def reset_gen():
if args.model in ['iagan_began_cs']:
gen = Generator128(64)
gen = load_trained_net(
gen,
('./checkpoints/celeba_began.withskips.bs32.cosine.min=0.25'
'.n_cuts=0/gen_ckpt.49.pt'))
gen = gen.eval().to(DEVICE)
img_size = 128
elif args.model in ['iagan_dcgan_cs']:
gen = dcgan_generator()
t = torch.load(('./dcgan_checkpoints/netG.epoch_24.n_cuts_0.bs_64'
'.b1_0.5.lr_0.0002.pt'))
gen.load_state_dict(t)
gen = gen.eval().to(DEVICE)
img_size = 64
elif args.model in ['iagan_vanilla_vae_cs']:
gen = VAE()
t = torch.load('./vae_checkpoints/vae_bs=128_beta=1.0/epoch_19.pt')
gen.load_state_dict(t)
gen = gen.eval().to(DEVICE)
gen = gen.decoder
img_size = 128
else:
raise NotImplementedError()
return gen, img_size
def main():
parser = _build_parser()
args = parser.parse_args()
logging.basicConfig(format="%(levelname)s: %(message)s",
level=logging.DEBUG)
device = torch.device('cpu')
if torch.cuda.is_available():
device = torch.device('cuda')
model = None
if args.model == 'vae':
model = VAE().double().to(device)
elif args.model == 'hm':
model = HM().double().to(device)
else:
logging.critical('model unimplemented: %s' % args.model)
return
if not args.out.exists():
args.out.mkdir(parents=True)
_, test_ds = build_datasets(args.im_path, train_test_split=1)
ckpt = torch.load(args.save_path, map_location=device)
model.load_state_dict(ckpt['model_state_dict'])
model.eval()
_sample_images(model, args.batch, args.samples, test_ds, args.out)
def __init__(self, binary_features, continuous_features, z_dim, hidden_dim,
hidden_layers, optimizer, activation, cuda):
pyro.clear_param_store()
vae = VAE(binary_features, continuous_features, z_dim, hidden_dim,
hidden_layers, activation, cuda)
vae = vae.double()
self.vae = vae
self.svi = SVI(vae.model, vae.guide, optimizer, loss=Trace_ELBO())
self.cuda = cuda
self.train_stats = Statistics()
self.test_stats = Statistics()
def get_model(name):
para = {'struct': [784,128,100],
'hidden_func': ['r','a'],
'hidden_func2': ['r','s'],
'n_category': 10,
'dropout': 0.0,
'task': 'gnr',
'img_size': [28,28],
'flatten': True}
if name == 'vae':
return VAE(**para)
elif name == 'mmd':
return MMDGM_VAE(**para)
else:
raise Exception("Enter a correct model name!")
def main():
(train_X, train_Y), (test_X,
test_Y) = load_fold(["data/input.fold.train1.csv"],
["data/input.fold.test1.csv"])
print(train_X.shape)
vae = VAE()
# Train VAE
vae.fit(train_X)
# Sampling 1 data point from each data point in train set
samples = vae.generate(X=train_X)
return
def main():
parser = _build_parser()
args = parser.parse_args()
logging.basicConfig(format="%(levelname)s: %(message)s", level=logging.DEBUG)
device = torch.device('cpu')
if torch.cuda.is_available():
device = torch.device('cuda')
model = None
if args.model == 'vae':
model = VAE().double().to(device)
else:
logging.critical('model unimplemented: %s' % args.model)
return
if not args.out.parent.exists():
args.out.parent.mkdir()
_, test_ds = build_datasets(args.im_path, train_test_split=1)
ckpt = torch.load(args.save_path, map_location=device)
model.load_state_dict(ckpt['model_state_dict'])
model.eval()
