batch_distill_lr_loss.py

## @package batch_distill_lr_loss
# Module caffe2.python.layers.batch_distill_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 BatchDistillLRLoss(ModelLayer):

    def __init__(
            self, model, input_record,
            name='batch_distill_lr_loss', teacher_weight=0.0,
            filter_invalid_teacher_label=False, **kwargs):

        super(BatchDistillLRLoss, self).__init__(model, name, input_record, **kwargs)

        assert teacher_weight >= 0 and teacher_weight <= 1, (
            'teacher_weight=%0.2f should be in [0, 1]' % teacher_weight
        )

        self._teacher_weight = teacher_weight
        self._filter_invalid_teacher_label = filter_invalid_teacher_label
        # hyper-parameter determines whether to filter out bad teacehr labels,
        # i.e., teacher labels that are zero.
        if self._filter_invalid_teacher_label:
            self.threshold = model.add_global_constant(
                str(model.net.NextScopedBlob('threshold')),
                [0.0],   # threshold for filtering teacher weight.
                dtype=np.float
            )
            self.neg_ONE = model.add_global_constant(
                str(model.net.NextScopedBlob('neg_ONE')),
                [-1.0],
                dtype=np.float
            )
            self.ONE = model._GetOne()
        assert schema.is_schema_subset(
            schema.Struct(
                ('teacher_label', schema.Scalar()),
                ('label', schema.Scalar()),
                ('logit', schema.Scalar()),
            ),
            input_record
        )
        self.tags.update([Tags.EXCLUDE_FROM_PREDICTION])

        self.output_schema = schema.Scalar(
            np.float32,
            self.get_next_blob_reference('output')
        )

    def add_ops(self, net):
        label = self.input_record.label()
        if self.input_record.label.field_type() != np.float32:
            label = net.Cast(
                label,
                net.NextScopedBlob('float_label'),
                to=core.DataType.FLOAT,
            )

        # Assuming 1-D input
        label = net.ExpandDims(label, net.NextScopedBlob('expanded_label'),
                               dims=[1])

        teacher_label = self.input_record.teacher_label()

        if self.input_record.teacher_label.field_type() != np.float32:
            teacher_label = net.Cast(
                teacher_label,
                net.NextScopedBlob('float_teacher_label'),
                to=core.DataType.FLOAT,
            )
        teacher_label = net.ExpandDims(
            teacher_label, net.NextScopedBlob('expanded_teacher_label'),
            dims=[1])

        true_xent = net.SigmoidCrossEntropyWithLogits(
            [self.input_record.logit(), label],
            net.NextScopedBlob('cross_entropy')
        )

        teacher_xent = net.SigmoidCrossEntropyWithLogits(
            [self.input_record.logit(), teacher_label],
            net.NextScopedBlob('teacher_cross_entropy')
        )
        if self._filter_invalid_teacher_label:
            squeezed_teacher_label = net.Squeeze(
                teacher_label,
                net.NextScopedBlob('squeezed_teacher_label'),
                dims=[1]
            )
            # blob used to contain the original teacher weights
            keep_weights = net.ConstantFill(
                [squeezed_teacher_label],
                net.NextScopedBlob('keep_weights'),
                value=self._teacher_weight,
                dtype=core.DataType.FLOAT
            )
            #blob used to zero out the teacher weights
            zero_weights = net.ConstantFill(
                [squeezed_teacher_label],
                net.NextScopedBlob('zero_weights'),
                value=0.0,
                dtype=core.DataType.FLOAT
            )

            #Indicating which teacher labels are bad, i.e., are zero.
            judge = net.GT(
                [squeezed_teacher_label, self.threshold],
                net.NextScopedBlob('judge'),
                broadcast=1
            )
            #zero out bad teacher weights corresponding to bad teacher labels.
            screened_teacher_weights = net.Conditional(
                [judge, keep_weights, zero_weights],
                net.NextScopedBlob('screened_teacher_weights')
            )
            neg_screened_teacher_weights = net.Mul(
                [screened_teacher_weights, self.neg_ONE],
                net.NextScopedBlob('neg_screened_teacher_weights'),
                broadcast=1
            )
            one_minus_screened_teacher_weights = net.Add(
                [neg_screened_teacher_weights, self.ONE],
                net.NextScopedBlob('one_minus_screened_teacher_weights'),
                broadcast=1
            )
            scaled_true_xent = net.Mul(
                [true_xent, one_minus_screened_teacher_weights],
                net.NextScopedBlob('scaled_cross_entropy'),
                broadcast=1
            )
            scaled_teacher_xent = net.Mul(
                [teacher_xent, screened_teacher_weights],
                net.NextScopedBlob('scaled_teacher_cross_entropy'),
                broadcast=1
            )
        else:
            scaled_true_xent = net.Scale(
                true_xent,
                net.NextScopedBlob('scaled_cross_entropy'),
                scale=float(1.0 - self._teacher_weight),
            )
            scaled_teacher_xent = net.Scale(
                teacher_xent,
                net.NextScopedBlob('scaled_teacher_cross_entropy'),
                scale=float(self._teacher_weight),
            )
        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')],
            )
            scaled_true_xent = net.Mul(
                [scaled_true_xent, weight_blob],
                net.NextScopedBlob('weighted_xent_label'),
            )
            scaled_teacher_xent = net.Mul(
                [scaled_teacher_xent, weight_blob],
                net.NextScopedBlob('weighted_xent_teacher'),
            )

        true_loss = net.AveragedLoss(
            scaled_true_xent,
            net.NextScopedBlob('true_loss')
        )
        teacher_loss = net.AveragedLoss(
            scaled_teacher_xent,
            net.NextScopedBlob('teacher_loss')
        )
        net.Add(
            [true_loss, teacher_loss],
            self.output_schema.field_blobs()
        )