-
Notifications
You must be signed in to change notification settings - Fork 0
/
weight_fns_test.py
249 lines (217 loc) · 9.39 KB
/
weight_fns_test.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
from absl.testing import absltest
import numpy as np
import numpy.random as npr
import torch
import torch.nn as nn
import weight_fns
class WeightFnTest(absltest.TestCase):
def test_hat_normalize(self):
blank = torch.tensor([2.0, 7.0])
lexical = torch.tensor([[0.0, 1.0], [3.0, 5.0]])
expect_blank = torch.tensor([-0.126928, -0.000912])
expect_lexical = torch.tensor([[-3.44019, -2.44019], [-9.12784, -7.12784]])
actual_blank, actual_lexical = weight_fns.hat_normalize(blank, lexical)
torch.testing.assert_close(actual_blank, expect_blank, rtol=1e-3, atol=1e-6)
torch.testing.assert_close(actual_lexical, expect_lexical, rtol=1e-6, atol=1e-6)
def test_log_softmax_normalize(self):
blank = torch.tensor([2.0, 7.0])
lexical = torch.tensor([[0.0, 1.0], [3.0, 5.0]])
expect_blank = torch.tensor([-0.407606, -0.142932])
expect_lexical = torch.tensor([[-2.407606, -1.407606], [-4.142932, -2.142932]])
actual_blank, actual_lexical = weight_fns.log_softmax_normalize(blank, lexical)
torch.testing.assert_close(actual_blank, expect_blank, rtol=1e-3, atol=1e-6)
torch.testing.assert_close(actual_lexical, expect_lexical, rtol=1e-6, atol=1e-6)
def test_NullCacher(self):
# Get cache and inspect for trainable paramaters
cacher = weight_fns.NullCacher()
num_params: int = sum(p.numel() for p in cacher.parameters() if p.requires_grad)
self.assertEqual(num_params, 0)
self.assertIsNone(cacher())
def test_TableWeightFn(self):
with self.subTest("batch ndim = 0"):
table = torch.arange(5 * 4 * 3).reshape([5, 4, 3])
weight_fn = weight_fns.TableWeightFn(table)
frame = torch.tensor([1.0, 2.0])
(blank, lexical) = weight_fn(None, frame)
num_params: int = sum(
p.numel() for p in weight_fn.parameters() if p.requires_grad
)
self.assertEqual(num_params, 0)
torch.testing.assert_close(blank, table[1, :, 0])
torch.testing.assert_close(lexical, table[1, :, 1:])
state = torch.tensor(3)
blank, lexical = weight_fn(None, frame, state)
torch.testing.assert_close(blank, table[1, 3, 0])
torch.testing.assert_close(lexical, table[1, 3, 1:])
with self.assertRaisesRegex(
ValueError,
r"frame should have batch_dims=\(\) but got torch.Size\(\[1\]\)",
):
weight_fn(None, frame[None])
with self.subTest("batch ndim = 1"):
table = torch.arange(2 * 5 * 4 * 3).reshape([2, 5, 4, 3])
weight_fn = weight_fns.TableWeightFn(table)
frame = torch.tensor([[1.0, 2.0], [4.0, 3.0]])
(blank, lexical) = weight_fn(None, frame)
num_params: int = sum(
p.numel() for p in weight_fn.parameters() if p.requires_grad
)
self.assertEqual(num_params, 0)
torch.testing.assert_close(
blank, torch.stack((table[0, 1, :, 0], table[1, 4, :, 0]))
)
torch.testing.assert_close(
lexical, torch.stack((table[0, 1, :, 1:], table[1, 4, :, 1:]))
)
state = torch.tensor([3, 2])
blank, lexical = weight_fn(None, frame, state)
torch.testing.assert_close(
blank, torch.stack((table[0, 1, 3, 0], table[1, 4, 2, 0]))
)
torch.testing.assert_close(
lexical, torch.stack((table[0, 1, 3, 1:], table[1, 4, 2, 1:]))
)
with self.assertRaisesRegex(
ValueError,
r"frame should have batch_dims=\(2,\) but got torch.Size\(\[1, 2\]\)",
):
weight_fn(None, frame[None])
class LocallyNormalizedWeightFnTest(absltest.TestCase):
def test_call(self):
weight_fn = weight_fns.LocallyNormalizedWeightFn(
weight_fns.JointWeightFn(vocab_size=3, hidden_size=8)
)
rngs = npr.default_rng(0).integers(low=0, high=1e9, size=5)
frame = torch.from_numpy(
npr.default_rng(rngs[0]).uniform(high=1, size=(2, 4)).astype(np.float32)
)
cache = torch.from_numpy(
npr.default_rng(rngs[1]).uniform(high=1, size=(6, 5)).astype(np.float32)
) # context embeddings.
