default=1000000,
help="Number of training iterations")
parser.add_argument('--iter_save',
type=int,
default=10000,
help="Save model every n iterations")
parser.add_argument('--run',
type=int,
default=0,
help="Run ID. In case you want to run replicates")
args = parser.parse_args()
layout = [('model={:s}', 'fsvae'), ('run={:04d}', args.run)]
model_name = '_'.join([t.format(v) for (t, v) in layout])
pprint(vars(args))
print('Model name:', model_name)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
train_loader, labeled_subset, test_set = ut.get_svhn_data(device)
fsvae = FSVAE(name=model_name).to(device)
writer = ut.prepare_writer(model_name, overwrite_existing=True)
train(model=fsvae,
train_loader=train_loader,
labeled_subset=labeled_subset,
device=device,
y_status='fullsup',
tqdm=tqdm.tqdm,
writer=writer,
iter_max=args.iter_max,
iter_save=args.iter_save)
def run(args, verbose=False):
layout = [
('{:s}', "vae2"),
('{:s}', args.model),
# ('x{:02d}', 24 if args.hourly==1 else 96),
# ('z{:02d}', args.z),
('k{:02d}', args.k),
('iw{:02d}', args.iw),
('vp{:02d}', args.var_pen),
('lr{:.4f}', args.lr),
('epo{:03d}', args.num_epochs),
('run{:02d}', args.run)
]
model_name = 'car' + '_'.join([t.format(v) for (t, v) in layout])
if verbose: pprint(vars(args))
print('Model name:', model_name)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# cloud
# root_dir = "../data/data15_final"
# Oskar
root_dir = "../data/CS236/data60/split" if (
args.hourly == 1) else "../data/CS236/data15_final"
# load train loader anyways - to get correct shift_scale values.
train_loader = torch.utils.data.DataLoader(
LoadDataset2(root_dir=root_dir,
mode='train',
shift_scale=None,
filter_ev=False,
log_car=(args.log_ev == 1),
smooth=args.smooth),
batch_size=args.batch,
shuffle=True,
)
shift_scale = train_loader.dataset.shift_scale
# load use-model
use_model = run_vae2.main({
"mode": 'load',
"model": 'ff-s-dec', # hardcode
"lr": 0.01, # hardcode
"k": 1, # hardcode
"iw": 0, # hardcode
"num_epochs": 20, # hardcode
"var_pen": 1, # hardcode
"run": 1, # hardcode
})
if args.k > 1:
print('36')
model = GMVAE2CAR(
nn=args.model,
name=model_name,
z_dim=args.z,
x_dim=24 if args.hourly == 1 else 96,
c_dim=use_model.z_dim,
warmup=(args.warmup == 1),
var_pen=args.var_pen,
use_model=use_model,
k=args.k,
y_dim=train_loader.dataset.dim_meta,
).to(device)
else:
model = VAE2CAR(
nn=args.model,
name=model_name,
z_dim=args.z,
x_dim=24 if args.hourly == 1 else 96,
c_dim=use_model.z_dim,
warmup=(args.warmup == 1),
var_pen=args.var_pen,
use_model=use_model,
y_dim=train_loader.dataset.dim_meta,
).to(device)
if args.mode == 'train':
split_set = LoadDataset2(
root_dir=root_dir,
mode='val',
shift_scale=shift_scale,
filter_ev=False,
log_car=(args.log_ev == 1),
smooth=None,
)
val_set = {
"x": torch.FloatTensor(split_set.car).to(device),
"y": torch.FloatTensor(split_set.meta).to(device),
"c": torch.FloatTensor(split_set.other).to(device),
}
_ = ut.prepare_writer(model_name, overwrite_existing=True)
# make sure not to train the first VAE
if not (args.finetune == 1):
for p in model.use_model.parameters():
p.requires_grad = False
train2(
model=model,
train_loader=train_loader,
val_set=val_set,
tqdm=tqdm.tqdm,
lr=args.lr,
lr_gamma=args.lr_gamma,
lr_milestone_every=args.lr_every,
iw=args.iw,
num_epochs=args.num_epochs,
is_car_model=True,
)
else:
ut.load_model_by_name(model, global_step=args.num_epochs)
if args.mode in ['val', 'test']:
model.set_to_eval()
split_set = LoadDataset2(
root_dir=root_dir,
mode=args.mode,
shift_scale=shift_scale,
filter_ev=False,
log_car=(args.log_ev == 1),
smooth=None,
)
val_set = {
"x": torch.FloatTensor(split_set.car).to(device),
"y": torch.FloatTensor(split_set.meta).to(device),
"c": torch.FloatTensor(split_set.other).to(device),
}
summaries = OrderedDict({
'epoch': args.num_epochs,
'loss': 0,
'kl_z': 0,
'rec_mse': 0,
'rec_var': 0,
'loss_type': 0,
'lr': args.lr,
'var_pen': model.var_pen,
})
ut.save_latent(model, val_set, mode=args.mode, is_car_model=True)
ut.evaluate_lower_bound2(model,
val_set,
run_iwae=True,
mode=args.mode,
repeats=10,
summaries=copy.deepcopy(summaries))
if args.mode == 'plot':
make_image_load(model, shift_scale["car"], (args.log_ev == 1))
# make_image_load_day(model, shift_scale["car"], (args.log_ev==1))
make_image_load_z(model, shift_scale["car"], (args.log_ev == 1))
make_image_load_z_use(model, shift_scale["car"], (args.log_ev == 1))
if args.mode == 'load':
if verbose: print(model)
return model
def run(args, verbose=False):
layout = [
('{:s}', "vae2"),
('{:s}', args.model),
# ('x{:02d}', 24 if args.hourly==1 else 96),
# ('z{:02d}', args.z),
('k{:02d}', args.k),
('iw{:02d}', args.iw),
('vp{:02d}', args.var_pen),
('lr{:.4f}', args.lr),
('epo{:03d}', args.num_epochs),
('run{:02d}', args.run),
]
model_name = '_'.join([t.format(v) for (t, v) in layout])
if verbose: pprint(vars(args))
print('Model name:', model_name)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# cloud
# root_dir = "../data/data15_final"
# Oskar
root_dir = "../data/CS236/data60/split" if (
args.hourly == 1) else "../data/CS236/data15_final"
# Will
#root_dir = '/Users/willlauer/Desktop/latent_load_gen/data/split'
