231 lines
9.1 KiB
Python
231 lines
9.1 KiB
Python
## @package fc
|
|
# Module caffe2.python.layers.fc
|
|
|
|
|
|
|
|
|
|
|
|
from caffe2.python.helpers.arg_scope import get_current_scope
|
|
from caffe2.python import schema
|
|
from caffe2.python.layers.layers import ModelLayer
|
|
from caffe2.python.layers.sampling_trainable_mixin import SamplingTrainableMixin
|
|
import math
|
|
import numpy as np
|
|
|
|
|
|
def get_fc_predictor_version(fc_version):
|
|
assert fc_version in ["fp32", "fp16"], (
|
|
"Only support fp32 and fp16 for the fully connected layer "
|
|
"in the predictor net, the provided FC precision is {}".format(fc_version)
|
|
)
|
|
return fc_version
|
|
|
|
|
|
class FC(SamplingTrainableMixin, ModelLayer):
|
|
|
|
def __init__(self, model, input_record, output_dims, weight_init=None,
|
|
bias_init=None, weight_optim=None, bias_optim=None, name='fc',
|
|
weight_reg=None, bias_reg=None, clip_param=None,
|
|
max_fc_size=None, axis=1, transposed=False,
|
|
uniform_weight_init_scale_numerator=1.0,
|
|
**kwargs):
|
|
super(FC, self).__init__(model, name, input_record, **kwargs)
|
|
assert isinstance(input_record, schema.Scalar), (
|
|
"Incorrect input type {}".format(input_record))
|
|
assert len(input_record.field_types()[0].shape) > 0, (
|
|
"FC expects limited dimensions of the input tensor")
|
|
assert axis >= 1, "axis {} should >= 1.".format(axis)
|
|
self.axis = axis
|
|
input_dims = np.prod(input_record.field_types()[0].shape[axis - 1:])
|
|
|
|
assert input_dims > 0, (
|
|
"FC expects input dimensions > 0, got {}".format(input_dims))
|
|
|
|
self.clip_args = None
|
|
if (clip_param is not None):
|
|
assert len(clip_param) == 2, (
|
|
'clip_param must be a tuple / list '
|
|
'of length 2 and in the form of (clip_min, clip max)'
|
|
)
|
|
clip_min, clip_max = clip_param
|
|
assert clip_min is not None or clip_max is not None, (
|
|
'clip_min, and clip_max in clip_param cannot both be None'
|
|
)
|
|
assert (
|
|
(clip_min is None or clip_max is None) or clip_min < clip_max
|
|
), (
|
|
'clip_param = [clip_min, clip_max] must have clip_min < clip_max'
|
|
)
|
|
self.clip_args = {}
|
|
if clip_min is not None:
|
|
self.clip_args['min'] = clip_min
|
|
if clip_max is not None:
|
|
self.clip_args['max'] = clip_max
|
|
|
|
if uniform_weight_init_scale_numerator is None:
|
|
uniform_weight_init_scale_numerator = 1.0
|
|
|
|
scale = math.sqrt(uniform_weight_init_scale_numerator / input_dims)
|
|
weight_init = weight_init if weight_init else (
|
|
'UniformFill', {'min': -scale, 'max': scale})
|
|
bias_init = bias_init if bias_init else (
|
|
'UniformFill', {'min': -scale, 'max': scale})
|
|
|
|
self.output_dim_vec = FC.calculate_fc_output_dims(
|
|
max_fc_size, input_dims, output_dims)
|
|
|
|
self.transposed = transposed
|
|
if self.output_dim_vec is None or len(self.output_dim_vec) == 1:
|
|
weight_shape = [input_dims, output_dims] if transposed else [output_dims, input_dims]
|
|
self.w = self.create_param(param_name='w',
|
|
shape=weight_shape,
|
|
initializer=weight_init,
|
|
optimizer=weight_optim,
|
|
regularizer=weight_reg)
|
|
|
|
self.b = self.create_param(param_name='b',
|
|
shape=[output_dims, ],
|
|
initializer=bias_init,
|
|
optimizer=bias_optim,
|
|
regularizer=bias_reg)
|
|
else:
|
|
self.w_vec = []
|
|
self.b_vec = []
|
|
|
|
for idx, output_dim in enumerate(self.output_dim_vec):
|
|
weight_shape = [input_dims, output_dim] if transposed else [output_dim, input_dims]
|
|
self.w_vec.append(self.create_param(param_name='w_sub_{}'.format(idx),
|
|
shape=weight_shape,
|
|
initializer=weight_init,
|
|
optimizer=weight_optim,
|
|
regularizer=weight_reg))
|
|
|
|
self.b_vec.append(self.create_param(param_name='b_sub_{}'.format(idx),
|
|
shape=[output_dim, ],
|
|
initializer=weight_init,
|
|
optimizer=weight_optim,
|
|
regularizer=weight_reg))