with torch.no_grad():
samps = model.sample(args.samples).reshape(-1, *IM_DIMS, 3)
loader = DataLoader(test_ds,
batch_size=args.batch,
num_workers=args.workers,
pin_memory=torch.cuda.is_available())
record = _init_record(samps)
with tqdm(total=TOTAL_IMAGES) as pbar:
for chunk in loader:
_update_winner(chunk.reshape(-1, *IM_DIMS, 3), record, pbar)
np.save(args.out, record['pair'])
print('final distances:', record['distance'])
def reset_gen(model):
if model == 'began':
gen = Generator128(64)
gen = load_trained_net(
gen,
('./checkpoints/celeba_began.withskips.bs32.cosine.min=0.25'
'.n_cuts=0/gen_ckpt.49.pt'))
gen = gen.eval().to(DEVICE)
img_size = 128
elif model == 'vae':
gen = VAE()
t = torch.load('./vae_checkpoints/vae_bs=128_beta=1.0/epoch_19.pt')
gen.load_state_dict(t)
gen = gen.eval().to(DEVICE)
gen = gen.decoder
img_size = 128
elif model == 'dcgan':
gen = dcgan_generator()
t = torch.load(('./dcgan_checkpoints/netG.epoch_24.n_cuts_0.bs_64'
'.b1_0.5.lr_0.0002.pt'))
gen.load_state_dict(t)
gen = gen.eval().to(DEVICE)
img_size = 64
return gen, img_size
def generator_samples(model):
if model == 'began':
g = Generator128(64).to('cuda:0')
g = load_trained_net(
g, ('./checkpoints/celeba_began.withskips.bs32.cosine.min=0.25'
'.n_cuts=0/gen_ckpt.49.pt'))
elif model == 'vae':
g = VAE().to('cuda:0')
g.load_state_dict(
torch.load('./vae_checkpoints/vae_bs=128_beta=1.0/epoch_19.pt'))
g = g.decoder
elif model == 'biggan':
g = BigGanSkip().to('cuda:0')
elif model == 'dcgan':
g = dcgan_generator().to('cuda:0')
g.load_state_dict(
torch.load(('./dcgan_checkpoints/netG.epoch_24.n_cuts_0.bs_64'
'.b1_0.5.lr_0.0002.pt')))
else:
raise NotImplementedError
nseed = 10
n_cuts_list = [0, 1, 2, 3, 4, 5]
fig, ax = plt.subplots(len(n_cuts_list),
nseed,
figsize=(10, len(n_cuts_list)))
for row, n_cuts in enumerate(n_cuts_list):
input_shapes = g.input_shapes[n_cuts]
z1_shape = input_shapes[0]
z2_shape = input_shapes[1]
for col in range(nseed):
torch.manual_seed(col)
np.random.seed(col)
if n_cuts == 0 and model == 'biggan':
class_vector = torch.tensor(
949, dtype=torch.long).to('cuda:0').unsqueeze(
0) # 949 = strawberry
embed = g.biggan.embeddings(
torch.nn.functional.one_hot(
class_vector, num_classes=1000).to(torch.float))
cond_vector = torch.cat(
(torch.randn(1, 128).to('cuda:0'), embed), dim=1)
img = orig_biggan_forward(
g.biggan.generator, cond_vector,
truncation=1.0).detach().cpu().squeeze(
0).numpy().transpose([1, 2, 0])
elif n_cuts > 0 and model == 'biggan':
z1 = torch.randn(1, *z1_shape).to('cuda:0')
class_vector = torch.tensor(
949, dtype=torch.long).to('cuda:0').unsqueeze(
0) # 949 = strawberry
embed = g.biggan.embeddings(
torch.nn.functional.one_hot(
class_vector, num_classes=1000).to(torch.float))
cond_vector = torch.cat(
(torch.randn(1, 128).to('cuda:0'), embed), dim=1)
z2 = cond_vector
img = g(
z1, z2, truncation=1.0,
n_cuts=n_cuts).detach().cpu().squeeze(0).numpy().transpose(
[1, 2, 0])
else:
z1 = torch.randn(1, *z1_shape).to('cuda:0')
if len(z2_shape) == 0:
z2 = None
else:
z2 = torch.randn(1, *z2_shape).to('cuda:0')
img = g(
z1, z2,
