def __init__(self, opt, logger):
super().__init__(opt, logger)
self.net_N = opt.net_N
self.x_dim = opt.x_dim
self.z_dim = opt.z_dim
self.n_classes = opt.n_classes
self.p_drop = opt.p_drop
self.modality = opt.modality
self.clf_weights = opt.clf_weights
# init networks
self.net_q = Q_net(self.net_N, self.x_dim, self.z_dim, self.p_drop)
self.net_p = P_net(self.net_N, self.x_dim, self.z_dim, self.p_drop)
self.net_c = C_net(self.net_N, self.z_dim, self.n_classes, self.p_drop)
self._nets = [self.net_q, self.net_p, self.net_c]
# optimizers
# self.optimizer and self.optimizer_params have been initialized in netinterface
self.optimizer_q = self.optimizer(self.net_q.parameters(),
lr=opt.lr,
**self.optimizer_params)
self.optimizer_p = self.optimizer(self.net_p.parameters(),
lr=opt.lr,
**self.optimizer_params)
self.optimizer_c = self.optimizer(self.net_c.parameters(),
lr=opt.lr,
**self.optimizer_params)
self._optimizers = [
self.optimizer_q, self.optimizer_p, self.optimizer_c
]
# schedulers
# self.scheduler and self.scheduler_params have been initialized in netinterface
self.scheduler_q = self.scheduler(self.optimizer_q,
**self.scheduler_params)
self.scheduler_p = self.scheduler(self.optimizer_p,
**self.scheduler_params)
self.scheduler_c = self.scheduler(self.optimizer_c,
**self.scheduler_params)
self._schedulers = [
self.scheduler_q, self.scheduler_p, self.scheduler_c
]
# general
self.opt = opt
self._metrics = [
'loss_mse', 'loss_survival', 'c_index'
] # log the autoencoder loss and the classification loss
if opt.log_time:
self._metrics += ['t_recon', 't_survival']
# init variables
self.init_vars(add_path=True)
# init weights
self.init_weight(self.net_q)
self.init_weight(self.net_p)
self.init_weight(self.net_c)
def __init__(self, opt, logger):
super().__init__(opt, logger)
self.net_N = opt.net_N
self.x_dim = opt.x_dim
self.z_dim = opt.z_dim
self.n_classes = opt.n_classes
self.p_drop = opt.p_drop
self.modality = opt.modality
self.n_modality = len(self.modality)
# init networks
self.net_q = Q_net(self.net_N, self.x_dim, self.z_dim, self.p_drop)
self.net_p = P_net(self.net_N, self.x_dim, self.z_dim, self.p_drop)
self.net_c = C_net(self.net_N, self.z_dim, self.n_classes, self.p_drop)
self._nets = [self.net_q, self.net_p, self.net_c]
# optimizers
self.optimizer_q = self.optimizer(self.net_q.parameters(),
lr=opt.lr,
**self.optimizer_params)
self.optimizer_p = self.optimizer(self.net_p.parameters(),
lr=opt.lr,
**self.optimizer_params)
self.optimizer_c = self.optimizer(self.net_c.parameters(),
lr=opt.lr,
**self.optimizer_params)
self._optimizers = [
self.optimizer_q, self.optimizer_p, self.optimizer_c
]
#schedulers
self.scheduler_q = self.scheduler(self.optimizer_q,
**self.scheduler_params)
self.scheduler_p = self.scheduler(self.optimizer_p,
**self.scheduler_params)
self.scheduler_c = self.scheduler(self.optimizer_c,
**self.scheduler_params)
self._schedulers = [
self.scheduler_q, self.scheduler_p, self.scheduler_c
]
# general
self.opt = opt
self._metrics = [
'loss_mse', 'loss_clf'
] # log the autoencoder loss and the classification loss
if opt.log_time:
self._metrics += ['t_recon', 't_clf']
# init variables
#self.input_names = ['features', 'labels']
self.init_vars(add_path=True)
# init weights
self.init_weight(self.net_q)
self.init_weight(self.net_p)
self.init_weight(self.net_c)
def __init__(self, opt, logger):
