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
|
import numpy
import pytest
from thinc import registry
from thinc.api import (
CategoricalCrossentropy,
CosineDistance,
L2Distance,
SequenceCategoricalCrossentropy,
)
# some simple arrays
scores0 = numpy.zeros((3, 3), dtype="f")
labels0 = numpy.asarray([0, 1, 1], dtype="i")
# a few more diverse ones to test realistic values
guesses1 = numpy.asarray([[0.1, 0.5, 0.6], [0.4, 0.6, 0.3], [1, 1, 1], [0, 0, 0]])
labels1 = numpy.asarray([2, 1, 0, 2])
labels1_full = numpy.asarray([[0, 0, 1], [0, 1, 0], [1, 0, 0], [0, 0, 1]])
labels1_strings = ["C", "B", "A", "C"]
guesses2 = numpy.asarray([[0.2, 0.3, 0.0]])
labels2 = numpy.asarray([1])
labels2_strings = ["B"]
eps = 0.0001
def test_loss():
d_scores = CategoricalCrossentropy().get_grad(scores0, labels0)
assert d_scores.dtype == "float32"
assert d_scores.shape == scores0.shape
d_scores = SequenceCategoricalCrossentropy().get_grad([scores0], [labels0])
assert d_scores[0].dtype == "float32"
assert d_scores[0].shape == scores0.shape
assert SequenceCategoricalCrossentropy().get_grad([], []) == []
@pytest.mark.parametrize(
"dist", [CategoricalCrossentropy(), CosineDistance(ignore_zeros=True), L2Distance()]
)
@pytest.mark.parametrize("vect", [scores0, guesses1, guesses2])
def test_equality(dist, vect):
assert int(dist.get_grad(vect, vect)[0][0]) == pytest.approx(0, eps)
assert dist.get_loss(vect, vect) == pytest.approx(0, eps)
@pytest.mark.parametrize(
"guesses, labels", [(guesses1, labels1), (guesses1, labels1_full)]
)
def test_categorical_crossentropy(guesses, labels):
d_scores = CategoricalCrossentropy(normalize=True).get_grad(guesses, labels)
assert d_scores.shape == guesses.shape
# The normalization divides the difference (e.g. 0.4) by the number of vectors (4)
assert d_scores[1][0] == pytest.approx(0.1, eps)
assert d_scores[1][1] == pytest.approx(-0.1, eps)
# The third vector predicted all labels, but only the first one was correct
assert d_scores[2][0] == pytest.approx(0, eps)
assert d_scores[2][1] == pytest.approx(0.25, eps)
assert d_scores[2][2] == pytest.approx(0.25, eps)
# The fourth vector predicted no labels but should have predicted the last one
assert d_scores[3][0] == pytest.approx(0, eps)
assert d_scores[3][1] == pytest.approx(0, eps)
assert d_scores[3][2] == pytest.approx(-0.25, eps)
loss = CategoricalCrossentropy(normalize=True).get_loss(guesses, labels)
assert loss == pytest.approx(0.239375, eps)
def test_crossentropy_incorrect_scores_targets():
labels = numpy.asarray([2])
guesses_neg = numpy.asarray([[-0.1, 0.5, 0.6]])
with pytest.raises(ValueError, match=r"Cannot calculate.*guesses"):
CategoricalCrossentropy(normalize=True).get_grad(guesses_neg, labels)
guesses_larger_than_one = numpy.asarray([[1.1, 0.5, 0.6]])
with pytest.raises(ValueError, match=r"Cannot calculate.*guesses"):
CategoricalCrossentropy(normalize=True).get_grad(
guesses_larger_than_one, labels
)
guesses_ok = numpy.asarray([[0.1, 0.4, 0.5]])
targets_neg = numpy.asarray([[-0.1, 0.5, 0.6]])
with pytest.raises(ValueError, match=r"Cannot calculate.*truth"):
CategoricalCrossentropy(normalize=True).get_grad(guesses_ok, targets_neg)
targets_larger_than_one = numpy.asarray([[2.0, 0.5, 0.6]])
with pytest.raises(ValueError, match=r"Cannot calculate.*truth"):
CategoricalCrossentropy(normalize=True).get_grad(
guesses_ok, targets_larger_than_one
)
@pytest.mark.parametrize(
"guesses, labels",
[(guesses1, [2, 1, 0, 2])],
)
def test_categorical_crossentropy_int_list_missing(guesses, labels):
d_scores = CategoricalCrossentropy(normalize=True, missing_value=0).get_grad(
guesses, labels
)
assert d_scores.shape == guesses.shape
