google-research

1
# coding=utf-8
2
# Copyright 2024 The Google Research Authors.
3
#
4
# Licensed under the Apache License, Version 2.0 (the "License");
5
# you may not use this file except in compliance with the License.
6
# You may obtain a copy of the License at
7
#
8
#     http://www.apache.org/licenses/LICENSE-2.0
9
#
10
# Unless required by applicable law or agreed to in writing, software
11
# distributed under the License is distributed on an "AS IS" BASIS,
12
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
# See the License for the specific language governing permissions and
14
# limitations under the License.
15

16
"""Dataclasses for network parameter containers."""
17

18
import dataclasses
19

20

21
@dataclasses.dataclass
22
class ConvScopeParams:
23
  """Parameters for tf.slim arg_scope for conv layers."""
24
  # Whether to use dropout for conv layers.
25
  dropout: bool = False
26

27
  # Dropout regularization strength for conv layers.
28
  dropout_keep_prob: float = 0.8
29

30
  # Whether to use batch_norm for conv layers.
31
  batch_norm: bool = True
32

33
  # Decay factor for batch_norm in conv layers.
34
  batch_norm_decay: float = 0.99
35

36
  # L2 regularization strength on conv and clf weights.
37
  l2_weight_decay: float = 0.00004
38

39

40
@dataclasses.dataclass
41
class InceptionV3FCNParams:
42
  """Parameters for configuring an InceptionV3FCN network."""
43
  # The receptive field size used by the network. Currently only two sizes are
44
  # supported: 911 and 129.
45
  receptive_field_size: int = 911
46

47
  # Prelogit dropout regularization strength.
48
  prelogit_dropout_keep_prob: float = 0.8
49

50
  # Scale number of filters in each Inception(V3) layer by this factor.
51
  # Minimum number of filters defaults to 16.
52
  depth_multiplier: float = 0.1
53

54
  # Minimum depth for the conv layers. Relevant only when depth_multiplier < 1.
55
  min_depth: int = 16
56

57
  # Stride used in inference. This stride should be a multiple of 16 as
58
  # InceptionV3 downsamples by 16 by the time it reaches its logits layer. If
59
  # set to 0, non-FCN mode is assumed and the output is squeezed from
60
  # (?, 1, 1, classes) to (?, classes).
61
  inception_fcn_stride: int = 0
62