-
Notifications
You must be signed in to change notification settings - Fork 276
/
view_types.py
391 lines (328 loc) · 13.8 KB
/
view_types.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
# Copyright 2020 Google LLC
#
# 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
#
# https://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.
"""View types for Tensorflow Model Analysis."""
import copy
from typing import Any, Dict, List, Sequence, NamedTuple, Optional, Union
from tensorflow_model_analysis import constants
from tensorflow_model_analysis.metrics import metric_types
from tensorflow_model_analysis.proto import config_pb2
from tensorflow_model_analysis.proto import metrics_for_slice_pb2
from tensorflow_model_analysis.slicer import slicer_lib as slicer
Plots = Any
PlotsBySubKey = Dict[str, Plots]
PlotsByOutputName = Dict[str, PlotsBySubKey]
class SlicedPlots(
NamedTuple('SlicedPlots', [('slice', slicer.SliceKeyType),
('plot', PlotsByOutputName)])):
"""A tuple containing the plots belonging to a slice.
Attributes:
slice: A 2-element tuple representing a slice. The first element is the key
of a feature (ex: 'color'), and the second element is the value (ex:
'green'). An empty tuple represents an 'overall' slice (i.e. one that
encompasses the entire dataset.
plot: A dict mapping `output_name` and `sub_key_id` to plot data. The data
contains histograms and confusion matrices, which can be rendered with the
`tfma.view.render_plot` function.
"""
MetricsByTextKey = Dict[str, Dict[str, Any]]
MetricsBySubKey = Dict[str, MetricsByTextKey]
MetricsByOutputName = Dict[str, MetricsBySubKey]
class SlicedMetrics(
NamedTuple('SlicedMetrics', [('slice', slicer.SliceKeyType),
('metrics', MetricsByOutputName)])):
"""A tuple containing the metrics belonging to a slice.
The metrics are stored in a nested dictionary with the following levels:
1. output_name: Optional output name associated with metric (for multi-output
models). '' by default.
2. sub_key: Optional sub key associated with metric (for multi-class models).
'' by default. See `tfma.metrics.SubKey` for more info.
3. metric_name: Name of the metric (`auc`, `accuracy`, etc).
4. metric_value: A dictionary containing the metric's value. See
[`tfma.proto.metrics_for_slice_pb2.MetricValue`](https://github.com/tensorflow/model-analysis/blob/cdb6790dcd7a37c82afb493859b3ef4898963fee/tensorflow_model_analysis/proto/metrics_for_slice.proto#L194)
for more info.
Below is a sample SlicedMetrics:
```python
(
(('color', 'green')),
{
'': { # default for single-output models
'': { # default sub_key for non-multiclass-classification models
'auc': {
'doubleValue': 0.7243943810462952
},
'accuracy': {
'doubleValue': 0.6488351225852966
}
}
}
}
)
```
Attributes:
slice: A 2-element tuple representing a slice. The first element is the key
of a feature (ex: 'color'), and the second element is the value (ex:
'green'). An empty tuple represents an 'overall' slice (i.e. one that
encompasses the entire dataset.
metrics: A nested dictionary containing metric names and values.
"""
AttributionsByFeatureKey = Dict[str, metrics_for_slice_pb2.MetricValue]
AttributionsByMetricName = Dict[str, AttributionsByFeatureKey]
AttributionsBySubKey = Dict[str, AttributionsByMetricName]
AttributionsByOutputName = Dict[str, AttributionsBySubKey]
class SlicedAttributions(
NamedTuple('SlicedAttributions',
[('slice', slicer.SliceKeyType),
('attributions', AttributionsByOutputName)])):
"""A tuple containing the attributions belonging to a slice.
The attributions are stored in a nested dictionary with the following levels:
1. output_name: Optional output name associated with metric (for multi-output
models). '' by default.
2. sub_key: Optional sub key associated with metric (for multi-class models).
'' by default. See `tfma.metrics.SubKey` for more info.
3. metric_name: Name of the metric (`auc`, `accuracy`, etc).
4. feature_key: Key of feature ('age', etc).