super().__init__(opt, logger)
self.net_N = opt.net_N
self.x_dim = opt.x_dim
self.z_dim = opt.z_dim
self.n_classes = opt.n_classes
self.p_drop = opt.p_drop
self.modality = opt.modality[
0] # get the fisrt element since changed this argument to list
self.clf_weights = opt.clf_weights
# init networks
self.net_q = Q_net(self.net_N, self.x_dim, self.z_dim, self.p_drop)
self.net_c = C_net(self.net_N, self.z_dim, self.n_classes, self.p_drop)
self._nets = [self.net_q, self.net_c]
# optimizers
# self.optimizer and self.optimizer_params have been initialized in netinterface
self.optimizer_q = self.optimizer(self.net_q.parameters(),
lr=opt.lr,
**self.optimizer_params)
self.optimizer_c = self.optimizer(self.net_c.parameters(),
lr=opt.lr,
**self.optimizer_params)
self._optimizers = [self.optimizer_q, self.optimizer_c]
# schedulers
# self.scheduler and self.scheduler_params have been initialized in netinterface
self.scheduler_q = self.scheduler(self.optimizer_q,
**self.scheduler_params)
self.scheduler_c = self.scheduler(self.optimizer_c,
**self.scheduler_params)
self._schedulers = [self.scheduler_q, self.scheduler_c]
# general
self.opt = opt
self._metrics = [
'loss_clf'
] # log the autoencoder loss and the classification loss
if opt.log_time:
self._metrics += ['t_clf']
# init variables
self.init_vars(add_path=True)
# init weights
self.init_weight(self.net_q)
self.init_weight(self.net_c)
def __init__(self, opt, logger):
super().__init__(opt, logger)
self.net_N = opt.net_N
self.x_dim = opt.x_dim
self.z_dim = opt.z_dim
self.n_classes = opt.n_classes
self.p_drop = opt.p_drop
self.modality = opt.modality
self.n_modality = len(self.modality)
self.clf_weights = opt.clf_weights
assert len(self.x_dim) == self.n_modality # check the length of x_dim
# init networks
self.net_q = [None] * self.n_modality
for i in range(self.n_modality):
self.net_q[i] = Q_net(self.net_N, self.x_dim[i], self.z_dim,
self.p_drop)
self.net_c = C_net(self.net_N, self.z_dim * self.n_modality,
self.n_classes, self.p_drop)
self._nets = self.net_q + [self.net_c]
# optimizers
self.optimizer_q = [None] * self.n_modality
for i in range(self.n_modality):
self.optimizer_q[i] = self.optimizer(self.net_q[i].parameters(),
lr=opt.lr,
**self.optimizer_params)
self.optimizer_c = self.optimizer(self.net_c.parameters(),
lr=opt.lr,
**self.optimizer_params)
self._optimizers = self.optimizer_q + [self.optimizer_c]
# scheduler
self.scheduler_q = [None] * self.n_modality
for i in range(self.n_modality):
self.scheduler_q[i] = self.scheduler(self.optimizer_q[i],
**self.scheduler_params)
self.scheduler_c = self.scheduler(self.optimizer_c,
**self.scheduler_params)
self._schedulers = self.scheduler_q + [self.scheduler_c]
# general
self.opt = opt
self._metrics = [
'loss_survival', 'c_index'
] # log the autoencoder loss and the classification loss
if opt.log_time:
self._metrics += ['t_survival']
# init variables
#self.input_names = ['features', 'labels']
self.init_vars(add_path=True)
# init weights
for i in range(self.n_modality):
self.init_weight(self.net_q[i])
self.init_weight(self.net_c)
def __init__(self, args):
self.args = args
self.policy = [Q_net(args) for _ in range(args.n_agents)]
self.hyperNet = HyperNet(args)
self.policy_target = [copy.deepcopy(p) for p in self.policy]
self.hyperNet_target = copy.deepcopy(self.hyperNet)
self.replayBuffer = ReplayBuffer(args)
