mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
a8695178aa
Summary:
To achive this, I modified the blob name scheme defined in a layer.
Before it was scope/fc_w and scope/fc_w_auto_0 (if there is another fc
within the same scope).
Now I change it to scope/fc/w and scope/fc_auto_0/w.
That is, we rely on the uniqueness of the scoped layer name to define
names for blobs.
I also overwrote the create_param method in LayerModelHelper to let it
use the resolved name for blobs given the sharingparameter context.
There are some details such as making the initializer more structured
that I need to finalize.
Reviewed By: kennyhorror
Differential Revision: D5435132
fbshipit-source-id: a0525f5ea0977e255dd5ea765b38913f5951d455
71 lines
2.0 KiB
Python
71 lines
2.0 KiB
Python
from __future__ import absolute_import
|
|
from __future__ import division
|
|
from __future__ import print_function
|
|
from __future__ import unicode_literals
|
|
|
|
import numpy as np
|
|
|
|
from caffe2.python import core, schema
|
|
from caffe2.python.layers.layers import ModelLayer
|
|
|
|
|
|
class MapToRange(ModelLayer):
|
|
"""
|
|
This layer aims to build a mapping from raw keys to indices within [0, max_index).
|
|
The mapping is continuously built during training. The mapping will be frozen during
|
|
evaluation and prediction. Unseen keys will be assigned to index 0.
|
|
"""
|
|
|
|
def __init__(
|
|
self, model,
|
|
input_record,
|
|
max_index,
|
|
name='map_to_range',
|
|
**kwargs
|
|
):
|
|
super(MapToRange, self).__init__(model, name, input_record, **kwargs)
|
|
|
|
assert max_index > 0
|
|
assert isinstance(input_record, schema.Scalar)
|
|
|
|
self.max_index = max_index
|
|
|
|
self.handler = self.create_param(
|
|
param_name='handler',
|
|
shape=None,
|
|
initializer=('LongIndexCreate', {'max_elements': self.max_index}),
|
|
optimizer=model.NoOptim
|
|
)
|
|
|
|
self.output_schema = schema.Struct(
|
|
('indices', schema.Scalar(
|
|
np.int64, self.get_next_blob_reference("indices")
|
|
)),
|
|
('handler', schema.Scalar(
|
|
np.void, self.handler
|
|
)),
|
|
)
|
|
|
|
def add_train_ops(self, net):
|
|
if self.input_record.field_type().base != np.int64:
|
|
keys = net.Cast(
|
|
self.input_record(),
|
|
net.NextScopedBlob("indices_before_mapping"),
|
|
to=core.DataType.INT64
|
|
)
|
|
else:
|
|
keys = self.input_record()
|
|
|
|
# Load keys into indices
|
|
indices = net.IndexGet([self.handler, keys],
|
|
self.output_schema.indices())
|
|
|
|
net.StopGradient(indices, indices)
|
|
|
|
def add_eval_ops(self, net):
|
|
net.IndexFreeze(self.handler, self.handler)
|
|
self.add_train_ops(net)
|
|
|
|
def add_ops(self, net):
|
|
self.add_eval_ops(net)
|