n_cuts=n_cuts).detach().cpu().squeeze(0).numpy().transpose(
[1, 2, 0])
if g.rescale:
img = (img + 1) / 2
ax[row, col].imshow(np.clip(img, 0, 1), aspect='auto')
ax[row, col].set_xticks([])
ax[row, col].set_yticks([])
ax[row, col].set_frame_on(False)
if col == 0:
ax[row, col].set_ylabel(f'{n_cuts}')
fig.subplots_adjust(0, 0, 1, 1, 0, 0)
os.makedirs('./figures/generator_samples', exist_ok=True)
plt.savefig((f'./figures/generator_samples/'
f'model={model}.pdf'),
dpi=300,
bbox_inches='tight')
def mgan_images(args):
if args.set_seed:
torch.manual_seed(0)
np.random.seed(0)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
os.makedirs(BASE_DIR, exist_ok=True)
if args.model in ['mgan_began_cs']:
gen = Generator128(64)
gen = load_trained_net(
gen, ('./checkpoints/celeba_began.withskips.bs32.cosine.min=0.25'
'.n_cuts=0/gen_ckpt.49.pt'))
gen = gen.eval().to(DEVICE)
img_size = 128
elif args.model in ['mgan_vanilla_vae_cs']:
gen = VAE()
t = torch.load('./vae_checkpoints/vae_bs=128_beta=1.0/epoch_19.pt')
gen.load_state_dict(t)
gen = gen.eval().to(DEVICE)
gen = gen.decoder
img_size = 128
elif args.model in ['mgan_dcgan_cs']:
gen = dcgan_generator()
t = torch.load(('./dcgan_checkpoints/netG.epoch_24.n_cuts_0.bs_64'
'.b1_0.5.lr_0.0002.pt'))
gen.load_state_dict(t)
gen = gen.eval().to(DEVICE)
img_size = 64
else:
raise NotImplementedError()
img_shape = (3, img_size, img_size)
metadata = recovery_settings[args.model]
n_cuts_list = metadata['n_cuts_list']
del (metadata['n_cuts_list'])
z_init_mode_list = metadata['z_init_mode']
limit_list = metadata['limit']
assert len(z_init_mode_list) == len(limit_list)
del (metadata['z_init_mode'])
del (metadata['limit'])
forwards = forward_models[args.model]
data_split = Path(args.img_dir).name
for img_name in tqdm(sorted(os.listdir(args.img_dir)),
desc='Images',
leave=True,
disable=args.disable_tqdm):
# Load image and get filename without extension
orig_img = load_target_image(os.path.join(args.img_dir, img_name),
img_size).to(DEVICE)
img_basename, _ = os.path.splitext(img_name)
for n_cuts in tqdm(n_cuts_list,
desc='N_cuts',
leave=False,
disable=args.disable_tqdm):
metadata['n_cuts'] = n_cuts
for i, (f, f_args_list) in enumerate(
tqdm(forwards.items(),
desc='Forwards',
leave=False,
disable=args.disable_tqdm)):
for f_args in tqdm(f_args_list,
desc=f'{f} Args',
leave=False,
disable=args.disable_tqdm):
f_args['img_shape'] = img_shape
forward_model = get_forward_model(f, **f_args)
for z_init_mode, limit in zip(
tqdm(z_init_mode_list,
desc='z_init_mode',
leave=False), limit_list):
metadata['z_init_mode'] = z_init_mode
metadata['limit'] = limit
# Before doing recovery, check if results already exist
# and possibly skip
recovered_name = 'recovered.pt'
results_folder = get_results_folder(
image_name=img_basename,
model=args.model,
n_cuts=n_cuts,
split=data_split,
forward_model=forward_model,
recovery_params=dict_to_str(metadata),
base_dir=BASE_DIR)
os.makedirs(results_folder, exist_ok=True)
recovered_path = results_folder / recovered_name
if os.path.exists(
recovered_path) and not args.overwrite:
print(
f'{recovered_path} already exists, skipping...'