self.preference_pool = Preference(args)
policy_param = [policy.parameters() for policy in self.policy]
self.optim = torch.optim.Adam(itertools.chain(
*policy_param, self.hyperNet.parameters()),
lr=self.args.learning_rate)
self.lr_scheduler = torch.optim.lr_scheduler.StepLR(self.optim,
step_size=10,
gamma=0.95,
last_epoch=-1)
class Model(NetInterface):
@classmethod
def add_arguments(cls, parser):
parser.add_argument('--net_N',
type=int,
default=128,
help="Number of neurons in hidden layers")
parser.add_argument('--x_dim',
type=int,
default=1000,
help="dimension of input features")
parser.add_argument('--z_dim',
type=int,
default=100,
help="dimension of hidden variables")
parser.add_argument('--n_classes',
type=int,
default=1,
help="number of nodes for classification")
parser.add_argument('--p_drop',
type=float,
default=0.2,
help="probability of dropout")
return parser, set()
def __init__(self, opt, logger):
super().__init__(opt, logger)
self.net_N = opt.net_N
self.x_dim = opt.x_dim
self.z_dim = opt.z_dim
self.n_classes = opt.n_classes
self.p_drop = opt.p_drop
self.modality = opt.modality
self.n_modality = len(self.modality)
# init networks
self.net_q = Q_net(self.net_N, self.x_dim, self.z_dim, self.p_drop)
self.net_p = P_net(self.net_N, self.x_dim, self.z_dim, self.p_drop)
self.net_c = C_net(self.net_N, self.z_dim, self.n_classes, self.p_drop)
self._nets = [self.net_q, self.net_p, self.net_c]
# optimizers
self.optimizer_q = self.optimizer(self.net_q.parameters(),
lr=opt.lr,
**self.optimizer_params)
self.optimizer_p = self.optimizer(self.net_p.parameters(),
lr=opt.lr,
**self.optimizer_params)
self.optimizer_c = self.optimizer(self.net_c.parameters(),
lr=opt.lr,
**self.optimizer_params)
self._optimizers = [
self.optimizer_q, self.optimizer_p, self.optimizer_c
]
#schedulers
self.scheduler_q = self.scheduler(self.optimizer_q,
**self.scheduler_params)
self.scheduler_p = self.scheduler(self.optimizer_p,
**self.scheduler_params)
self.scheduler_c = self.scheduler(self.optimizer_c,
**self.scheduler_params)
self._schedulers = [
self.scheduler_q, self.scheduler_p, self.scheduler_c
]
# general
self.opt = opt
self._metrics = [
'loss_mse', 'loss_clf'
] # log the autoencoder loss and the classification loss
if opt.log_time:
self._metrics += ['t_recon', 't_clf']
# init variables
#self.input_names = ['features', 'labels']
self.init_vars(add_path=True)
# init weights
self.init_weight(self.net_q)
self.init_weight(self.net_p)
self.init_weight(self.net_c)
def __str__(self):
s = "Concat + Autoencoder"
return s
def _train_on_batch(self, epoch, batch_idx, batch):
net_q, net_p, net_c = self.net_q, self.net_p, self.net_c
opt_q, opt_p, opt_c = self.optimizer_q, self.optimizer_p, self.optimizer_c
net_q.train()
net_p.train()
net_c.train()
# read in X
X_list = [None] * self.n_modality
for i in range(self.n_modality):
X_list[i] = batch[self.modality[i]]
X_list[i] = X_list[i].view(X_list[i].shape[0], -1)
X_list = [tmp.cuda() for tmp in X_list]
# Concatenate X before feeding into Autoencoder
X = torch.cat(X_list, dim=1).cuda()
y = batch['labels'].cuda()
batch_size = X.shape[0]
batch_log = {'size': batch_size}
# Stage 1: train Q and P with reconstruction loss
net_q.zero_grad()
net_p.zero_grad()
for p in net_q.parameters():
p.requires_grad = True
for p in net_p.parameters():
p.requires_grad = True
for p in net_c.parameters():
