/
encoded.py
411 lines (344 loc) · 16.2 KB
/
encoded.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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
# 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.
"""Factory for aggregations parameterized by tensor_encoding Encoders."""
import collections
from collections.abc import Callable
import typing
import tensorflow as tf
import tree
from tensorflow_federated.python.aggregators import factory
from tensorflow_federated.python.common_libs import py_typecheck
from tensorflow_federated.python.core.environments.tensorflow_frontend import tensorflow_computation
from tensorflow_federated.python.core.impl.federated_context import federated_computation
from tensorflow_federated.python.core.impl.federated_context import intrinsics
from tensorflow_federated.python.core.impl.types import array_shape
from tensorflow_federated.python.core.impl.types import computation_types
from tensorflow_federated.python.core.impl.types import placements
from tensorflow_federated.python.core.impl.types import type_conversions
from tensorflow_federated.python.core.templates import aggregation_process
from tensorflow_federated.python.core.templates import measured_process
from tensorflow_model_optimization.python.core.internal import tensor_encoding as te
_EncoderConstructor = Callable[[tf.TensorSpec], te.core.GatherEncoder]
# TODO: b/173001046 - Enable parameterization by an inner aggregation factory.
# This will only be possible when the encoders fully commute with sum and
# requires different implementation of the core logic, not directly using
# `intrinsics.federated_aggregate`.
class EncodedSumFactory(factory.UnweightedAggregationFactory):
"""`UnweightedAggregationFactory` for encoded sum.
The created `tff.templates.AggregationProcess` aggregates values placed at
`CLIENTS` according to the provided `GatherEncoder` objects, and outputs the
result placed at `SERVER`.
A `GatherEncoder` defines encode and decode methods, which are applied by
`CLIENTS` before aggregation and by `SERVER` after aggregation. Usually these
are such that the encoded representation communicated from `CLIENTS` to
`SERVER` is smaller than communicating the original value, possibly at a cost
of communicating slightly inaccurate values. A common example is to apply
quantization. In addition, `GatherEncoder` enables the decoding logic to be
split based on which part of its decode part commutes with sum, which is
automatically utilized by the implementation.
NOTE: The current implementation does not allow for specification of an
"inner" aggregation factory. This will be possible in a future version, when
used together with a `GatherEncoder` of which the entire decode logic commutes
with sum, as per its `fully_commutes_with_sum` property. Contributions
welcome.
"""
def __init__(self, encoder_fn: _EncoderConstructor):
"""Initializes `EncodedSumFactory`.
This class is initialized with an `encoder_fn` function, which given a
`tf.TensorSpec`, returns an instance of `GatherEncoder` which is to be used
for encoding and aggregating a value matching the provided spec.
An example where this pattern is practically useful, is when encoding a
collection of values, such as all variables of a model, it is usually best
to only encode variables with larger number of elements, as those with small
number of elements are often more sensitive to inaccuracy, and can provide
only relatively small gain in terms of compression.
The `encoder_fn` will be used during the call to `create` of the factory,
and applied based on the provided `value_type`.
Args:
encoder_fn: A one-arg callable, mapping a `tf.TensorSpec`, to a
`GatherEncoder`.
"""
self._encoder_fn = encoder_fn
@classmethod
def quantize_above_threshold(
cls, quantization_bits=8, threshold=20000, **kwargs
):
"""Quantization of values with at least `threshold` elements.
Given a `value_type` in the `create` method, this classmethod configures the
`EncodedSumFactory` to apply uniform quantization to all instances of
`tff.TensorType` in the `value_type` which have more than `threshold`
elements.
Precisely, for each tensor `t`, this operation corresponds to
`t = round((t - min(t)) / (max(t) - min(t)) * (2**quantizaton_bits - 1))`.
If a type does not have more than `threshold` elements, it is summed
directly without being modified.
Args:
quantization_bits: A integer specifying the quantization bitwidth.
threshold: A non-negative integer. Only tensors with more than this number
of elements are quantized.
**kwargs: Keyword arguments.
Returns:
An `EncodedSumFactory`.
"""
_check_quantization_bits(quantization_bits)
_check_threshold(threshold)
def encoder_fn(value_spec):
if array_shape.num_elements_in_shape(value_spec.shape) > threshold:
return te.encoders.as_gather_encoder(
te.encoders.uniform_quantization(quantization_bits, **kwargs),
value_spec,
)
return te.encoders.as_gather_encoder(te.encoders.identity(), value_spec)
return cls(encoder_fn)
def create(
self, value_type: factory.ValueType
) -> aggregation_process.AggregationProcess:
type_args = typing.get_args(factory.ValueType)
py_typecheck.check_type(value_type, type_args)
encoders = self._encoders_for_value_type(value_type)
init_fn = _encoded_init_fn(encoders)
next_fn = _encoded_next_fn(
init_fn.type_signature.result, value_type, encoders
)
return aggregation_process.AggregationProcess(init_fn, next_fn)
def _encoders_for_value_type(self, value_type):
value_specs = type_conversions.type_to_tf_tensor_specs(value_type)
encoders = tf.nest.map_structure(self._encoder_fn, value_specs)
return encoders
def _encoded_init_fn(encoders):
"""Creates `init_fn` for the process returned by `EncodedSumFactory`.