)
continue
if args.run_name is not None:
current_run_name = (
f'{img_basename}.{forward_model}'
f'.{dict_to_str(metadata)}'
f'.{args.run_name}')
else:
current_run_name = None
recovered_img, distorted_img, _ = mgan_recover(
orig_img, gen, n_cuts, forward_model,
metadata['optimizer'], z_init_mode, limit,
metadata['z_lr'], metadata['n_steps'],
metadata['z_number'], metadata['restarts'],
args.run_dir, current_run_name, args.disable_tqdm)
# Make images folder
img_folder = get_images_folder(split=data_split,
image_name=img_basename,
img_size=img_size,
base_dir=BASE_DIR)
os.makedirs(img_folder, exist_ok=True)
# Save original image if needed
original_img_path = img_folder / 'original.pt'
if not os.path.exists(original_img_path):
torch.save(orig_img, original_img_path)
# Save distorted image if needed
if forward_model.viewable:
distorted_img_path = img_folder / f'{forward_model}.pt'
if not os.path.exists(distorted_img_path):
torch.save(distorted_img, distorted_img_path)
# Save recovered image and metadata
torch.save(recovered_img, recovered_path)
pickle.dump(
metadata,
open(results_folder / 'metadata.pkl', 'wb'))
p = psnr(recovered_img, orig_img)
pickle.dump(p, open(results_folder / 'psnr.pkl', 'wb'))
from dataset.cfd import build_datasets
# from dataset.utk import build_datasets
from model.vae import VAE
# from model.vae_gm import GMVAE
from util import lda_analysis
data_path = Path('../data/cfd')
# data_path = Path('../data/utk')
model_path = Path('../save/vae/vae_jan19_final.pt')
save_path = Path('../save/vae/cfd/latent')
# <codecell>
if not save_path.exists():
save_path.mkdir(parents=True)
model = VAE()
ckpt = torch.load(model_path, map_location=torch.device('cpu'))
model.load_state_dict(ckpt['model_state_dict'])
model.eval()
_, test_ds = build_datasets(data_path, train_test_split=1)
# test_ds = [test_ds[i] for i in range(10)]
# <codecell>
test_len = len(test_ds)
ldims = model.latent_dims
mu_points = np.zeros((test_len, ldims))
var_points = np.zeros((test_len, ldims))
feats = []
default=0.4,
metavar='DR',
help='aux loss coef (default: 0.4)')
parser.add_argument('--use-trained',
type=bool,
default=False,
metavar='UT',
help='load pretrained model (default: False)')
args = parser.parse_args()
batch_loader = BatchLoader()
parameters = Parameters(batch_loader.vocab_size)
vae = VAE(parameters.vocab_size, parameters.embed_size,
parameters.latent_size, parameters.decoder_rnn_size,
parameters.decoder_rnn_num_layers)
if args.use_trained:
vae.load_state_dict(t.load('trained_VAE'))
if args.use_cuda:
vae = vae.cuda()
optimizer = Adam(vae.parameters(), args.learning_rate)
for iteration in range(args.num_iterations):
'''Train step'''
input, decoder_input, target = batch_loader.next_batch(
args.batch_size, 'train', args.use_cuda)
target = target.view(-1)
logits, aux_logits, kld = vae(args.dropout, input, decoder_input)
def __init__(self, config):
super(Cycle_GAN, self).__init__()
self.config = config
self.logger = get_logger("CycleGAN")
self.output_names = [
'real_A', 'fake_B', 'rec_A', 'real_B', 'fake_A', 'rec_B'
]
self.sigma = self.config['variational']['sigma']
assert "sketch" in self.config["model_type"] and "face" in self.config[
"model_type"], "This CycleGAN iterator only works with model_type:'sketch2face'"
assert config[
"iterator"] == "iterator.cycle_gan.CycleGAN", "This CycleGAN model only works with with the Cycle_GAN iterator."