p.requires_grad = False
t0 = time()
z = net_q(X)
X_recon = net_p(z)
loss_mse = nn.functional.mse_loss(X_recon, X) # Mean square error
loss_mse.backward()
opt_q.step()
opt_p.step()
t_recon = time() - t0
batch_log['loss_mse'] = loss_mse.item()
# Stage 2: train Q and C with classification loss
net_q.zero_grad()
net_c.zero_grad()
for p in net_q.parameters():
p.requires_grad = True
for p in net_c.parameters():
p.requires_grad = True
for p in net_p.parameters():
p.requires_grad = False
t0 = time()
z = net_q(X)
pred = net_c(z)
# calculate loss with cross entropy loss
criterion = nn.CrossEntropyLoss()
loss_clf = criterion(pred, y)
loss_clf.backward()
opt_q.step()
opt_c.step()
t_clf = time() - t0
batch_log['loss_clf'] = loss_clf.item()
if self.opt.log_time:
batch_log['t_recon'] = t_recon
batch_log['t_clf'] = t_clf
return batch_log
def _vali_on_batch(self, epoch, batch_idx, batch):
self.net_q.eval()
self.net_c.eval()
# read in X
X_list = [None] * self.n_modality
for i in range(self.n_modality):
X_list[i] = batch[self.modality[i]]
X_list[i] = X_list[i].view(X_list[i].shape[0], -1)
X_list = [tmp.cuda() for tmp in X_list]
# concatenate X before feeding into Autoencoder
X = torch.cat(X_list, dim=1).cuda()
y = batch['labels'].cuda()
batch_size = X.shape[0]
batch_log = {'size': batch_size}
with torch.no_grad():
z = self.net_q(X)
pred = self.net_c(z)
criterion = nn.CrossEntropyLoss()
loss_clf = criterion(pred, y)
batch_log['loss_clf'] = loss_clf.item()
batch_log['loss'] = loss_clf.item()
return batch_log
class Model(NetInterface):
@classmethod
def add_arguments(cls, parser):
parser.add_argument('--net_N',
type=int,
default=128,
help="Number of neurons in hidden layers")
parser.add_argument('--x_dim',
type=int,
default=1000,
help="dimension of input features")
parser.add_argument('--z_dim',
type=int,
default=100,
help="dimension of hidden variables")
parser.add_argument('--n_classes',
type=int,
default=1,
help="number of nodes for classification")
parser.add_argument('--p_drop',
type=float,
default=0.2,
help="probability of dropout")
parser.add_argument('--clf_weights',
type=float,
nargs='+',
default=0.7,
help="classification weight for each class")
return parser, set()
def __init__(self, opt, logger):
super().__init__(opt, logger)
self.net_N = opt.net_N
self.x_dim = opt.x_dim
self.z_dim = opt.z_dim
self.n_classes = opt.n_classes
self.p_drop = opt.p_drop
self.modality = opt.modality[
0] # get the fisrt element since changed this argument to list
self.clf_weights = opt.clf_weights
# init networks
self.net_q = Q_net(self.net_N, self.x_dim, self.z_dim, self.p_drop)
self.net_c = C_net(self.net_N, self.z_dim, self.n_classes, self.p_drop)
self._nets = [self.net_q, self.net_c]
# optimizers
# self.optimizer and self.optimizer_params have been initialized in netinterface
self.optimizer_q = self.optimizer(self.net_q.parameters(),
lr=opt.lr,
**self.optimizer_params)
self.optimizer_c = self.optimizer(self.net_c.parameters(),
lr=opt.lr,
**self.optimizer_params)
self._optimizers = [self.optimizer_q, self.optimizer_c]
# schedulers
# self.scheduler and self.scheduler_params have been initialized in netinterface
self.scheduler_q = self.scheduler(self.optimizer_q,
**self.scheduler_params)
self.scheduler_c = self.scheduler(self.optimizer_c,
**self.scheduler_params)
self._schedulers = [self.scheduler_q, self.scheduler_c]
