Comparator Utils#

utils #

FUNCTION DESCRIPTION
get_layer_names

Get names of all layers in a model.
filter_layers_by_type

Filter layers in a model by their type.
match_layers_by_name

Match layers with the same names.
match_layers_by_position

Match layers by their position in the model.
match_layers_by_depth

Match layers by their relative depth in the network.
match_layers_by_type

Match all layers of a specific type between two models.

Functions#

get_layer_names(model: torch.nn.Module, include_container_modules: bool = False) -> list[str] #

Get names of all layers in a model.

PARAMETER DESCRIPTION
model

Model to inspect.

TYPE: Module

include_container_modules

If True, include container modules (Sequential, ModuleList, etc.). If False, only include leaf modules (Conv2d, Linear, etc.).

TYPE: bool, optional, by default False DEFAULT: False

RETURNS DESCRIPTION
list[str]

List of layer names.

Examples:

>>> import torch.nn as nn
>>> model = nn.Sequential(nn.Linear(10, 20), nn.ReLU(), nn.Linear(20, 5))
>>> get_layer_names(model)
['0', '1', '2']

Custom model:

>>> class Net(nn.Module):
...     def __init__(self):
...         super().__init__()
...         self.conv1 = nn.Conv2d(3, 64, 3)
...         self.fc = nn.Linear(64, 10)
...
...     def forward(self, x):
...         return self.fc(self.conv1(x).flatten(1))
>>> model = Net()
>>> get_layer_names(model)
['conv1', 'fc']
Source code in spectre/comparator/utils.py 
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
def get_layer_names(
    model: torch.nn.Module, include_container_modules: bool = False
) -> list[str]:
    """
    Get names of all layers in a model.

    Parameters
    ----------
    model : nn.Module
        Model to inspect.

    include_container_modules : bool, optional, by default False
        If True, include container modules (Sequential, ModuleList, etc.).
        If False, only include leaf modules (Conv2d, Linear, etc.).

    Returns
    -------
    list[str]
        List of layer names.

    Examples
    --------
    >>> import torch.nn as nn
    >>> model = nn.Sequential(nn.Linear(10, 20), nn.ReLU(), nn.Linear(20, 5))
    >>> get_layer_names(model)
    ['0', '1', '2']

    Custom model:

    >>> class Net(nn.Module):
    ...     def __init__(self):
    ...         super().__init__()
    ...         self.conv1 = nn.Conv2d(3, 64, 3)
    ...         self.fc = nn.Linear(64, 10)
    ...
    ...     def forward(self, x):
    ...         return self.fc(self.conv1(x).flatten(1))
    >>> model = Net()
    >>> get_layer_names(model)
    ['conv1', 'fc']
    """
    if include_container_modules:
        return [name for name, _ in model.named_modules() if name != ""]
    else:
        # Only include leaf modules (no children)
        layer_names = []
        for name, module in model.named_modules():
            if name != "" and len(list(module.children())) == 0:
                layer_names.append(name)
        return layer_names

filter_layers_by_type(model: torch.nn.Module, layer_types: list[type] | type) -> dict[str, torch.nn.Module] #

Filter layers in a model by their type.

PARAMETER DESCRIPTION
model

Model to filter.

TYPE: Module

layer_types

Layer type(s) to filter for (e.g., nn.Conv2d, nn.Linear).

TYPE: list[type] | type

RETURNS DESCRIPTION
dict[str, Module]

Dictionary mapping layer names to modules of specified type(s).

Examples:

>>> import torch.nn as nn
>>> model = nn.Sequential(nn.Conv2d(3, 64, 3), nn.ReLU(), nn.Linear(64, 10))
>>> filter_layers_by_type(model, nn.Linear)
{'2': Linear(in_features=64, out_features=10, bias=True)}

Multiple types:

>>> filter_layers_by_type(model, [nn.Conv2d, nn.Linear])
{'0': Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1)),
 '2': Linear(in_features=64, out_features=10, bias=True)}
Source code in spectre/comparator/utils.py 
 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
def filter_layers_by_type(
    model: torch.nn.Module, layer_types: list[type] | type
) -> dict[str, torch.nn.Module]:
    """
    Filter layers in a model by their type.

    Parameters
    ----------
    model : nn.Module
        Model to filter.

    layer_types : list[type] | type
        Layer type(s) to filter for (e.g., nn.Conv2d, nn.Linear).

    Returns
    -------
    dict[str, nn.Module]
        Dictionary mapping layer names to modules of specified type(s).