5. metric_value: A dictionary containing the metric's value. See
[tfma.proto.metrics_for_slice_pb2.MetricValue](https://github.com/tensorflow/model-analysis/blob/cdb6790dcd7a37c82afb493859b3ef4898963fee/tensorflow_model_analysis/proto/metrics_for_slice.proto#L194)
for more info.
Below is a sample SlicedAttributions:
```python
(
(('color', 'green')),
{
'': { # default for single-output models
'': { # default sub_key for non-multiclass-classification models
'total_attributions': {
'feature1': {
'doubleValue': 100.32
},
'feature2': {
'doubleValue': 54.2
}
}
}
}
}
)
```
Attributes:
slice: A 2-element tuple representing a slice. The first element is the key
of a feature (ex: 'color'), and the second element is the value (ex:
'green'). An empty tuple represents an 'overall' slice (i.e. one that
encompasses the entire dataset.
attributions: A nested dictionary containing attribution names and values.
"""
class EvalResult(
NamedTuple('EvalResult', [('slicing_metrics', List[SlicedMetrics]),
('plots', List[SlicedPlots]),
('attributions', List[SlicedAttributions]),
('config', config_pb2.EvalConfig),
('data_location', str), ('file_format', str),
('model_location', str)])):
"""The result of a single model analysis run.
Attributes:
slicing_metrics: a list of `tfma.SlicedMetrics`, containing metric values
for each slice.
plots: List of slice-plot pairs.
attributions: List of SlicedAttributions containing attribution values for
each slice.
config: The config containing slicing and metrics specification.
data_location: Optional location for data used with config.
file_format: Optional format for data used with config.
model_location: Optional location(s) for model(s) used with config.
"""
def get_metrics_for_slice(
self,
slice_name: slicer.SliceKeyType = (),
output_name: str = '',
class_id: Optional[int] = None,
k: Optional[int] = None,
top_k: Optional[int] = None) -> Union[MetricsByTextKey, None]:
"""Get metric names and values for a slice.
Args:
slice_name: A tuple of the form (column, value), indicating which slice to
get metrics from. Optional; if excluded, return overall metrics.
output_name: The name of the output. Optional, only used for multi-output
models.
class_id: Used with multi-class metrics to identify a specific class ID.
k: Used with multi-class metrics to identify the kth predicted value.
top_k: Used with multi-class and ranking metrics to identify top-k
predicted values.
Returns:
Dictionary containing metric names and values for the specified slice.
"""
if all(v is None for v in [class_id, k, top_k]):
sub_key = ''
else:
sub_key = str(metric_types.SubKey(class_id, k, top_k))
def equals_slice_name(slice_key):
if not slice_key:
return not slice_name
else:
return slice_key == slice_name
for slicing_metric in self.slicing_metrics:
slice_key = slicing_metric[0]
slice_val = slicing_metric[1]
if equals_slice_name(slice_key):
return slice_val[output_name][sub_key]
# if slice could not be found, return None
return None
def get_metrics_for_all_slices(
self,
output_name: str = '',
class_id: Optional[int] = None,
k: Optional[int] = None,
top_k: Optional[int] = None) -> Dict[str, MetricsByTextKey]:
"""Get metric names and values for every slice.
Args:
output_name: The name of the output (optional, only used for multi-output
models).
class_id: Used with multi-class metrics to identify a specific class ID.
k: Used with multi-class metrics to identify the kth predicted value.
top_k: Used with multi-class and ranking metrics to identify top-k
predicted values.
Returns:
Dictionary mapping slices to metric names and values.
"""
if all(v is None for v in [class_id, k, top_k]):
sub_key = ''
else:
sub_key = str(metric_types.SubKey(class_id, k, top_k))
sliced_metrics = {}
for slicing_metric in self.slicing_metrics:
slice_name = slicing_metric[0]
metrics = slicing_metric[1][output_name][sub_key]
sliced_metrics[slice_name] = {
metric_name: metric_value
for metric_name, metric_value in metrics.items()
}
return sliced_metrics # pytype: disable=bad-return-type
def get_metric_names(self) -> Sequence[str]:
"""Get names of metrics.