The state for the `EncodedSumFactory` is directly derived from the state of
the `GatherEncoder` objects that parameterize the functionality.
Args:
encoders: A collection of `GatherEncoder` objects.
Returns:
A no-arg `tff.Computation` returning initial state for `EncodedSumFactory`.
"""
init_fn_tf = tensorflow_computation.tf_computation(
lambda: tf.nest.map_structure(lambda e: e.initial_state(), encoders)
)
init_fn = federated_computation.federated_computation(
lambda: intrinsics.federated_eval(init_fn_tf, placements.SERVER)
)
return init_fn
def _encoded_next_fn(server_state_type, value_type, encoders):
"""Creates `next_fn` for the process returned by `EncodedSumFactory`.
The structure of the implementation is roughly as follows:
* Extract params for encoding/decoding from state (`get_params_fn`).
* Encode values to be aggregated, placed at clients (`encode_fn`).
* Call `federated_aggregate` operator, with decoding of the part which does
not commute with sum, placed in its `accumulate_fn` arg.
* Finish decoding the summed value placed at server (`decode_after_sum_fn`).
* Update the state placed at server (`update_state_fn`).
Args:
server_state_type: A `tff.Type` of the expected state placed at server.
value_type: An unplaced `tff.Type` of the value to be aggregated.
encoders: A collection of `GatherEncoder` objects.
Returns:
A `tff.Computation` for `EncodedSumFactory`, with the type signature of
`(server_state_type, value_type@CLIENTS) ->
MeasuredProcessOutput(server_state_type, value_type@SERVER, ()@SERVER)`
"""
@tensorflow_computation.tf_computation(server_state_type.member)
def get_params_fn(state):
params = tree.map_structure_up_to(
encoders, lambda e, s: e.get_params(s), encoders, state
)
encode_params = _slice(encoders, params, 0)
decode_before_sum_params = _slice(encoders, params, 1)
decode_after_sum_params = _slice(encoders, params, 2)
return encode_params, decode_before_sum_params, decode_after_sum_params
encode_params_type = get_params_fn.type_signature.result[0]
decode_before_sum_params_type = get_params_fn.type_signature.result[1]
decode_after_sum_params_type = get_params_fn.type_signature.result[2]
# TODO: b/139844355 - Get rid of decode_before_sum_params.
# We pass decode_before_sum_params to the encode method, because TFF currently
# does not have a mechanism to make a tff.SERVER placed value available inside
# of intrinsics.federated_aggregate - in production, this could mean an
# intermediary aggregator node. So currently, we send the params to clients,
# and ask them to send them back as part of the encoded structure.
@tensorflow_computation.tf_computation(
value_type, encode_params_type, decode_before_sum_params_type
)
def encode_fn(x, encode_params, decode_before_sum_params):
encoded_structure = tree.map_structure_up_to(
encoders, lambda e, *args: e.encode(*args), encoders, x, encode_params
)
encoded_x = _slice(encoders, encoded_structure, 0)
state_update_tensors = _slice(encoders, encoded_structure, 1)
return encoded_x, decode_before_sum_params, state_update_tensors
state_update_tensors_type = encode_fn.type_signature.result[2]
# This is not a @tensorflow_computation.tf_computation because it will be used
# below when bulding the tensorflow_computation.tf_computations that will
# compose a intrinsics.federated_aggregate...
def decode_before_sum_tf_function(encoded_x, decode_before_sum_params):
part_decoded_x = tree.map_structure_up_to(
encoders,
lambda e, *args: e.decode_before_sum(*args),
encoders,
encoded_x,
decode_before_sum_params,
)
one = tf.constant((1,), tf.int32)
return part_decoded_x, one
# ...however, result type is needed to build the subsequent tf_computations.
@tensorflow_computation.tf_computation(encode_fn.type_signature.result[0:2])
def tmp_decode_before_sum_fn(encoded_x, decode_before_sum_params):
return decode_before_sum_tf_function(encoded_x, decode_before_sum_params)
part_decoded_x_type = tmp_decode_before_sum_fn.type_signature.result
del tmp_decode_before_sum_fn # Only needed for result type.