latent_dim = self.config["latent_dim"]
min_channels = self.config['conv']['n_channel_start']
max_channels = self.config['conv']['n_channel_max']
sketch_shape = [32, 1]
face_shape = [self.config['data']['transform']['resolution'], 3]
sigma = self.config['variational']['sigma']
num_extra_conv_sketch = self.config['conv']['sketch_extra_conv']
num_extra_conv_face = self.config['conv']['face_extra_conv']
BlockActivation = nn.ReLU()
FinalActivation = nn.Tanh()
batch_norm_enc = batch_norm_dec = self.config['batch_norm']
drop_rate_enc = self.config['dropout']['enc_rate']
drop_rate_dec = self.config['dropout']['dec_rate']
bias_enc = self.config['bias']['enc']
bias_dec = self.config['bias']['dec']
num_latent_layer = self.config['variational'][
'num_latent_layer'] if 'num_latent_layer' in self.config[
'variational'] else 0
# load the right networks of the model
self.netG_A = VAE(latent_dim,
min_channels,
max_channels,
*sketch_shape,
*face_shape,
sigma,
num_extra_conv_sketch,
num_extra_conv_face,
BlockActivation,
FinalActivation,
batch_norm_enc,
batch_norm_dec,
drop_rate_enc,
drop_rate_dec,
bias_enc,
bias_dec,
num_latent_layer=num_latent_layer)
self.netD_A = Discriminator_sketch()
self.netG_B = VAE(latent_dim,
min_channels,
max_channels,
*face_shape,
*sketch_shape,
sigma,
num_extra_conv_face,
num_extra_conv_sketch,
BlockActivation,
FinalActivation,
batch_norm_enc,
batch_norm_dec,
drop_rate_enc,
drop_rate_dec,
bias_enc,
bias_dec,
num_latent_layer=num_latent_layer)
self.netD_B = Discriminator_face()
def __init__(self, config):
super(Cycle_WGAN, self).__init__()
assert config[
"iterator"] == "iterator.cycle_wgan.Cycle_WGAN", "This model only works with the iterator: 'iterator.cycle_wgan.Cycle_WGAN"
assert "sketch" in config["model_type"] and "face" in config[
"model_type"], "This model only works with model_type: 'sketch2face'"
if "debug_log_level" in config and config["debug_log_level"]:
LogSingleton.set_log_level("debug")
# get logger and config
self.logger = get_logger("CycleWGAN")
self.config = config
set_random_state(self.config)
self.logger.info("WGAN_GradientPenalty init model.")
self.output_names = [
'real_A', 'fake_B', 'rec_A', 'real_B', 'fake_A', 'rec_B'
]
self.sigma = self.config['variational']['sigma']
self.num_latent_layer = self.config['variational'][
'num_latent_layer'] if "variational" in self.config and "num_latent_layer" in self.config[
"variational"] else 0
latent_dim = self.config["latent_dim"]
min_channels = self.config['conv']['n_channel_start']
max_channels = self.config['conv']['n_channel_max']
sketch_shape = [32, 1]
face_shape = [self.config['data']['transform']['resolution'], 3]
sigma = self.config['variational']['sigma']
num_extra_conv_sketch = self.config['conv']['sketch_extra_conv']
num_extra_conv_face = self.config['conv']['face_extra_conv']
block_activation = nn.ReLU()
final_activation = nn.Tanh()
batch_norm_enc = batch_norm_dec = self.config['batch_norm']
drop_rate_enc = self.config['dropout'][
'enc_rate'] if "dropout" in self.config and "enc_rate" in self.config[
"dropout"] else 0
drop_rate_dec = self.config['dropout'][
'dec_rate'] if "dropout" in self.config and "dec_rate" in self.config[
"dropout"] else 0
bias_enc = self.config['bias'][
'enc'] if "bias" in self.config and "enc" in self.config[
"bias"] else True
bias_dec = self.config['bias'][
'dec'] if "bias" in self.config and "dec" in self.config[
"bias"] else True
num_latent_layer = self.config['variational'][
'num_latent_layer'] if "variational" in self.config and "num_latent_layer" in self.config[
"variational"] else 0
## cycle A ##
self.netG_A = VAE(latent_dim=latent_dim,
min_channels=min_channels,
max_channels=max_channels,
in_size=sketch_shape[0],
in_channels=sketch_shape[1],
out_size=face_shape[0],
out_channels=face_shape[1],
sigma=sigma,
num_extra_conv_enc=num_extra_conv_sketch,
num_extra_conv_dec=num_extra_conv_face,
block_activation=block_activation,
final_activation=final_activation,
batch_norm_enc=batch_norm_enc,
batch_norm_dec=batch_norm_dec,
drop_rate_enc=drop_rate_enc,
drop_rate_dec=drop_rate_dec,
bias_enc=bias_enc,
bias_dec=bias_dec,
same_max_channels=False,
num_latent_layer=num_latent_layer)
self.netD_A = WGAN_Discriminator_sketch()
## cycle B ##
self.netG_B = VAE(latent_dim=latent_dim,
min_channels=min_channels,
max_channels=max_channels,
in_size=face_shape[0],