# load train loader anyways - to get correct shift_scale values.
train_loader = torch.utils.data.DataLoader(
LoadDataset2(root_dir=root_dir,
mode='train',
shift_scale=None,
filter_ev=False,
smooth=args.smooth),
batch_size=args.batch,
shuffle=True,
)
shift_scale = train_loader.dataset.shift_scale
if args.k > 1:
model = GMVAE2(
nn=args.model,
z_dim=args.z,
name=model_name,
x_dim=24 if args.hourly == 1 else 96,
warmup=(args.warmup == 1),
var_pen=args.var_pen,
k=args.k,
y_dim=train_loader.dataset.dim_meta,
).to(device)
else:
model = VAE2(nn=args.model,
z_dim=args.z,
name=model_name,
x_dim=24 if args.hourly == 1 else 96,
y_dim=train_loader.dataset.dim_meta,
warmup=(args.warmup == 1),
var_pen=args.var_pen).to(device)
if args.mode == 'train':
split_set = LoadDataset2(
root_dir=root_dir,
mode='val',
shift_scale=shift_scale,
filter_ev=False,
smooth=None,
)
val_set = {
"x": torch.FloatTensor(split_set.other).to(device),
"y": torch.FloatTensor(split_set.meta).to(device),
"c": None,
}
# maybe in future use Tensorboard?
_ = ut.prepare_writer(model_name, overwrite_existing=True)
train2(
model=model,
train_loader=train_loader,
val_set=val_set,
tqdm=tqdm.tqdm,
# writer=writer,
lr=args.lr,
lr_gamma=args.lr_gamma,
lr_milestone_every=args.lr_every,
iw=args.iw,
num_epochs=args.num_epochs)
else:
ut.load_model_by_name(model, global_step=args.num_epochs)
if args.mode in ['val', 'test']:
model.set_to_eval()
split_set = LoadDataset2(
root_dir=root_dir,
mode=args.mode,
shift_scale=shift_scale,
filter_ev=False,
smooth=None,
)
val_set = {
"x": torch.FloatTensor(split_set.other).to(device),
"y": torch.FloatTensor(split_set.meta).to(device),
"c": None,
}
summaries = OrderedDict({
'epoch': args.num_epochs,
'loss': 0,
'kl_z': 0,
'rec_mse': 0,
'rec_var': 0,
'loss_type': 0,
'lr': args.lr,
'var_pen': model.var_pen,
})
ut.save_latent(model, val_set, mode=args.mode, is_car_model=False)
ut.evaluate_lower_bound2(model,
val_set,
run_iwae=True,
mode=args.mode,
repeats=10,
summaries=summaries)
if args.mode == 'plot':
# print(shift_scale["other"])
# print(shift_scale)
make_image_load(model, shift_scale["other"], (args.log_ev == 1))
# make_image_load_day(model, shift_scale["other"], (args.log_ev==1))
make_image_load_z(model, shift_scale["other"], (args.log_ev == 1))
if args.mode == 'load':
if verbose: print(model)
return model
help="Save model every n iterations")
parser.add_argument('--run',
type=int,
default=0,
help="Run ID. In case you want to run replicates")
parser.add_argument('--sample', type=int, default=0, help="Sample")
args = parser.parse_args()
layout = [('model={:s}', 'fsvae'), ('run={:04d}', args.run)]
model_name = '_'.join([t.format(v) for (t, v) in layout])
pprint(vars(args))
print('Model name:', model_name)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
device = torch.device('cpu')
fsvae = FSVAE(name=model_name).to(device)
writer = ut.prepare_writer(model_name, overwrite_existing=args.sample == 0)
if args.sample:
with torch.no_grad():
ut.load_model_by_name(fsvae, global_step=args.iter_max)
images = fsvae.sample_clipped_x(20)
save_image(images.view(200, 3, 32, 32), 'fsvae_sample.png', nrow=20)
else:
train_loader, labeled_subset, test_set = ut.get_svhn_data(device)
train(model=fsvae,
train_loader=train_loader,
labeled_subset=labeled_subset,
device=device,
y_status='fullsup',
tqdm=tqdm.tqdm,
writer=writer,