@tensorflow_computation.tf_computation(
part_decoded_x_type, decode_after_sum_params_type
)
def decode_after_sum_fn(summed_values, decode_after_sum_params):
part_decoded_aggregated_x, num_summands = summed_values
return tree.map_structure_up_to(
encoders,
lambda e, x, params: e.decode_after_sum(x, params, num_summands),
encoders,
part_decoded_aggregated_x,
decode_after_sum_params,
)
@tensorflow_computation.tf_computation(
server_state_type.member, state_update_tensors_type
)
def update_state_fn(state, state_update_tensors):
return tree.map_structure_up_to(
encoders,
lambda e, *args: e.update_state(*args),
encoders,
state,
state_update_tensors,
)
# Computations for intrinsics.federated_aggregate.
def _accumulator_value(values, state_update_tensors):
return collections.OrderedDict(
values=values, state_update_tensors=state_update_tensors
)
@tensorflow_computation.tf_computation
def zero_fn():
values = tf.nest.map_structure(
lambda s: tf.zeros(s.shape, s.dtype),
type_conversions.type_to_tf_tensor_specs(part_decoded_x_type),
)
state_update_tensors = tf.nest.map_structure(
lambda s: tf.zeros(s.shape, s.dtype),
type_conversions.type_to_tf_tensor_specs(state_update_tensors_type),
)
return _accumulator_value(values, state_update_tensors)
accumulator_type = zero_fn.type_signature.result
state_update_aggregation_modes = tf.nest.map_structure(
lambda e: tuple(e.state_update_aggregation_modes), encoders
)
@tensorflow_computation.tf_computation(
accumulator_type, encode_fn.type_signature.result
)
def accumulate_fn(acc, encoded_x):
value, params, state_update_tensors = encoded_x
part_decoded_value = decode_before_sum_tf_function(value, params)
new_values = tf.nest.map_structure(
tf.add, acc['values'], part_decoded_value
)
new_state_update_tensors = tf.nest.map_structure(
_accmulate_state_update_tensor,
acc['state_update_tensors'],
state_update_tensors,
state_update_aggregation_modes,
)
return _accumulator_value(new_values, new_state_update_tensors)
@tensorflow_computation.tf_computation(accumulator_type, accumulator_type)
def merge_fn(acc1, acc2):
new_values = tf.nest.map_structure(tf.add, acc1['values'], acc2['values'])
new_state_update_tensors = tf.nest.map_structure(
_accmulate_state_update_tensor,
acc1['state_update_tensors'],
acc2['state_update_tensors'],
state_update_aggregation_modes,
)
return _accumulator_value(new_values, new_state_update_tensors)
@tensorflow_computation.tf_computation(accumulator_type)
def report_fn(acc):
return acc
@federated_computation.federated_computation(
server_state_type,
computation_types.FederatedType(value_type, placements.CLIENTS),
)
def next_fn(state, value):
encode_params, decode_before_sum_params, decode_after_sum_params = (
intrinsics.federated_map(get_params_fn, state)
)
encode_params = intrinsics.federated_broadcast(encode_params)
decode_before_sum_params = intrinsics.federated_broadcast(
decode_before_sum_params
)
encoded_values = intrinsics.federated_map(
encode_fn, [value, encode_params, decode_before_sum_params]
)
aggregated_values = intrinsics.federated_aggregate(
encoded_values, zero_fn(), accumulate_fn, merge_fn, report_fn
)
decoded_values = intrinsics.federated_map(
decode_after_sum_fn, [aggregated_values.values, decode_after_sum_params]
)
updated_state = intrinsics.federated_map(
update_state_fn, [state, aggregated_values.state_update_tensors]
)
empty_metrics = intrinsics.federated_value((), placements.SERVER)
return measured_process.MeasuredProcessOutput(
state=updated_state, result=decoded_values, measurements=empty_metrics
)
return next_fn
def _slice(encoders, nested_value, idx):
"""Takes a slice of nested values.
We use a collection of encoders to encode a collection of values. When a
method of the encoder returns a tuple, e.g., encode / decode params of the
get_params method, we need to recover the matching collection of encode params
and collection of decode params. This method is a utility to achieve this.
Args:
encoders: A collection of encoders.
nested_value: A collection of indexable values of the same structure as
`encoders`.
idx: An integer. Index of the values in `nested_value` along which to take
the slice.
Returns:
A collection of values of the same structure as `encoders`.
"""
return tree.map_structure_up_to(encoders, lambda t: t[idx], nested_value)
def _accmulate_state_update_tensor(a, b, mode):
"""Accumulates state_update_tensors according to aggregation mode."""
if mode == te.core.StateAggregationMode.SUM:
return a + b
elif mode == te.core.StateAggregationMode.MIN:
return tf.minimum(a, b)
elif mode == te.core.StateAggregationMode.MAX:
return tf.maximum(a, b)
elif mode == te.core.StateAggregationMode.STACK:
raise NotImplementedError(
'StateAggregationMode.STACK is not supported yet.'
)
else:
raise ValueError('Not supported state aggregation mode: %s' % mode)
def _check_threshold(threshold):
py_typecheck.check_type(threshold, int)
if threshold < 0:
raise ValueError(
f'The threshold must be nonnegative. Provided threshold: {threshold}'
)
def _check_quantization_bits(quantization_bits):
py_typecheck.check_type(quantization_bits, int)
if not 1 <= quantization_bits <= 16:
raise ValueError(
'The quantization_bits must be in range [1, 16]. '
f'Provided quantization_bits: {quantization_bits}'
)