# general
self.opt = opt
self._metrics = [
'loss_clf'
] # log the autoencoder loss and the classification loss
if opt.log_time:
self._metrics += ['t_clf']
# init variables
self.init_vars(add_path=True)
# init weights
self.init_weight(self.net_q)
self.init_weight(self.net_c)
def __str__(self):
s = "Autoencoder"
return s
def _train_on_batch(self, epoch, batch_idx, batch):
net_q, net_c = self.net_q, self.net_c
opt_q, opt_c = self.optimizer_q, self.optimizer_c
net_q.train()
net_c.train()
X = batch[self.modality].cuda()
#print(X.shape) # (batchsize, 1, 28, 28)
# Flatten the X from 28x28 to 784
X = X.view(X.shape[0], -1)
#print(X.shape) # (batchsize, 784)
y = batch['labels'].cuda()
batch_size = X.shape[0]
batch_log = {'size': batch_size}
net_q.zero_grad()
net_c.zero_grad()
for p in net_q.parameters():
p.requires_grad = True
for p in net_c.parameters():
p.requires_grad = True
t0 = time()
z = net_q(X)
pred = net_c(z)
criterion = nn.CrossEntropyLoss()
# the cross-entropy loss combines the 1)LogSoftmax() and 2) Negative log-likelihood loss NLLLoss
loss_clf = criterion(pred, y)
loss_clf.backward()
opt_q.step()
opt_c.step()
t_clf = time() - t0
batch_log['loss_clf'] = loss_clf.item()
if self.opt.log_time:
batch_log['t_clf'] = t_clf
return batch_log
def _vali_on_batch(self, epoch, batch_idx, batch):
self.net_q.eval()
self.net_c.eval()
X = batch[self.modality].cuda()
X = X.view(X.shape[0], -1) # flatten the X
y = batch['labels'].cuda()
batch_size = X.shape[0]
batch_log = {'size': batch_size}
with torch.no_grad():
z = self.net_q(X)
pred = self.net_c(z)
criterion = nn.CrossEntropyLoss()
# the cross-entropy loss combines the 1)LogSoftmax() and 2) Negative log-likelihood loss NLLLoss
loss_clf = criterion(pred, y)
batch_log['loss'] = loss_clf.item()
batch_log['loss_clf'] = loss_clf.item()
return batch_log
class Model(NetInterface):
@classmethod
def add_arguments(cls, parser):
parser.add_argument('--net_N',
type=int,
default=128,
help="Number of neurons in hidden layers")
parser.add_argument('--x_dim',
type=int,
default=1000,
help="dimension of input features")
parser.add_argument('--z_dim',
type=int,
default=100,
help="dimension of hidden variables")
parser.add_argument('--n_classes',
type=int,
default=1,
help="number of nodes for classification")
parser.add_argument('--p_drop',
type=float,
default=0.2,
help="probability of dropout")
parser.add_argument('--clf_weights',
type=float,
nargs='+',
default=0.7,
help="classification weight for each class")
return parser, set()
def __init__(self, opt, logger):
super().__init__(opt, logger)
self.net_N = opt.net_N
self.x_dim = opt.x_dim
self.z_dim = opt.z_dim
self.n_classes = opt.n_classes
self.p_drop = opt.p_drop
self.modality = opt.modality
self.clf_weights = opt.clf_weights
# init networks
self.net_q = Q_net(self.net_N, self.x_dim, self.z_dim, self.p_drop)
self.net_p = P_net(self.net_N, self.x_dim, self.z_dim, self.p_drop)
self.net_c = C_net(self.net_N, self.z_dim, self.n_classes, self.p_drop)
self._nets = [self.net_q, self.net_p, self.net_c]
# optimizers
# self.optimizer and self.optimizer_params have been initialized in netinterface
self.optimizer_q = self.optimizer(self.net_q.parameters(),
lr=opt.lr,
**self.optimizer_params)
self.optimizer_p = self.optimizer(self.net_p.parameters(),
lr=opt.lr,
**self.optimizer_params)
self.optimizer_c = self.optimizer(self.net_c.parameters(),