    Examples
    --------
    >>> import torch.nn as nn
    >>> model = nn.Sequential(nn.Conv2d(3, 64, 3), nn.ReLU(), nn.Linear(64, 10))
    >>> filter_layers_by_type(model, nn.Linear)
    {'2': Linear(in_features=64, out_features=10, bias=True)}

    Multiple types:

    >>> filter_layers_by_type(model, [nn.Conv2d, nn.Linear])
    {'0': Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1)),
     '2': Linear(in_features=64, out_features=10, bias=True)}
    """
    if not isinstance(layer_types, list):
        layer_types = [layer_types]

    filtered = OrderedDict()
    for name, module in model.named_modules():
        if name != "" and any(isinstance(module, t) for t in layer_types):
            filtered[name] = module

    return filtered

match_layers_by_name(layers1: list[str], layers2: list[str]) -> list[tuple[str, str]] #

Match layers with the same names.

PARAMETER DESCRIPTION
layers1

Layer names from first model.

TYPE: list[str]

layers2

Layer names from second model.

TYPE: list[str]

RETURNS DESCRIPTION
list[tuple[str, str]]

List of (layer1_name, layer2_name) pairs for matching names.

Examples:

>>> layers1 = ["conv1", "relu", "conv2"]
>>> layers2 = ["conv1", "bn", "conv2", "relu"]
>>> match_layers_by_name(layers1, layers2)
[('conv1', 'conv1'), ('relu', 'relu'), ('conv2', 'conv2')]
Source code in spectre/comparator/utils.py 
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
def match_layers_by_name(
    layers1: list[str], layers2: list[str]
) -> list[tuple[str, str]]:
    """
    Match layers with the same names.

    Parameters
    ----------
    layers1 : list[str]
        Layer names from first model.

    layers2 : list[str]
        Layer names from second model.

    Returns
    -------
    list[tuple[str, str]]
        List of (layer1_name, layer2_name) pairs for matching names.

    Examples
    --------
    >>> layers1 = ["conv1", "relu", "conv2"]
    >>> layers2 = ["conv1", "bn", "conv2", "relu"]
    >>> match_layers_by_name(layers1, layers2)
    [('conv1', 'conv1'), ('relu', 'relu'), ('conv2', 'conv2')]
    """
    common_names = set(layers1) & set(layers2)
    # Preserve order from layers1
    return [(name, name) for name in layers1 if name in common_names]

match_layers_by_position(layers1: list[str], layers2: list[str]) -> list[tuple[str, str]] #

Match layers by their position in the model.

PARAMETER DESCRIPTION
layers1

Layer names from first model.

TYPE: list[str]

layers2

Layer names from second model.

TYPE: list[str]

RETURNS DESCRIPTION
list[tuple[str, str]]

List of (layer1_name, layer2_name) pairs matched by position. Length is min(len(layers1), len(layers2)).

Examples:

>>> layers1 = ["layer1", "layer2", "layer3"]
>>> layers2 = ["conv1", "conv2"]
>>> match_layers_by_position(layers1, layers2)
[('layer1', 'conv1'), ('layer2', 'conv2')]
Source code in spectre/comparator/utils.py 
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
def match_layers_by_position(
    layers1: list[str], layers2: list[str]
) -> list[tuple[str, str]]:
    """
    Match layers by their position in the model.

    Parameters
    ----------
    layers1 : list[str]
        Layer names from first model.

    layers2 : list[str]
        Layer names from second model.

    Returns
    -------
    list[tuple[str, str]]
        List of (layer1_name, layer2_name) pairs matched by position.
        Length is min(len(layers1), len(layers2)).

    Examples
    --------
    >>> layers1 = ["layer1", "layer2", "layer3"]
    >>> layers2 = ["conv1", "conv2"]
    >>> match_layers_by_position(layers1, layers2)
    [('layer1', 'conv1'), ('layer2', 'conv2')]
    """
    n_pairs = min(len(layers1), len(layers2))
    return [(layers1[i], layers2[i]) for i in range(n_pairs)]

match_layers_by_depth(model1: torch.nn.Module, model2: torch.nn.Module) -> list[tuple[str, str]] #

Match layers by their relative depth in the network.

This function computes the depth of each layer (number of parent modules) and matches layers at similar relative positions in the network hierarchy.

PARAMETER DESCRIPTION
model1

First model.

TYPE: Module

model2

Second model.

TYPE: Module

RETURNS DESCRIPTION
list[tuple[str, str]]

List of (layer1_name, layer2_name) pairs matched by depth.

Examples:

>>> import torch.nn as nn
>>> model1 = nn.Sequential(
...     nn.Sequential(nn.Linear(10, 20), nn.ReLU()),
...     nn.Linear(20, 5),
... )
>>> model2 = nn.Sequential(nn.Linear(10, 15), nn.ReLU(), nn.Linear(15, 5))
>>> match_layers_by_depth(model1, model2)
[('0.0', '0'), ('0.1', '1'), ('1', '2')]
Source code in spectre/comparator/utils.py 
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
def match_layers_by_depth(
    model1: torch.nn.Module, model2: torch.nn.Module
) -> list[tuple[str, str]]:
    """
    Match layers by their relative depth in the network.

