def __init__(self, model, data, args, axis='label', verbose=False):
self.axis = axis
# self.flow = deeper_flow(layer_num=5, feature_size=20).to(args.device)
self.flow = NafFlow().to(
args.device
) # This flow model is too simple, might need more layers and latents?
flow_optim = optim.Adam(self.flow.parameters(), lr=1e-3)
k = args.knn
assert k % 2 == 0
inputs, labels = data
inputs = inputs.to(args.device)
labels = labels.to(args.device).flatten()
for iteration in range(5000):
flow_optim.zero_grad()
outputs = model(inputs).flatten()
if axis == 'label':
ranking = torch.argsort(labels)
else:
assert axis == 'prediction'
ranking = torch.argsort(outputs)
sorted_labels = labels[ranking]
sorted_outputs = outputs[ranking]
smoothed_outputs = F.conv1d(
sorted_outputs.view(1, 1, -1),
weight=(1. / (k + 1) * torch.ones(
1, 1, k + 1, device=args.device, requires_grad=False)),
padding=k // 2).flatten()
smoothed_labels = F.conv1d(
sorted_labels.view(1, 1, -1),
weight=(1. / (k + 1) * torch.ones(
1, 1, k + 1, device=args.device, requires_grad=False)),
padding=k // 2).flatten()
# print(smoothed_outputs.shape)
# print(smoothed_labels.shape)
# loss_bias = smoothed_labels - smoothed_outputs
if axis == 'label':
adjusted_labels, _ = self.flow(smoothed_labels.view(-1, 1))
# adjusted_outputs = self.flow.invert(smoothed_outputs.view(-1, 1))
loss_bias = (adjusted_labels.view(-1) -
smoothed_outputs).pow(2).mean()
elif axis == 'prediction':
adjusted_outputs, _ = self.flow(smoothed_outputs.view(-1, 1))
loss_bias = (smoothed_labels -
adjusted_outputs.view(-1)).pow(2).mean()
loss_bias.backward()
flow_optim.step()
if verbose and iteration % 100 == 0:
print("Iteration %d, loss_bias=%.5f" % (iteration, loss_bias))
def __init__(self,
model,
args,
re_calib=False,
re_bias_f=False,
re_bias_y=False,
verbose=False):
self.args = args
self.model = model # regression model
self.flow = NafFlow(feature_size=40).to(
args.device
) # This flow model is too simple, might need more layers and latents?
self.flow_optim = optim.Adam(self.flow.parameters(), lr=1e-3)
self.flow_scheduler = torch.optim.lr_scheduler.StepLR(
self.flow_optim,
step_size=args.num_iter // (args.flow_skip * 20),
gamma=0.9)
self.re_calib = re_calib
self.re_bias_f = re_bias_f
self.re_bias_y = re_bias_y
assert self.args.knn % 2 == 0
assert re_calib or re_bias_f or re_bias_y
class RecalibratorOnline:
def __init__(self,
model,
args,
re_calib=False,
re_bias_f=False,
re_bias_y=False,
verbose=False):
self.args = args
self.model = model # regression model
self.flow = NafFlow(feature_size=40).to(
args.device
) # This flow model is too simple, might need more layers and latents?
