mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-08 07:39:33 +01:00
b6f130aa70
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17236 Following the paper in https://papers.nips.cc/paper/7141-what-uncertainties-do-we-need-in-bayesian-deep-learning-for-computer-vision.pdf, approximate the classification case with the regression formulation. For the LRLoss, add penalty based on the variance and regularization on the variance with a tunable parameter lambda. Reviewed By: chocjy Differential Revision: D14077106 fbshipit-source-id: 4405d8995cebdc7275a0dd07857d32a8915d78ef
240 lines
8.2 KiB
Python
240 lines
8.2 KiB
Python
## @package batch_lr_loss
|
|
# Module caffe2.python.layers.batch_lr_loss
|
|
from __future__ import absolute_import
|
|
from __future__ import division
|
|
from __future__ import print_function
|
|
from __future__ import unicode_literals
|
|
|
|
from caffe2.python import core, schema
|
|
from caffe2.python.layers.layers import (
|
|
ModelLayer,
|
|
)
|
|
from caffe2.python.layers.tags import (
|
|
Tags
|
|
)
|
|
import numpy as np
|
|
|
|
|
|
class BatchLRLoss(ModelLayer):
|
|
|
|
def __init__(
|
|
self,
|
|
model,
|
|
input_record,
|
|
name='batch_lr_loss',
|
|
average_loss=True,
|
|
jsd_weight=0.0,
|
|
pos_label_target=1.0,
|
|
neg_label_target=0.0,
|
|
homotopy_weighting=False,
|
|
log_D_trick=False,
|
|
unjoined_lr_loss=False,
|
|
uncertainty_penalty=1.0,
|
|
**kwargs
|
|
):
|
|
super(BatchLRLoss, self).__init__(model, name, input_record, **kwargs)
|
|
|
|
self.average_loss = average_loss
|
|
|
|
assert (schema.is_schema_subset(
|
|
schema.Struct(
|
|
('label', schema.Scalar()),
|
|
('logit', schema.Scalar())
|
|
),
|
|
input_record
|
|
))
|
|
|
|
self.jsd_fuse = False
|
|
assert jsd_weight >= 0 and jsd_weight <= 1
|
|
if jsd_weight > 0 or homotopy_weighting:
|
|
assert 'prediction' in input_record
|
|
self.init_weight(jsd_weight, homotopy_weighting)
|
|
self.jsd_fuse = True
|
|
self.homotopy_weighting = homotopy_weighting
|
|
|
|
assert pos_label_target <= 1 and pos_label_target >= 0
|
|
assert neg_label_target <= 1 and neg_label_target >= 0
|
|
assert pos_label_target >= neg_label_target
|
|
self.pos_label_target = pos_label_target
|
|
self.neg_label_target = neg_label_target
|
|
|
|
assert not (log_D_trick and unjoined_lr_loss)
|
|
self.log_D_trick = log_D_trick
|
|
self.unjoined_lr_loss = unjoined_lr_loss
|
|
assert uncertainty_penalty >= 0
|
|
self.uncertainty_penalty = uncertainty_penalty
|
|
|
|
self.tags.update([Tags.EXCLUDE_FROM_PREDICTION])
|
|
|
|
self.output_schema = schema.Scalar(
|
|
np.float32,
|
|
self.get_next_blob_reference('output')
|
|
)
|
|
|
|
def init_weight(self, jsd_weight, homotopy_weighting):
|
|
if homotopy_weighting:
|
|
self.mutex = self.create_param(
|
|
param_name=('%s_mutex' % self.name),
|
|
shape=None,
|
|
initializer=('CreateMutex', ),
|
|
optimizer=self.model.NoOptim,
|
|
)
|
|
self.counter = self.create_param(
|
|
param_name=('%s_counter' % self.name),
|
|
shape=[1],
|
|
initializer=(
|
|
'ConstantFill', {
|
|
'value': 0,
|
|
'dtype': core.DataType.INT64
|
|
}
|
|
),
|
|
optimizer=self.model.NoOptim,
|
|
)
|
|
self.xent_weight = self.create_param(
|
|
param_name=('%s_xent_weight' % self.name),
|
|
shape=[1],
|
|
initializer=(
|
|
'ConstantFill', {
|
|
'value': 1.,
|
|
'dtype': core.DataType.FLOAT
|
|
}
|
|
),
|
|
optimizer=self.model.NoOptim,
|
|
)
|
|
self.jsd_weight = self.create_param(
|
|
param_name=('%s_jsd_weight' % self.name),
|
|
shape=[1],
|
|
initializer=(
|
|
'ConstantFill', {
|
|
'value': 0.,
|
|
'dtype': core.DataType.FLOAT
|
|
}
|
|
),
|
|
optimizer=self.model.NoOptim,
|
|
)
|
|
else:
|
|
self.jsd_weight = self.model.add_global_constant(
|
|
'%s_jsd_weight' % self.name, jsd_weight