|
|
if axis == 1:
|
|
output_shape = (output_dims, )
|
|
else:
|
|
output_shape = list(input_record.field_types()[0].shape)[0: axis - 1]
|
|
output_shape = tuple(output_shape + [output_dims])
|
|
|
|
self.output_schema = schema.Scalar(
|
|
(np.float32, output_shape),
|
|
self.get_next_blob_reference('output')
|
|
)
|
|
|
|
@staticmethod
|
|
def calculate_fc_output_dims(max_fc_size, input_dim, output_dim):
|
|
|
|
if not max_fc_size or max_fc_size < 0:
|
|
return None
|
|
|
|
assert max_fc_size >= input_dim, "Currently we split along the output " \
|
|
"dimension. So we need max_fc_size >= input_dim. But, max_fc_size: " \
|
|
"{}, input_dim: {}".format(max_fc_size, input_dim)
|
|
|
|
output_dim_allowed = int(np.floor(max_fc_size / input_dim))
|
|
num_fc = int(np.floor((output_dim - 1) / output_dim_allowed) + 1)
|
|
|
|
output_dim_vec = [output_dim_allowed] * (num_fc - 1)
|
|
|
|
output_dim_vec.append(output_dim - sum(output_dim_vec))
|
|
|
|
return output_dim_vec
|
|
|
|
def _insert_fc_ops(self, net, params, outputs, version):
|
|
"""
|
|
Args:
|
|
net: the caffe2 net to insert operator
|
|
params: weight and bias for FC
|
|
outputs: the output blobs
|
|
version: support fp32 and fp16 for now.
|
|
"""
|
|
if version == "fp32":
|
|
if self.transposed:
|
|
return net.FCTransposed(
|
|
self.input_record.field_blobs() + params,
|
|
outputs,
|
|
axis=self.axis,
|
|
**self.kwargs
|
|
)
|
|
else:
|
|
return net.FC(
|
|
self.input_record.field_blobs() + params,
|
|
outputs,
|
|
axis=self.axis,
|
|
**self.kwargs
|
|
)
|
|
elif version == "fp16":
|
|
return net.FbFCPacked(
|
|
self.input_record.field_blobs() + params,
|
|
outputs,
|
|
axis=self.axis,
|
|
**self.kwargs
|
|
)
|
|
else:
|
|
raise Exception("unsupported FC type version {}".format(version))
|
|
|
|
def _add_ops(self, net, params, version):
|
|
"""
|
|
Args:
|
|
params : the weight and bias,
|
|
passed by either add_ops or add_train_ops function
|
|
version : fp16 or fp32, might support in8 in the future.
|
|
"""
|
|
if self.clip_args is not None:
|
|
clipped_params = [net.NextScopedBlob(
|
|
'clipped_%s' % str(p)) for p in params]
|
|
for p, cp in zip(params, clipped_params):
|
|
net.Clip([p], [cp], **self.clip_args)
|
|
params = clipped_params
|
|
|
|
if self.output_dim_vec is None or len(self.output_dim_vec) == 1:
|
|
self._insert_fc_ops(net, params, self.output_schema.field_blobs(), version)
|
|
else:
|
|
w_vec = params[:int(len(params) / 2)]
|
|
b_vec = params[int(len(params) / 2):]
|
|
|
|
assert len(w_vec) == len(b_vec)
|
|
|
|
output_blob_vec = []
|
|
|
|
for i in range(len(self.output_dim_vec)):
|
|
output_blob = net.NextScopedBlob(
|
|
'output_sub_{}'.format(i))
|
|
insert_ret = self._insert_fc_ops(
|
|
net, [w_vec[i], b_vec[i]], [output_blob], version
|
|
)
|
|
output_blob_vec.append(insert_ret)
|
|
net.Concat(output_blob_vec,
|
|
self.output_schema.field_blobs() +
|
|
[self.output_schema.field_blobs()[0] + "_concat_dims"])
|
|
|
|
def add_ops(self, net):
|
|
"""Both the predict net and the eval net will call this function
|
|
"""
|
|
version_info = get_current_scope().get(
|
|
get_fc_predictor_version.__name__, {'fc_version': 'fp32'}
|
|
)
|
|
predictor_fc_fp_version = version_info['fc_version']
|
|
self._add_ops(net, self.param_blobs, predictor_fc_fp_version)
|
|
|
|
def add_train_ops(self, net):
|
|
# use the train_param_blobs to be consistent with the SamplingTrain unittest
|
|
self._add_ops(net, self.train_param_blobs, "fp32")
|
|
|
|
def get_fp16_compatible_parameters(self):
|
|
if self.output_dim_vec is None or len(self.output_dim_vec) == 1:
|
|
return [self.w]
|
|
else:
|
|
return self.w_vec
|
|
|
|
@property
|
|
def param_blobs(self):
|
|
if self.output_dim_vec is None or len(self.output_dim_vec) == 1:
|
|
return [self.w, self.b]
|
|
else:
|
|
return self.w_vec + self.b_vec
|