self.flow_optim = optim.Adam(self.flow.parameters(), lr=1e-3)
self.flow_scheduler = torch.optim.lr_scheduler.StepLR(
self.flow_optim,
step_size=args.num_iter // (args.flow_skip * 20),
gamma=0.9)
self.re_calib = re_calib
self.re_bias_f = re_bias_f
self.re_bias_y = re_bias_y
assert self.args.knn % 2 == 0
assert re_calib or re_bias_f or re_bias_y
def train_step(self, data):
k = self.args.knn
inputs, labels = data
inputs = inputs.to(self.args.device)
labels = labels.to(self.args.device).flatten()
self.flow_optim.zero_grad()
loss_all = 0.0
outputs = self.model(inputs).flatten()
for objective in range(2):
if objective == 0 and self.re_bias_f:
ranking = torch.argsort(outputs)
elif objective == 1 and self.re_bias_y:
ranking = torch.argsort(labels)
else:
continue
sorted_labels = labels[ranking]
sorted_outputs = outputs[ranking]
smoothed_outputs = F.conv1d(
sorted_outputs.view(1, 1, -1),
weight=(1. / (k + 1) * torch.ones(
1, 1, k + 1, device=self.args.device,
requires_grad=False)),
padding=k // 2).flatten()
smoothed_labels = F.conv1d(
sorted_labels.view(1, 1, -1),
weight=(1. / (k + 1) * torch.ones(
1, 1, k + 1, device=self.args.device,
requires_grad=False)),
padding=k // 2).flatten()
adjusted_outputs, _ = self.flow(smoothed_outputs.view(-1, 1))
loss_bias = (smoothed_labels -
adjusted_outputs.view(-1)).pow(2).mean()
loss_all += loss_bias
if self.re_calib:
labels = torch.sort(labels.flatten())[0]
outputs = torch.sort(outputs.flatten())[0]
adjusted_outputs, _ = self.flow(outputs.view(-1, 1))
loss_bias = (labels - adjusted_outputs.view(-1)).pow(2).mean()
loss_all += loss_bias
loss_all.backward()
self.flow_optim.step()
self.flow_scheduler.step()
return loss_all
def adjust(self, original_y):
original_shape = original_y.shape
adjusted_output, _ = self.flow(original_y.view(-1, 1))
return adjusted_output.view(original_shape)
def __init__(self,
model,
data,
args,
re_calib=False,
re_bias_f=False,
re_bias_y=False,
verbose=False):
self.args = args
self.re_calib = re_calib
self.re_bias_f = re_bias_f
self.re_bias_y = re_bias_y
self.model = model # regression model
self.flow = NafFlow(feature_size=40).to(
args.device
) # This flow model is too simple, might need more layers and latents?
flow_optim = optim.Adam(self.flow.parameters(), lr=1e-3)
# flow_scheduler = torch.optim.lr_scheduler.StepLR(flow_optim, step_size=20, gamma=0.9)
k = args.knn
assert k % 2 == 0
assert re_calib or re_bias_f or re_bias_y
inputs, labels = data
inputs = inputs.to(self.args.device)
labels = labels.to(self.args.device).flatten()
for iteration in range(5000):
flow_optim.zero_grad()
loss_all = 0.0
outputs = model(inputs).flatten()
for objective in range(2):
if objective == 0 and self.re_bias_f:
ranking = torch.argsort(outputs)
elif objective == 1 and self.re_bias_y:
ranking = torch.argsort(labels)
else:
continue
sorted_labels = labels[ranking]
sorted_outputs = outputs[ranking]
smoothed_outputs = F.conv1d(
sorted_outputs.view(1, 1, -1),
weight=(1. / (k + 1) * torch.ones(
1, 1, k + 1, device=args.device, requires_grad=False)),
padding=k // 2).flatten()
smoothed_labels = F.conv1d(
sorted_labels.view(1, 1, -1),
weight=(1. / (k + 1) * torch.ones(
1, 1, k + 1, device=args.device, requires_grad=False)),
padding=k // 2).flatten()
adjusted_outputs, _ = self.flow(smoothed_outputs.view(-1, 1))
loss_bias = (smoothed_labels -
adjusted_outputs.view(-1)).pow(2).mean()
loss_all += loss_bias
if re_calib:
labels = torch.sort(labels.flatten())[0]
outputs = torch.sort(outputs.flatten())[0]
adjusted_outputs, _ = self.flow(outputs.view(-1, 1))
loss_bias = (labels - adjusted_outputs.view(-1)).pow(2).mean()
loss_all += loss_bias
loss_all.backward()
flow_optim.step()
# flow_scheduler.step()
if verbose and iteration % 100 == 0:
print("Iteration %d, loss_bias=%.5f" % (iteration, loss_bias))