|
|
)
|
|
self.xent_weight = self.model.add_global_constant(
|
|
'%s_xent_weight' % self.name, 1. - jsd_weight
|
|
)
|
|
|
|
def update_weight(self, net):
|
|
net.AtomicIter([self.mutex, self.counter], [self.counter])
|
|
# iter = 0: lr = 1;
|
|
# iter = 1e6; lr = 0.5^0.1 = 0.93
|
|
# iter = 1e9; lr = 1e-3^0.1 = 0.50
|
|
net.LearningRate([self.counter], [self.xent_weight], base_lr=1.0,
|
|
policy='inv', gamma=1e-6, power=0.1,)
|
|
net.Sub(
|
|
[self.model.global_constants['ONE'], self.xent_weight],
|
|
[self.jsd_weight]
|
|
)
|
|
return self.xent_weight, self.jsd_weight
|
|
|
|
def add_ops(self, net):
|
|
# numerically stable log-softmax with crossentropy
|
|
label = self.input_record.label()
|
|
# mandatory cast to float32
|
|
# self.input_record.label.field_type().base is np.float32 but
|
|
# label type is actually int
|
|
label = net.Cast(
|
|
label,
|
|
net.NextScopedBlob('label_float32'),
|
|
to=core.DataType.FLOAT)
|
|
label = net.ExpandDims(label, net.NextScopedBlob('expanded_label'),
|
|
dims=[1])
|
|
if self.pos_label_target != 1.0 or self.neg_label_target != 0.0:
|
|
label = net.StumpFunc(
|
|
label,
|
|
net.NextScopedBlob('smoothed_label'),
|
|
threshold=0.5,
|
|
low_value=self.neg_label_target,
|
|
high_value=self.pos_label_target,
|
|
)
|
|
xent = net.SigmoidCrossEntropyWithLogits(
|
|
[self.input_record.logit(), label],
|
|
net.NextScopedBlob('cross_entropy'),
|
|
log_D_trick=self.log_D_trick,
|
|
unjoined_lr_loss=self.unjoined_lr_loss
|
|
)
|
|
# fuse with JSD
|
|
if self.jsd_fuse:
|
|
jsd = net.BernoulliJSD(
|
|
[self.input_record.prediction(), label],
|
|
net.NextScopedBlob('jsd'),
|
|
)
|
|
if self.homotopy_weighting:
|
|
self.update_weight(net)
|
|
loss = net.WeightedSum(
|
|
[xent, self.xent_weight, jsd, self.jsd_weight],
|
|
net.NextScopedBlob('loss'),
|
|
)
|
|
else:
|
|
loss = xent
|
|
|
|
if 'log_variance' in self.input_record.fields:
|
|
# mean (0.5 * exp(-s) * loss + 0.5 * penalty * s)
|
|
log_variance_blob = self.input_record.log_variance()
|
|
|
|
log_variance_blob = net.ExpandDims(
|
|
log_variance_blob, net.NextScopedBlob('expanded_log_variance'),
|
|
dims=[1]
|
|
)
|
|
|
|
neg_log_variance_blob = net.Negative(
|
|
[log_variance_blob],
|
|
net.NextScopedBlob('neg_log_variance')
|
|
)
|
|
|
|
# enforce less than 88 to avoid OverflowError
|
|
neg_log_variance_blob = net.Clip(
|
|
[neg_log_variance_blob],
|
|
net.NextScopedBlob('clipped_neg_log_variance'),
|
|
max=88.0
|
|
)
|
|
|
|
exp_neg_log_variance_blob = net.Exp(
|
|
[neg_log_variance_blob],
|
|
net.NextScopedBlob('exp_neg_log_variance')
|
|
)
|
|
|
|
exp_neg_log_variance_loss_blob = net.Mul(
|
|
[exp_neg_log_variance_blob, loss],
|
|
net.NextScopedBlob('exp_neg_log_variance_loss')
|
|
)
|
|
|
|
penalized_uncertainty = net.Scale(
|
|
log_variance_blob, net.NextScopedBlob("penalized_unceratinty"),
|
|
scale=float(self.uncertainty_penalty)
|
|
)
|
|
|
|
loss_2x = net.Add(
|
|
[exp_neg_log_variance_loss_blob, penalized_uncertainty],
|
|
net.NextScopedBlob('loss')
|
|
)
|
|
loss = net.Scale(loss_2x, net.NextScopedBlob("loss"), scale=0.5)
|
|
|
|
if 'weight' in self.input_record.fields:
|
|
weight_blob = self.input_record.weight()
|
|
if self.input_record.weight.field_type().base != np.float32:
|
|
weight_blob = net.Cast(
|
|
weight_blob,
|
|
weight_blob + '_float32',
|
|
to=core.DataType.FLOAT
|
|
)
|
|
weight_blob = net.StopGradient(
|
|
[weight_blob],
|
|
[net.NextScopedBlob('weight_stop_gradient')],
|
|
)
|
|
loss = net.Mul(
|
|
[loss, weight_blob],
|
|
net.NextScopedBlob('weighted_cross_entropy'),
|
|
)
|
|
|
|
if self.average_loss:
|
|
net.AveragedLoss(loss, self.output_schema.field_blobs())
|
|
else:
|
|
net.ReduceFrontSum(loss, self.output_schema.field_blobs())
|