def fitXy(self, x, y, batch_size=100, n_epochs=10, print_every=10,
valdata=None):
'''
fit a model to x, y data by batch
'''
n, d = x.shape
time_start = time.time()
losses = []
cost = 0
for epoch in range(n_epochs):
x, y = data_shuffle(x, y)
num_batches = math.ceil(n / batch_size)
for k in range(num_batches):
start, end = k * batch_size, min((k + 1) * batch_size, n)
x_batch, y_batch = prepareData(x[start:end], y[start:end])
y_hat, regret = self.step(x_batch, y_batch)
m = end-start
cost += 1 / (k+1) * (regret/m - cost)
if print_every != 0 and k % print_every == 0:
losses.append(cost)
print('%.2f%% (%s) %.4f' % ((epoch * n + (k+1) * (end-start)) /
(n_epochs * n) * 100, # progress
timeSince(time_start), # time
cost)) # cost
torch.save(self.model, 'models/%s.pt' % self.name)
np.save('models/%s.loss' % self.name, losses)
return losses
def fitData(self, data, batch_size=100, n_epochs=10, print_every=10,
valdata=None):
'''
fit a model to x, y data by batch
print_every is 0 if do not wish to print
'''
time_start = time.time()
losses = []
best_valloss, best_valindex = np.inf, 0
vallosses = []
n = len(data.dataset)
cost = 0
for epoch in range(n_epochs):
for k, (x_batch, y_batch) in enumerate(data):
x_batch, y_batch = to_var(x_batch), to_var(y_batch)
y_hat, regret = self.step(x_batch, y_batch)
m = x_batch.size(0)
cost += 1 / (k+1) * (regret/m - cost)
if print_every != 0 and k % print_every == 0:
losses.append(cost)
# progress, time, avg loss, auc
to_print = ('%.2f%% (%s) %.4f %.4f' % ((epoch * n + (k+1) * m) /
(n_epochs * n) * 100,
timeSince(time_start),
cost,
model_auc(self.model,
data)))
if valdata is not None:
valloss = calc_loss(self.model, valdata, self.loss)
vallosses.append(valloss)
np.save('models/%s.valloss' % self.name, vallosses)
to_print += " %.4f" % model_auc(self.model, valdata)
if valloss <= best_valloss:
best_valloss = valloss
best_valindex = len(vallosses) - 1
torch.save(self.model, 'models/%s.pt' % self.name)
else:
torch.save(self.model, 'models/%s.pt' % self.name)
print(to_print)
np.save('models/%s.loss' % self.name, losses)
cost = 0
return losses, vallosses
def fit(self, data, batch_size=100, n_epochs=10,
valdata=None, val_theta=None, use_auc=False):
'''
fit a model to x, y data by batch
val_theta: for recovering heterogeneous subpopulation
'''
savedir = os.path.dirname('nonlinear_models/%s' % self.name)
os.system('mkdir -p %s' % savedir)
self.writer = SummaryWriter(log_dir=self.log_dir)
time_start = time.time()
losses = []
vallosses = [1000]
best_valloss, best_valindex = 1000, 0 # for early stopping
n = len(data.dataset)
cost = 0
self.count = 0
for epoch in range(n_epochs):
for k, (x_batch, y_batch) in enumerate(data):
x_batch, y_batch = to_var(x_batch).float(), to_var(y_batch).float()
y_hat, regret = self.step(x_batch, y_batch)
m = x_batch.size(0)
cost += 1 / (k+1) * (regret - cost)
if self.print_every != 0 and self.count % self.print_every == 0:
losses.append(cost)
# progress, time, avg loss, auc
duration = timeSince(time_start)
if int(duration.split('m')[0]) >= self.max_time:
return losses
to_print = ('%.2f%% (%s) %.4f' % ((epoch * n + (k+1) * m) /
(n_epochs * n) * 100,
duration,
cost))
print(to_print)
# if self.draw_plot:
# self.plotMTL()
# self.plot(x_batch, y_batch, silence=self.silence, inrange=True)
if valdata is not None:
if use_auc:
acc = reportAuc(self, valdata)
else:
acc = reportAcc(self,valdata)
valloss = -acc
vallosses.append(valloss)
if valloss <= best_valloss:
best_valloss = valloss
best_valindex = len(vallosses) - 1
torch.save(self.weightNet,
'nonlinear_models/%s.pt' % self.name)
np.save('nonlinear_models/%s.loss' % self.name, losses)
if len(vallosses) - best_valindex > self.n_early_stopping:
print('early stop at iteration', self.count)
return losses
if use_auc:
# note acc here is auc
self.writer.add_scalar('data/val_auc', acc,
self.count)
else:
self.writer.add_scalar('data/val_acc', acc,
self.count)
if val_theta is not None:
sim = self.evaluate_subpopulation(val_theta, valdata)
self.writer.add_scalar('data/subpopulation_cosine',
sim, self.count)
self.writer.add_scalar('weight/grad_norm', gradNorm(self.weightNet),
self.count)
self.writer.add_scalar('data/train_loss', cost, self.count)
# for tag, value in self.weightNet.named_parameters():
# tag = tag.replace('.', '/')
# self.writer.add_histogram(tag, to_np(value), self.count)
# if value.grad is not None:
# self.writer.add_histogram(tag+'/grad', to_np(value.grad),
# self.count)
cost = 0
self.count += 1
# if self.draw_plot:
# self.plot(x_batch, y_batch, inrange=True, silence=self.silence)
return losses
def fit(self, data, batch_size=100, n_epochs=10,
valdata=None, val_theta=None):
'''
fit a model to x, y data by batch
val_theta: for recovering heterogeneous subpopulation
'''
savedir = os.path.dirname('nonlinear_models/%s' % self.name)
os.system('mkdir -p %s' % savedir)
self.writer = SummaryWriter(log_dir=self.log_dir)
time_start = time.time()
losses = []
vallosses = [1000]
best_valloss, best_valindex = 1000, 0 # for early stopping
n = len(data.dataset)
cost = 0
self.count = 0
for epoch in range(n_epochs):
for k, (x_batch, y_batch) in enumerate(data):
x_batch, y_batch = to_var(x_batch).float(), to_var(y_batch).float()
y_hat, regret = self.step(x_batch, y_batch)
m = x_batch.size(0)
cost += 1 / (k+1) * (regret - cost)
if self.print_every != 0 and self.count % self.print_every == 0:
losses.append(cost)
# progress, time, avg loss, auc
duration = timeSince(time_start)
if int(duration.split('m')[0]) >= self.max_time:
return losses
to_print = ('%.2f%% (%s) %.4f' % ((epoch * n + (k+1) * m) /
(n_epochs * n) * 100,
duration,
cost))
print(to_print)
if valdata is not None:
_mse = reportMSE(self,valdata,is_autoencoder=True)
valloss = _mse
vallosses.append(valloss)
if valloss <= best_valloss:
best_valloss = valloss
best_valindex = len(vallosses) - 1
torch.save(self.autoencoder,
'nonlinear_models/%s.pt' % self.name)
np.save('nonlinear_models/%s.loss' % self.name, losses)
if len(vallosses) - best_valindex > self.n_early_stopping:
print('early stop at iteration', self.count)
return losses
self.writer.add_scalar('data/val_mse', _mse, self.count)
self.writer.add_scalar('model/grad_norm', gradNorm(self.autoencoder),
self.count)
# self.writer.add_scalar('data/train_loss', cost, self.count)
cost = 0
self.count += 1
return losses