Returns:
List of metric names.
"""
metric_names = set()
for slicing_metric in self.slicing_metrics:
for output_name in slicing_metric[1]:
for metrics in slicing_metric[1][output_name].values():
metric_names.update(metrics)
return list(metric_names)
def get_attributions_for_slice(
self,
slice_name: slicer.SliceKeyType = (),
metric_name: str = '',
output_name: str = '',
class_id: Optional[int] = None,
k: Optional[int] = None,
top_k: Optional[int] = None) -> Union[AttributionsByFeatureKey, None]:
"""Get attribution features names and values for a slice.
Args:
slice_name: A tuple of the form (column, value), indicating which slice to
get attributions from. Optional; if excluded, use overall slice.
metric_name: Name of metric to get attributions for. Optional if only one
metric used.
output_name: The name of the output. Optional, only used for multi-output
models.
class_id: Used with multi-class models to identify a specific class ID.
k: Used with multi-class models to identify the kth predicted value.
top_k: Used with multi-class models to identify top-k attribution values.
Returns:
Dictionary containing feature keys and values for the specified slice.
Raises:
ValueError: If metric_name is required.
"""
if class_id or k or top_k:
sub_key = str(metric_types.SubKey(class_id, k, top_k))
else:
sub_key = ''
def equals_slice_name(slice_key):
if not slice_key:
return not slice_name
else:
return slice_key == slice_name
for sliced_attributions in self.attributions:
slice_key = sliced_attributions[0]
slice_val = sliced_attributions[1]
if equals_slice_name(slice_key):
if metric_name:
return slice_val[output_name][sub_key][metric_name]
elif len(slice_val[output_name][sub_key]) == 1:
return list(slice_val[output_name][sub_key].values())[0]
else:
raise ValueError(
'metric_name must be one of the following: {}'.format(
slice_val[output_name][sub_key].keys()))
# if slice could not be found, return None
return None
def get_attributions_for_all_slices(
self,
metric_name: str = '',
output_name: str = '',
class_id: Optional[int] = None,
k: Optional[int] = None,
top_k: Optional[int] = None) -> Dict[str, AttributionsByFeatureKey]:
"""Get attribution feature keys and values for every slice.
Args:
metric_name: Name of metric to get attributions for. Optional if only one
metric used.
output_name: The name of the output (optional, only used for multi-output
models).
class_id: Used with multi-class metrics to identify a specific class ID.
k: Used with multi-class metrics to identify the kth predicted value.
top_k: Used with multi-class and ranking metrics to identify top-k
predicted values.
Returns:
Dictionary mapping slices to attribution feature keys and values.
"""
if class_id or k or top_k:
sub_key = str(metric_types.SubKey(class_id, k, top_k))
else:
sub_key = ''
all_sliced_attributions = {}
for sliced_attributions in self.attributions:
slice_name = sliced_attributions[0]
attributions = sliced_attributions[1][output_name][sub_key]
if metric_name:
attributions = attributions[metric_name]
elif len(attributions) == 1:
attributions = list(attributions.values())[0]
else:
raise ValueError('metric_name must be one of the following: {}'.format(
attributions.keys()))
all_sliced_attributions[slice_name] = copy.copy(attributions)
return all_sliced_attributions # pytype: disable=bad-return-type
def get_slice_names(self) -> Sequence[str]:
"""Get names of slices.
Returns:
List of slice names.
"""
return [slicing_metric[0] for slicing_metric in self.slicing_metrics] # pytype: disable=bad-return-type
class EvalResults:
"""The results from multiple TFMA runs, or a TFMA run on multiple models."""
def __init__(self,
results: List[EvalResult],
mode: str = constants.UNKNOWN_EVAL_MODE):
supported_modes = [
constants.DATA_CENTRIC_MODE,
constants.MODEL_CENTRIC_MODE,
]
if mode not in supported_modes:
raise ValueError('Mode ' + mode + ' must be one of ' +
str(supported_modes))
self._results = results
self._mode = mode
def get_results(self) -> List[EvalResult]:
return self._results
def get_mode(self) -> str:
return self._mode