/
aggregation_process.py
191 lines (165 loc) · 8.42 KB
/
aggregation_process.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
# Copyright 2020, The TensorFlow Federated Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Defines a template for a stateful process that aggregates values."""
from tensorflow_federated.python.common_libs import structure
from tensorflow_federated.python.core.impl.computation import computation_base
from tensorflow_federated.python.core.impl.types import computation_types
from tensorflow_federated.python.core.impl.types import placements
from tensorflow_federated.python.core.templates import errors
from tensorflow_federated.python.core.templates import measured_process
# Index of the argument to next_fn representing value to be aggregated.
_INPUT_PARAM_INDEX = 1
class AggregationNotFederatedError(TypeError):
"""`TypeError` for aggregation functions not being federated."""
class AggregationPlacementError(TypeError):
"""`TypeError` for aggregation types not being placed as expected."""
class AggregationProcess(measured_process.MeasuredProcess):
"""A stateful process that aggregates values.
This class inherits the constraints documented by
`tff.templates.MeasuredProcess`.
A `tff.templates.AggregationProcess` is a `tff.templates.MeasuredProcess`
that formalizes the type signature of `initialize_fn` and `next_fn` for
aggregation.
Compared to the `tff.templates.MeasuredProcess`, this class requires a second
input argument, which is a value placed at `CLIENTS` and to be aggregated.
The `result` field of returned `tff.templates.MeasuredProcessOutput`,
representing the aggregate, must be placed at `SERVER` and does not
necessarily need to have type signature equal to the type signature of the
second input argument.
The intended composition pattern for `tff.templates.AggregationProcess` is
that of nesting. An aggregation will broadly consist of three logical parts:
- A pre-aggregation computation placed at `CLIENTS`.
- Actual aggregation.
- A post-aggregation computation placed at `SERVER`.
The second step can be realized by direct application of appropriate intrinsic
such as `tff.federated_sum`, or by delegation to (one or more) "inner"
aggregation processes.
Both `initialize` and `next` must be `tff.Computation`s with the following
type signatures:
- initialize: `( -> S@SERVER)`
- next: `(<S@SERVER, V@CLIENTS, *> ->
<state=S@SERVER, result=V'@SERVER, measurements=M@SERVER>)`
where `*` represents optional other arguments placed at `CLIENTS`. This can be
used for weighted aggregation, where the third parameter is the weight.
Note that while the value type to be aggregated will often be preserved
(i.e., `V == V'`), it is not required. An example is sampling-based
aggregation.
"""
def __init__(
self,
initialize_fn: computation_base.Computation,
next_fn: computation_base.Computation,
):
"""Creates a `tff.templates.AggregationProcess`.
Args:
initialize_fn: A no-arg `tff.Computation` that returns the initial state
of the aggregation process. The returned state must be a server-placed
federated value. Let the type of this state be called `S@SERVER`.
next_fn: A `tff.Computation` that represents the iterated function.
`next_fn` must accept at least two arguments, the first of which is of a
type assignable from the state type `S@SERVER` and the second of which
is client-placed data of type `V@CLIENTS`. `next_fn` must return a
`MeasuredProcessOutput` where the `state` attribute is assignable to the
first argument and the `result` is value placed at `SERVER`.
Raises:
TypeError: If `initialize_fn` and `next_fn` are not instances of
`tff.Computation`.
TemplateInitFnParamNotEmptyError: If `initialize_fn` has any input
arguments.
TemplateStateNotAssignableError: If the `state` returned by either
`initialize_fn` or `next_fn` is not assignable to the first input
argument of `next_fn`.
TemplateNotMeasuredProcessOutputError: If `next_fn` does not return a
`MeasuredProcessOutput`.
TemplateNextFnNumArgsError: If `next_fn` does not have at least two
input arguments.
AggregationNotFederatedError: If `initialize_fn` and `next_fn` are not
computations operating on federated types.
AggregationPlacementError: If the placements of `initialize_fn` and
`next_fn` are not matching the expected type signature.
"""
# Calling super class __init__ first ensures that
# next_fn.type_signature.result is a `MeasuredProcessOutput`, make our
# validation here easier as that must be true.
super().__init__(initialize_fn, next_fn, next_is_multi_arg=True)
if not isinstance(
initialize_fn.type_signature.result, computation_types.FederatedType
):
raise AggregationNotFederatedError(
'Provided `initialize_fn` must return a federated type, but found '
f'return type:\n{initialize_fn.type_signature.result}\nTip: If you '
'see a collection of federated types, try wrapping the returned '
'value in `tff.federated_zip` before returning.'
)
next_types = structure.flatten(
next_fn.type_signature.parameter
) + structure.flatten(next_fn.type_signature.result)
non_federated_types = [
t
for t in next_types
if not isinstance(t, computation_types.FederatedType)
]
if non_federated_types:
offending_types_str = '\n- '.join(str(t) for t in non_federated_types)
raise AggregationNotFederatedError(
'Provided `next_fn` must both be a *federated* computations, that '
'is, operate on `tff.FederatedType`s, but found\n'
f'next_fn with type signature:\n{next_fn.type_signature}\n'
f'The non-federated types are:\n {offending_types_str}.'
)
if initialize_fn.type_signature.result.placement != placements.SERVER:
raise AggregationPlacementError(
'The state controlled by an `AggregationProcess` must be placed at '
f'the SERVER, but found type: {initialize_fn.type_signature.result}.'
)
# Note that state of next_fn being placed at SERVER is now ensured by the
# assertions in base class which would otherwise raise
# errors.TemplateStateNotAssignableError.
next_fn_param = next_fn.type_signature.parameter
next_fn_result = next_fn.type_signature.result
if len(next_fn_param) < 2:
raise errors.TemplateNextFnNumArgsError(
'The `next_fn` must have at least two input arguments, but found '
f'the following input type: {next_fn_param}.'
)
if next_fn_param[_INPUT_PARAM_INDEX].placement != placements.CLIENTS:
raise AggregationPlacementError(
'The second input argument of `next_fn` must be placed at CLIENTS '
f'but found {next_fn_param[_INPUT_PARAM_INDEX]}.'
)
if next_fn_result.result.placement != placements.SERVER:
raise AggregationPlacementError(
'The "result" attribute of return type of `next_fn` must be placed '
f'at SERVER, but found {next_fn_result.result}.'
)
if next_fn_result.measurements.placement != placements.SERVER:
raise AggregationPlacementError(
'The "measurements" attribute of return type of `next_fn` must be '
f'placed at SERVER, but found {next_fn_result.measurements}.'
)
@property
def next(self) -> computation_base.Computation:
"""A `tff.Computation` that runs one iteration of the process.
Its first argument should always be the current state (originally produced
by the `initialize` attribute), the second argument must be the input placed
at `CLIENTS`, and the return type must be a
`tff.templates.MeasuredProcessOutput` with each field placed at `SERVER`.
Returns:
A `tff.Computation`.
"""
return super().next
@property
def is_weighted(self) -> bool:
"""True if `next` takes a third argument for weights."""
return len(self.next.type_signature.parameter) == 3 # pytype: disable=wrong-arg-types