    This function computes the depth of each layer (number of parent modules)
    and matches layers at similar relative positions in the network hierarchy.

    Parameters
    ----------
    model1 : torch.nn.Module
        First model.

    model2 : torch.nn.Module
        Second model.

    Returns
    -------
    list[tuple[str, str]]
        List of (layer1_name, layer2_name) pairs matched by depth.

    Examples
    --------
    >>> import torch.nn as nn
    >>> model1 = nn.Sequential(
    ...     nn.Sequential(nn.Linear(10, 20), nn.ReLU()),
    ...     nn.Linear(20, 5),
    ... )
    >>> model2 = nn.Sequential(nn.Linear(10, 15), nn.ReLU(), nn.Linear(15, 5))
    >>> match_layers_by_depth(model1, model2)
    [('0.0', '0'), ('0.1', '1'), ('1', '2')]
    """

    def get_layer_depths(model: torch.nn.Module) -> dict[int, list[str]]:
        """Get depth (number of dots in name) for each layer."""
        depths: dict[int, list[str]] = {}
        for name in get_layer_names(model):
            depth = name.count(".")
            if depth not in depths:
                depths[depth] = []
            depths[depth].append(name)
        return depths

    depths1 = get_layer_depths(model1)
    depths2 = get_layer_depths(model2)

    # Match layers at each depth level
    pairs = []
    for depth in sorted(set(depths1.keys()) & set(depths2.keys())):
        layers_at_depth1 = depths1[depth]
        layers_at_depth2 = depths2[depth]
        n_pairs = min(len(layers_at_depth1), len(layers_at_depth2))
        for i in range(n_pairs):
            pairs.append((layers_at_depth1[i], layers_at_depth2[i]))

    return pairs

match_layers_by_type(model1: torch.nn.Module, model2: torch.nn.Module, layer_type: type | list[type] = torch.nn.Linear) -> list[tuple[str, str]] #

Match all layers of a specific type between two models.

PARAMETER DESCRIPTION
model1

First model.

TYPE: Module

model2

Second model.

TYPE: Module

layer_type

Layer type(s) to match (e.g., torch.nn.Conv2d, torch.nn.Linear).

TYPE: type | list[type], optional, by default torch.nn.Linear DEFAULT: Linear

RETURNS DESCRIPTION
list[tuple[str, str]]

List of (layer1_name, layer2_name) pairs for matching layer types. Matched by position within layers of that type.

Examples:

>>> import torch.nn as nn
>>> model1 = nn.Sequential(nn.Conv2d(3, 64, 3), nn.ReLU(), nn.Linear(64, 10))
>>> model2 = nn.Sequential(nn.Conv2d(3, 32, 3), nn.ReLU(), nn.Linear(32, 10))
>>> match_layers_by_type(model1, model2, nn.Linear)
[('2', '2')]

Multiple types:

>>> match_layers_by_type(model1, model2, [nn.Conv2d, nn.Linear])
[('0', '0'), ('2', '2')]
Source code in spectre/comparator/utils.py 
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
def match_layers_by_type(
    model1: torch.nn.Module,
    model2: torch.nn.Module,
    layer_type: type | list[type] = torch.nn.Linear,
) -> list[tuple[str, str]]:
    """
    Match all layers of a specific type between two models.

    Parameters
    ----------
    model1 : torch.nn.Module
        First model.

    model2 : torch.nn.Module
        Second model.

    layer_type : type | list[type], optional, by default torch.nn.Linear
        Layer type(s) to match (e.g., torch.nn.Conv2d, torch.nn.Linear).

    Returns
    -------
    list[tuple[str, str]]
        List of (layer1_name, layer2_name) pairs for matching layer types.
        Matched by position within layers of that type.

    Examples
    --------
    >>> import torch.nn as nn
    >>> model1 = nn.Sequential(nn.Conv2d(3, 64, 3), nn.ReLU(), nn.Linear(64, 10))
    >>> model2 = nn.Sequential(nn.Conv2d(3, 32, 3), nn.ReLU(), nn.Linear(32, 10))
    >>> match_layers_by_type(model1, model2, nn.Linear)
    [('2', '2')]

    Multiple types:

    >>> match_layers_by_type(model1, model2, [nn.Conv2d, nn.Linear])
    [('0', '0'), ('2', '2')]
    """
    layers1 = filter_layers_by_type(model1, layer_type)
    layers2 = filter_layers_by_type(model2, layer_type)

    # Match by position within layers of specified type
    keys1 = list(layers1.keys())
    keys2 = list(layers2.keys())
    n_pairs = min(len(keys1), len(keys2))

    return [(keys1[i], keys2[i]) for i in range(n_pairs)]