in_channels=face_shape[1],
out_size=sketch_shape[0],
out_channels=sketch_shape[1],
sigma=sigma,
num_extra_conv_enc=num_extra_conv_sketch,
num_extra_conv_dec=num_extra_conv_face,
block_activation=block_activation,
final_activation=final_activation,
batch_norm_enc=batch_norm_enc,
batch_norm_dec=batch_norm_dec,
drop_rate_enc=drop_rate_enc,
drop_rate_dec=drop_rate_dec,
bias_enc=bias_enc,
bias_dec=bias_dec,
same_max_channels=False,
num_latent_layer=num_latent_layer)
self.netD_B = WGAN_Discriminator_face(input_resolution=face_shape[0])
batch_size=2,
shuffle=True,
num_workers=2,
collate_fn=collate_fn)
#
# Model
#
device = torch.device('cuda')
epochs = 50
hidden_size = 1024
latent_size = 512
model = VAE(h_dim=hidden_size, z_dim=latent_size)
model = nn.DataParallel(model)
model = model.to(device)
criterion = vae_loss_function
optimizer = torch.optim.Adam(model.parameters(), lr=0.0001)
#
# Training
#
model.train()
for epoch in range(epochs):
loss, bce, kld = 0, 0, 0
max_batches = len(unlabeled_trainloader)
args = parser.parse_args()
args.device = torch.device(
"cuda" if args.gpu and torch.cuda.is_available() else "cpu")
with open('config') as file:
params = yaml.load(file, Loader=yaml.Loader)
layer = '_'.join([str(l) for l in args.layer])
path = '{}/{}/{}'.format(params['dataset_dir'], params['dataset'],
params['feature'])
print('feature data path: ' + path)
X = io.mmread(path).A.astype('float32')
args.n, args.d = X.shape
# X = normalize(X, norm='l2', axis=0)
vae = VAE(args)
if args.load:
path = '{}/model/vae_{}_{}'.format(params['dataset_dir'],
params['dataset'], layer)
vae.load_state_dict(torch.load(path))
vae.to(args.device)
# psm = PSM(torch.from_numpy(R.astype('float32')).to(args.device), args).to(args.device)
loader = DataLoader(np.arange(args.n), batch_size=1, shuffle=True)
optimizer = optim.Adam(vae.parameters(), lr=args.lr)
evaluator = Evaluator({'recall', 'dcg_cut'})
# variational()
# maximum()
def main():
parser = _build_parser()
args = parser.parse_args()
logging.basicConfig(format="%(levelname)s: %(message)s",
level=logging.DEBUG)
device = torch.device('cpu')
if torch.cuda.is_available():
device = torch.device('cuda')
model = None
save_path = args.save
if args.model == 'vae':
model = VAE(args.dim)
logging.info('training VAE with dims: {}'.format(args.dim))
elif args.model == 'ae':
model = AE(args.dim)
logging.info('training AE with dims: {}'.format(args.dim))
elif args.model == 'hm':
model = HM(args.color)
elif args.model == 'gmvae':
model = GMVAE()
else:
logging.critical('model unimplemented: %s' % args.model)
return
if not save_path.exists():
save_path.mkdir(parents=True)
model = model.float()
model.to(device)
optimizer = optim.Adam(model.parameters())
train_ds, test_ds = build_datasets(args.path)
losses = []
for e in range(args.epochs):
logging.info('epoch: %d of %d' % (e + 1, args.epochs))
loader = DataLoader(train_ds,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=torch.cuda.is_available())
total_batches = len(train_ds) // args.batch_size
log_every = total_batches // 50 + 1
save_every = 1 # hardcoded for now
for i, x in enumerate(loader):
x = x.to(device)
optimizer.zero_grad()
output = model(x)
total_loss = model.loss_function(output)
if type(total_loss) is dict: # TODO: generalize loss handling
total_loss = total_loss['loss']
total_loss.backward()
optimizer.step()
if i % log_every == 0:
model.eval()
loss = _eval(model, test_ds, device)
model.train()
logging.info('[batch %d/%d] ' % (i + 1, total_batches) +
model.print_loss(loss))
# TODO: generalize printing
# print_params = (i+1, total_batches, loss['loss'], loss['mse'], loss['kld'])
# logging.info('[batch %d/%d] loss: %f, mse: %f, kld: %f' % print_params)
# print_params = (i+1, total_batches, loss)
# logging.info('[batch %d/%d] loss: %f' % print_params)
losses.append({'iter': i, 'epoch': e, 'loss': loss})
if e % save_every == 0:
torch.save(
{
'epoch': e + 1,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss
}, save_path / ('epoch_%d.pt' % (e + 1)))
model.eval()
loss = _eval(model, test_ds, device)
model.train()
logging.info('final loss: %s' % loss)
losses.append({'iter': 0, 'epoch': e + 1, 'loss': loss})
with open(save_path / 'loss.pk', 'wb') as pkf:
pickle.dump(losses, pkf)
torch.save(
{
'epoch': args.epochs,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss
}, save_path / 'final.pt')
print('done!')