# The normalization divides the difference (e.g. 0.4) by the number of vectors (4)
assert d_scores[1][0] == pytest.approx(0.1, eps)
assert d_scores[1][1] == pytest.approx(-0.1, eps)
# Label 0 is masked, because it represents the missing value
assert d_scores[2][0] == 0.0
assert d_scores[2][1] == 0.0
assert d_scores[2][2] == 0.0
# The fourth vector predicted no labels but should have predicted the last one
assert d_scores[3][0] == pytest.approx(0, eps)
assert d_scores[3][1] == pytest.approx(0, eps)
assert d_scores[3][2] == pytest.approx(-0.25, eps)
loss = CategoricalCrossentropy(normalize=True, missing_value=0).get_loss(
guesses, labels
)
assert loss == pytest.approx(0.114375, eps)
@pytest.mark.parametrize(
"guesses, labels", [(guesses1, labels1), (guesses1, labels1_full)]
)
def test_categorical_crossentropy_missing(guesses, labels):
d_scores = CategoricalCrossentropy(normalize=True, missing_value=0).get_grad(
guesses, labels
)
assert d_scores.shape == guesses.shape
# The normalization divides the difference (e.g. 0.4) by the number of vectors (4)
assert d_scores[1][0] == pytest.approx(0.1, eps)
assert d_scores[1][1] == pytest.approx(-0.1, eps)
# Label 0 is masked, because it represents the missing value
assert d_scores[2][0] == 0.0
assert d_scores[2][1] == 0.0
assert d_scores[2][2] == 0.0
# The fourth vector predicted no labels but should have predicted the last one
assert d_scores[3][0] == pytest.approx(0, eps)
assert d_scores[3][1] == pytest.approx(0, eps)
assert d_scores[3][2] == pytest.approx(-0.25, eps)
loss = CategoricalCrossentropy(normalize=True, missing_value=0).get_loss(
guesses, labels
)
assert loss == pytest.approx(0.114375, eps)
@pytest.mark.parametrize(
"guesses, labels, names",
[
([guesses1, guesses2], [labels1, labels2], []),
([guesses1, guesses2], [labels1_full, labels2], []),
([guesses1, guesses2], [labels1_strings, labels2_strings], ["A", "B", "C"]),
],
)
def test_sequence_categorical_crossentropy(guesses, labels, names):
d_scores = SequenceCategoricalCrossentropy(normalize=False, names=names).get_grad(
guesses, labels
)
d_scores1 = d_scores[0]
d_scores2 = d_scores[1]
assert d_scores1.shape == guesses1.shape
assert d_scores2.shape == guesses2.shape
assert d_scores1[1][0] == pytest.approx(0.4, eps)
assert d_scores1[1][1] == pytest.approx(-0.4, eps)
# The normalization divides the difference (e.g. 0.4) by the number of seqs
d_scores = SequenceCategoricalCrossentropy(normalize=True, names=names).get_grad(
guesses, labels
)
d_scores1 = d_scores[0]
d_scores2 = d_scores[1]
assert d_scores1[1][0] == pytest.approx(0.2, eps)
assert d_scores1[1][1] == pytest.approx(-0.2, eps)
# The third vector predicted all labels, but only the first one was correct
assert d_scores1[2][0] == pytest.approx(0, eps)
assert d_scores1[2][1] == pytest.approx(0.5, eps)
assert d_scores1[2][2] == pytest.approx(0.5, eps)
# The fourth vector predicted no labels but should have predicted the last one
assert d_scores1[3][0] == pytest.approx(0, eps)
assert d_scores1[3][1] == pytest.approx(0, eps)
assert d_scores1[3][2] == pytest.approx(-0.5, eps)
# Test the second batch
assert d_scores2[0][0] == pytest.approx(0.1, eps)
assert d_scores2[0][1] == pytest.approx(-0.35, eps)
loss = SequenceCategoricalCrossentropy(normalize=True, names=names).get_loss(
guesses, labels
)
assert loss == pytest.approx(1.09, eps)
@pytest.mark.parametrize(
"guesses, labels, names",
[
([guesses1], [["A", "!A", "", "!C"]], ["A", "B", "C"]),
],
)
def test_sequence_categorical_missing_negative(guesses, labels, names):
d_scores = SequenceCategoricalCrossentropy(
normalize=False, names=names, neg_prefix="!", missing_value=""
).get_grad(guesses, labels)
