial
fixed_z = torch.cuda.FloatTensor(
np.random.normal(0, 1, (25, args.latent_dim)))
gen_avg_param = copy_params(gen_net)
start_epoch = 0
best_fid = 1e4
# set writer
if args.load_path:
print(f"=> resuming from {args.load_path}")
assert os.path.exists(args.load_path)
checkpoint_file = os.path.join(args.load_path, "Model",
"checkpoint.pth")
assert os.path.exists(checkpoint_file)
checkpoint = torch.load(checkpoint_file)
start_epoch = checkpoint["epoch"]
best_fid = checkpoint["best_fid"]
gen_net.load_state_dict(checkpoint["gen_state_dict"])
dis_net.load_state_dict(checkpoint["dis_state_dict"])
gen_optimizer.load_state_dict(checkpoint["gen_optimizer"])
dis_optimizer.load_state_dict(checkpoint["dis_optimizer"])
avg_gen_net = deepcopy(gen_net)
avg_gen_net.load_state_dict(checkpoint["avg_gen_state_dict"])
gen_avg_param = copy_params(avg_gen_net)
del avg_gen_net
args.path_helper = checkpoint["path_helper"]
logger = create_logger(args.path_helper["log_path"])
logger.info(
f"=> loaded checkpoint {checkpoint_file} (epoch {start_epoch})")
else:
# create new log dir
assert args.exp_name
args.path_helper = set_log_dir("logs", args.exp_name)
logger = create_logger(args.path_helper["log_path"])
logger.info(args)
writer_dict = {
"writer": SummaryWriter(args.path_helper["log_path"]),
"train_global_steps": start_epoch * len(train_loader),
"valid_global_steps": start_epoch // args.val_freq,
}
# train loop
for epoch in tqdm(range(int(start_epoch), int(args.max_epoch)),
desc="total progress"):
lr_schedulers = (gen_scheduler,
dis_scheduler) if args.lr_decay else None
train(
args,
gen_net,
dis_net,
gen_optimizer,
dis_optimizer,
gen_avg_param,
train_loader,
epoch,
writer_dict,
lr_schedulers,
)
if epoch and epoch % args.val_freq == 0 or epoch == int(
args.max_epoch) - 1:
backup_param = copy_params(gen_net)
load_params(gen_net, gen_avg_param)
inception_score, fid_score = validate(args, fixed_z, fid_stat,
gen_net, writer_dict)
logger.info(
f"Inception score: {inception_score}, FID score: {fid_score} || @ epoch {epoch}."
)
load_params(gen_net, backup_param)
if fid_score < best_fid:
best_fid = fid_score
is_best = True
else:
is_best = False
else:
is_best = False
avg_gen_net = deepcopy(gen_net)
load_params(avg_gen_net, gen_avg_param)
save_checkpoint(
{
"epoch": epoch + 1,
"gen_model": args.gen_model,
"dis_model": args.dis_model,
"gen_state_dict": gen_net.state_dict(),
"dis_state_dict": dis_net.state_dict(),
"avg_gen_state_dict": avg_gen_net.state_dict(),
"gen_optimizer": gen_optimizer.state_dict(),
"dis_optimizer": dis_optimizer.state_dict(),
"best_fid": best_fid,
"path_helper": args.path_helper,
},
is_best,
args.path_helper["ckpt_path"],
)
del avg_gen_net
"""
import argparse
import matplotlib.pyplot as plt
from functions import load_params, norm
def parse_arguments():
parser = argparse.ArgumentParser(description='Plots chosen terrains parameters and makes a simple analysis.')
parser.add_argument('-t', '--terrains', type=int, default=range(1, 16), nargs='+', choices=range(1, 16),
help='Terrains to be plotted (integers)')
return parser.parse_args()
if __name__ == '__main__':
terrains, qualities, ranges, env = load_params('terrain_types', 'terrain_qualities', 'qualities_ranges', 'env')
args = parse_arguments()
terrains_to_use = [terrains[str(i)] for i in sorted(args.terrains)]
''' Terrains Parameters '''
plt.matshow([[norm(env[terrain][quality], ranges[quality][1]) for terrain in terrains_to_use] for quality in qualities],
vmin=0.0, vmax=1.0)
plt.xticks(range(len(terrains_to_use)), terrains_to_use, rotation=45)
plt.yticks(range(len(qualities)), qualities)
plt.colorbar()
plt.suptitle('Chosen Terrains Parameters')
for t_i, terrain in enumerate(terrains_to_use):
for q_i, quality in enumerate(qualities):
plt.text(t_i, q_i, norm(env[terrain][quality], ranges[quality][1]), va='center', ha='center')
plt.savefig('../../results/png/terrains_parameters.png', bbox_inches='tight')
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
'''
Run our simulation of the system.
'''
from functions import load_params
from node import Node
from service_provider import ServiceProvider
from adversaries import load_adversary
from observer import Observer
if __name__ == '__main__':
params = load_params()
T = [[] for _ in params['percent_adv']]
for p, p_adv in enumerate(params['percent_adv']):
print(f"Running sim with {p_adv:%} adversaries...")
nodes = []
for i in range(params['number_nodes']):
if i < p_adv * params['number_nodes']:
nodes.append(load_adversary(params['adv_type'])(i, params))
else:
nodes.append(Node(i, params))
observer = Observer(params['number_nodes'] + 1, params)
nodes.append(observer)
sp = ServiceProvider()
for e in range(params['epochs']):
for n in nodes[:-1]:
n.transact(nodes, sp)
for i, structure in enumerate(kitt_mnist['structure_mean']):
kitt_mnist['structure_mean'][i] = [int(nn) for nn in structure.tolist()]
print '\n ## Loaded mnist pruned nets ('+str(n_obs)+' obs) ## --------------------'
print '@ Pruning steps:\t', max_len
print '@ Mean n synapses:\t', kitt_mnist['n_syn_mean'][-1]
print '@ Mean structure:\t', kitt_mnist['structure_mean'][-1]
print '@ Mean accuracy:\t', kitt_mnist['acc_mean'][-1]
print '----------------------------------------------------'
return kitt_mnist
if __name__ == '__main__':
sensors_ranges, terrain_types, all_sensors = load_params('sensors_ranges', 'terrain_types', 'sensors')
# loading examples
terrains = [terrain_types[str(i)] for i in [6, 8, 15]]
data = read_data(noises=['no_noise'], terrains=terrains, sensors=all_sensors, n_samples=10)
samples = dict()
for terrain in terrains:
samples[terrain] = [[] for i in range(len(data['no_noise'][terrain][all_sensors[0]]))]
for sensor in all_sensors:
for i_sample, sample_terrain in enumerate(data['no_noise'][terrain][sensor]):
samples[terrain][i_sample] += prepare_signal(signal=sample_terrain[10:40 + 10], sen=sensor)
nets = load_amter(na='nn')
#nets = load_mnist()
net = nets['net'][-1]
structure = net[0]