lr=opt.lr,
**self.optimizer_params)
self._optimizers = [
self.optimizer_q, self.optimizer_p, self.optimizer_c
]
# schedulers
# self.scheduler and self.scheduler_params have been initialized in netinterface
self.scheduler_q = self.scheduler(self.optimizer_q,
**self.scheduler_params)
self.scheduler_p = self.scheduler(self.optimizer_p,
**self.scheduler_params)
self.scheduler_c = self.scheduler(self.optimizer_c,
**self.scheduler_params)
self._schedulers = [
self.scheduler_q, self.scheduler_p, self.scheduler_c
]
# general
self.opt = opt
self._metrics = [
'loss_mse', 'loss_survival', 'c_index'
] # log the autoencoder loss and the classification loss
if opt.log_time:
self._metrics += ['t_recon', 't_survival']
# init variables
self.init_vars(add_path=True)
# init weights
self.init_weight(self.net_q)
self.init_weight(self.net_p)
self.init_weight(self.net_c)
def __str__(self):
s = "Autoencoder"
return s
def _train_on_batch(self, epoch, batch_idx, batch):
net_q, net_p, net_c = self.net_q, self.net_p, self.net_c
opt_q, opt_p, opt_c = self.optimizer_q, self.optimizer_p, self.optimizer_c
net_q.train()
net_p.train()
net_c.train()
X = batch[self.modality].cuda()
survival_time = batch['days'].cuda()
survival_event = batch['event'].cuda()
batch_size = X.shape[0]
batch_log = {'size': batch_size}
# Stage 1: train Q and P with reconstruction loss
net_q.zero_grad()
net_p.zero_grad()
for p in net_q.parameters():
p.requires_grad = True
for p in net_p.parameters():
p.requires_grad = True
for p in net_c.parameters():
p.requires_grad = False
t0 = time()
z = net_q(X)
X_recon = net_p(z)
loss_mse = nn.functional.mse_loss(X_recon, X) # Mean square error
loss_mse.backward()
opt_q.step()
opt_p.step()
t_recon = time() - t0
batch_log['loss_mse'] = loss_mse.item()
# Stage 2: train Q and C with classification loss
if not survival_event.sum(
0): # skip the batch if all instances are negative
batch_log['loss_survial'] = torch.Tensor([float('nan')])
return batch_log
net_q.zero_grad()
net_c.zero_grad()
for p in net_q.parameters():
p.requires_grad = True
for p in net_c.parameters():
p.requires_grad = True
for p in net_p.parameters():
p.requires_grad = False
t0 = time()
z = net_q(X)
pred = net_c(z)
loss_survival = neg_par_log_likelihood(pred, survival_time,
survival_event)
loss_survival.backward()
opt_q.step()
opt_c.step()
t_survival = time() - t0
c_index = CIndex(pred, survival_event, survival_time)
batch_log['loss_survival'] = loss_survival.item()
batch_log['c_index'] = c_index
if self.opt.log_time:
batch_log['t_recon'] = t_recon
batch_log['t_survival'] = t_survival
return batch_log
def _vali_on_batch(self, epoch, batch_idx, batch):
self.net_q.eval()
self.net_c.eval()
# load the data
X = batch[self.modality].cuda()
survival_time = batch['days'].cuda()
survival_event = batch['event'].cuda()
batch_size = X.shape[0]
batch_log = {'size': batch_size}
if not survival_event.sum(
0): # skip the batch if all instances are negative
batch_log['loss_survival'] = torch.Tensor([float('nan')])
return batch_log
with torch.no_grad():
z = self.net_q(X)
pred = self.net_c(z)
loss_survival = neg_par_log_likelihood(pred, survival_time,
survival_event)
c_index = CIndex(pred, survival_event, survival_time)
batch_log['loss'] = loss_survival.item()
batch_log['loss_survival'] = loss_survival.item()
batch_log['c_index'] = c_index
return batch_log