d_scores0 = d_scores[0]
# [0.1, 0.5, 0.6] should be A
assert d_scores0[0][0] == pytest.approx(-0.9, eps)
assert d_scores0[0][1] == pytest.approx(0.5, eps)
assert d_scores0[0][2] == pytest.approx(0.6, eps)
# [0.4, 0.6, 0.3] should NOT be A
assert d_scores0[1][0] == pytest.approx(0.4, eps)
assert d_scores0[1][1] == pytest.approx(0.0, eps)
assert d_scores0[1][2] == pytest.approx(0.0, eps)
# [1, 1, 1] has missing gold label
assert d_scores0[2][0] == pytest.approx(0.0, eps)
assert d_scores0[2][1] == pytest.approx(0.0, eps)
assert d_scores0[2][2] == pytest.approx(0.0, eps)
# [0.0, 0.0, 0.0] should NOT be C
assert d_scores0[3][0] == pytest.approx(0.0, eps)
assert d_scores0[3][1] == pytest.approx(0.0, eps)
assert d_scores0[3][2] == pytest.approx(0.0, eps)
def test_L2():
# L2 loss = 2²+4²=20 (or normalized: 1²+2²=5)
vec1 = numpy.asarray([[1, 2], [8, 9]])
vec2 = numpy.asarray([[1, 2], [10, 5]])
d_vecs = L2Distance().get_grad(vec1, vec2)
assert d_vecs.shape == vec1.shape
numpy.testing.assert_allclose(
d_vecs[0], numpy.zeros(d_vecs[0].shape), rtol=eps, atol=eps
)
loss_not_normalized = L2Distance(normalize=False).get_loss(vec1, vec2)
assert loss_not_normalized == pytest.approx(20, eps)
loss_normalized = L2Distance(normalize=True).get_loss(vec1, vec2)
assert loss_normalized == pytest.approx(5, eps)
def test_cosine_orthogonal():
# These are orthogonal, i.e. loss is 1
vec1 = numpy.asarray([[0, 2], [0, 5]])
vec2 = numpy.asarray([[8, 0], [7, 0]])
d_vecs = CosineDistance(normalize=True).get_grad(vec1, vec2)
assert d_vecs.shape == vec1.shape
assert d_vecs[0][0] < 0
assert d_vecs[0][1] > 0
assert d_vecs[1][0] < 0
assert d_vecs[1][1] > 0
loss_not_normalized = CosineDistance(normalize=False).get_loss(vec1, vec2)
assert loss_not_normalized == pytest.approx(2, eps)
loss_normalized = CosineDistance(normalize=True).get_loss(vec1, vec2)
assert loss_normalized == pytest.approx(1, eps)
def test_cosine_equal():
# These 3 vectors are equal when measured with Cosine similarity, i.e. loss is 0
vec1 = numpy.asarray([[1, 2], [8, 9], [3, 3]])
vec2 = numpy.asarray([[1, 2], [80, 90], [300, 300]])
d_vec1 = CosineDistance().get_grad(vec1, vec2)
assert d_vec1.shape == vec1.shape
numpy.testing.assert_allclose(d_vec1, numpy.zeros(d_vec1.shape), rtol=eps, atol=eps)
loss_not_normalized = CosineDistance(normalize=False).get_loss(vec1, vec2)
assert loss_not_normalized == pytest.approx(0, eps)
loss_normalized = CosineDistance(normalize=True).get_loss(vec1, vec2)
assert loss_normalized == pytest.approx(0, eps)
def test_cosine_unmatched():
vec1 = numpy.asarray([[1, 2, 3]])
vec2 = numpy.asarray([[1, 2]])
with pytest.raises(ValueError):
CosineDistance().get_grad(vec1, vec2)
@pytest.mark.parametrize(
"name,kwargs,args",
[
("CategoricalCrossentropy.v1", {}, (scores0, labels0)),
("SequenceCategoricalCrossentropy.v1", {}, ([scores0], [labels0])),
("CategoricalCrossentropy.v2", {"neg_prefix": "!"}, (scores0, labels0)),
("CategoricalCrossentropy.v3", {"neg_prefix": "!"}, (scores0, labels0)),
(
"SequenceCategoricalCrossentropy.v2",
{"neg_prefix": "!"},
([scores0], [labels0]),
),
(
"SequenceCategoricalCrossentropy.v3",
{"neg_prefix": "!"},
([scores0], [labels0]),
),
("L2Distance.v1", {}, (scores0, scores0)),
(
"CosineDistance.v1",
{"normalize": True, "ignore_zeros": True},
(scores0, scores0),
),
],
)
def test_loss_from_config(name, kwargs, args):
"""Test that losses are loaded and configured correctly from registry
(as partials)."""
cfg = {"test": {"@losses": name, **kwargs}}
func = registry.resolve(cfg)["test"]
loss = func.get_grad(*args)
if isinstance(loss, (list, tuple)):
loss = loss[0]
assert loss.ndim == 2
func.get_loss(*args)
func(*args)
|