def main():
file_name = 'data/CUB/attributes.txt'
f = open(file_name, 'r')
content = f.readlines()
atts = []
for item in content:
atts.append(item.strip().split(' ')[1])
CUDA = False
if torch.cuda.is_available():
CUDA = True
print('cuda available')
torch.backends.cudnn.benchmark = True
config = config_process(parser.parse_args())
print(config)
# pkl_name='./pkl/{}_{}_{}_{}_task_id_{}_finetune_{}_{}'.format(
# config['dataset'], config['method'], config['softmax_method'],config['arch'],
# config['task_id'], config['finetune'], '.pkl')
#
# with open(pkl_name,'rb') as f:
# feat_dict=pkl.load(f)
with open('pkl/task_0_train.pkl', 'rb') as f:
task_0_train = pkl.load(f)
with open('pkl/task_1_train.pkl', 'rb') as f:
task_1_train = pkl.load(f)
###### task 0:seen training data and unseen test data
examples, labels, class_map = image_load(config['class_file'],
config['image_label'])
###### task 0: seen test data
examples_0, labels_0, class_map_0 = image_load(config['class_file'],
config['test_seen_classes'])
datasets = split_byclass(config, examples, labels,
np.loadtxt(config['attributes_file']), class_map)
datasets_0 = split_byclass(config, examples_0, labels_0,
np.loadtxt(config['attributes_file']),
class_map)
print('load the task 0 train: {} the task 1 as test: {}'.format(
len(datasets[0][0]), len(datasets[0][1])))
print('load task 0 test data {}'.format(len(datasets_0[0][0])))
classes_text_embedding = torch.eye(312, dtype=torch.float32)
test_attr = classes_text_embedding[:, :]
train_attr = F.normalize(datasets[0][3])
# test_attr=F.normalize(datasets[0][4])
best_cfg = config
best_cfg['n_classes'] = datasets[0][3].size(0)
best_cfg['n_train_lbl'] = datasets[0][3].size(0)
best_cfg['n_test_lbl'] = datasets[0][4].size(0)
task_0_train_set = grab_data(best_cfg, task_0_train, datasets[0][2], True)
task_1_train_set = grab_data(best_cfg, task_1_train, datasets[0][2], False)
base_model = models.__dict__[config['arch']](pretrained=True)
if config['arch'].startswith('resnet'):
FE_model = nn.Sequential(*list(base_model.children())[:-1])
else:
print('untested')
raise NotImplementedError
print('load pretrained FE_model')
FE_path = './ckpts/{}_{}_{}_task_id_{}_finetune_{}_{}'.format(
config['dataset'], config['softmax_method'], config['arch'],
config['task_id'], config['finetune'], 'checkpoint.pth')
FE_model.load_state_dict(torch.load(FE_path)['state_dict_FE'])
for name, para in FE_model.named_parameters():
para.requires_grad = False
vae = VAE(encoder_layer_sizes=config['encoder_layer_sizes'],
latent_size=config['latent_size'],
decoder_layer_sizes=config['decoder_layer_sizes'],
num_labels=config['num_labels'])
vae2 = VAE(encoder_layer_sizes=config['encoder_layer_sizes'],
latent_size=config['latent_size'],
decoder_layer_sizes=config['decoder_layer_sizes'],
num_labels=config['num_labels'])
vae_path = './ckpts/{}_{}_{}_task_id_{}_finetune_{}_{}'.format(
config['dataset'], config['method'], config['arch'], config['task_id'],
