| |
| |
|
|
|
|
|
|
|
|
|
|
| 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 BatchMSELoss(ModelLayer): |
|
|
| def __init__(self, model, input_record, name='batch_mse_loss', **kwargs): |
| super(BatchMSELoss, self).__init__(model, name, input_record, **kwargs) |
|
|
| assert schema.is_schema_subset( |
| schema.Struct( |
| ('label', schema.Scalar()), |
| ('prediction', 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): |
| prediction = self.input_record.prediction() |
| label = self.input_record.label.field_blobs() |
| if self.input_record.label.field_type().base != ( |
| self.input_record.prediction.field_type().base): |
|
|
| label = net.Cast( |
| label, |
| net.NextScopedBlob('cast_label'), |
| to=schema.data_type_for_dtype( |
| self.input_record.prediction.field_type() |
| ) |
| ) |
|
|
| label = net.ExpandDims(label, 1, dims=[1]) |
|
|
| label = net.StopGradient( |
| label, |
| net.NextScopedBlob('stopped_label') |
| ) |
|
|
| l2dist = net.SquaredL2Distance( |
| [label, prediction], |
| net.NextScopedBlob('l2') |
| ) |
|
|
| 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')], |
| ) |
| l2dist = net.Mul( |
| [l2dist, weight_blob], |
| net.NextScopedBlob('weighted_l2_distance'), |
| ) |
|
|
| net.AveragedLoss(l2dist, self.output_schema.field_blobs()) |
|
|