config['finetune'], 'ckpt.pth')
vae_path = './ckpts/{}_{}_{}_{}_task_id_{}_finetune_{}_{}'.format(
config['dataset'], 'vae', 'softmax_distill', config['arch'], 1,
config['finetune'], 'ckpt.pth')
vae2_path = './ckpts/{}_{}_{}_{}_task_id_{}_finetune_{}_{}'.format(
config['dataset'], 'vae_distill', 'softmax_distill', config['arch'], 1,
config['finetune'], 'ckpt.pth')
vae.load_state_dict(torch.load(vae_path))
vae2.load_state_dict(torch.load(vae2_path))
for name, para in vae.named_parameters():
para.requires_grad = False
for name, para in vae2.named_parameters():
para.requires_grad = False
if CUDA:
FE_model = FE_model.cuda()
vae = vae.cuda()
vae2 = vae2.cuda()
ATTR_NUM = 312
SYN_NUM = config['syn_num']
attr_feat, attr_lbl = generate_syn_feature(ATTR_NUM, vae, test_attr,
SYN_NUM, config)
attr_feat2, attr_lbl2 = generate_syn_feature(ATTR_NUM, vae2, test_attr,
SYN_NUM, config)
with open('attr_tsne_data/attr_vae_time_2_fe_distill.pkl', 'wb') as f:
pkl.dump(attr_feat, f)
with open('attr_tsne_data/attr_vae_time_2_double_distill.pkl', 'wb') as g:
pkl.dump(attr_feat2, g)
colors = cm.rainbow(np.linspace(0, 1, ATTR_NUM))
fig = plt.figure(figsize=(16, 9))
tsne = TSNE(n_components=2)
feat = torch.cat((attr_feat, attr_feat2))
tsne_results = tsne.fit_transform(feat)
color_ind = colors[attr_lbl]
ax = fig.add_subplot(1, 1, 1)
for i in range(ATTR_NUM):
ax.scatter(tsne_results[i * SYN_NUM:(i + 1) * SYN_NUM, 0],
tsne_results[i * SYN_NUM:(i + 1) * SYN_NUM, 1],
label=atts[i],
c=np.tile(colors[i].reshape(1, -1), (SYN_NUM, 1)),
s=20,
marker='X')
result = tsne_results[ATTR_NUM * SYN_NUM:, :]
for j in range(ATTR_NUM):
ax.scatter(result[j * SYN_NUM:(j + 1) * SYN_NUM, 0],
result[j * SYN_NUM:(j + 1) * SYN_NUM, 1],
label=atts[j],
c=np.tile(colors[j].reshape(1, -1), (SYN_NUM, 1)),
s=20,
marker='o')
# ax.scatter(tsne_results[:, 0], tsne_results[:, 1],c=color_ind, s=20, marker='X')
plt.legend()
plt.show()
device = torch.device('cuda:0')
train_loader, img_size = get_dataloader('./data/celeba_preprocessed',
args.batch_size,
n_train=-1,
train=True)
test_loader, _ = get_dataloader('./data/celeba_preprocessed',
args.batch_size,
n_train=-1,
train=False)
save_img_every_n = 20
latent_dim = 512
lr = 2e-4
model = VAE(latent_dim).to(device)
optimizer = optim.AdamW(model.parameters(), lr=lr, betas=(0.9, 0.999))
scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9)
# Reconstruction + KL divergence losses summed over all elements and batch
def loss_function(recon_x, x, mu, logvar):
reconstr = F.mse_loss(recon_x, x,
reduction='none').sum(3).sum(2).sum(1).mean(0)
# see Appendix B from VAE paper:
# Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014
# https://arxiv.org/abs/1312.6114
# 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
KLD = torch.mean(-0.5 *
torch.sum(1 + logvar - mu.pow(2) - logvar.exp(), dim=1),
def test_new(self):
model = VAE()
self.assertNotEqual(model, None)