with self.subTest("all context states"):
blank, lexical = weight_fn(cache, frame)
torch.testing.assert_close(blank.shape, torch.Size([2, 6]))
torch.testing.assert_close(lexical.shape, torch.Size([2, 6, 3]))
torch.testing.assert_close(
torch.exp(blank) + torch.sum(torch.exp(lexical), dim=-1),
torch.ones_like(blank),
rtol=1e-4,
atol=1e-4,
)
with self.subTest("per-state"):
state = torch.tensor([2, 4])
blank_per_state, lexical_per_state = weight_fn(cache, frame, state)
torch.testing.assert_close(
blank_per_state,
blank[torch.tensor([0, 1]), state],
rtol=1e-6,
atol=1e-6,
)
torch.testing.assert_close(
lexical_per_state,
lexical[torch.tensor([0, 1]), state],
rtol=1e-6,
atol=1e-6,
)
class JointWeightFnTest(absltest.TestCase):
def test_call(self):
weight_fn = weight_fns.JointWeightFn(vocab_size=3, hidden_size=8)
rngs = npr.default_rng(0).integers(low=0, high=1e9, size=5)
frame = torch.from_numpy(
npr.default_rng(rngs[0]).uniform(high=1, size=(2, 4)).astype(np.float32)
)
cache = torch.from_numpy(
npr.default_rng(rngs[1]).uniform(high=1, size=(6, 5)).astype(np.float32)
) # context embeddings.
with self.subTest("all context states"):
blank, lexical = weight_fn(cache, frame)
torch.testing.assert_close(blank.shape, torch.Size([2, 6]))
torch.testing.assert_close(lexical.shape, torch.Size([2, 6, 3]))
with self.subTest("per-state"):
state = torch.tensor([2, 4])
blank_per_state, lexical_per_state = weight_fn(cache, frame, state)
torch.testing.assert_close(
blank_per_state,
blank[torch.tensor([0, 1]), state],
rtol=1e-6,
atol=1e-6,
)
torch.testing.assert_close(
lexical_per_state,
lexical[torch.tensor([0, 1]), state],
rtol=1e-6,
atol=1e-6,
)
def test_SharedEmbCacher(self):
cacher = weight_fns.SharedEmbCacher(num_context_states=4, embedding_size=5)
cacher()
param = cacher.get_parameter("context_embeddings")
torch.testing.assert_close(param.size(), torch.Size([4, 5]))
torch.testing.assert_close(param, cacher())
def test_SharedRNNCacher(self):
pad = -2.0
start = -1.0
class FakeRNNCell(nn.RNNCellBase):
"""Test RNN cell that remembers past inputs."""
def __init__(self, features: int) -> None:
self.features = features
@property
def num_features(self):
return 1
def __call__(self, carry, inputs):
carry = torch.cat([carry[..., 1:], inputs[..., :1]], dim=-1)
return carry, carry
@staticmethod
def initialize_carry(self, size):
batch_dims = size[:1]
return torch.full((*batch_dims, self.features), pad)
params = nn.Parameter(
torch.broadcast_to(torch.tensor([start, 1.0, 2.0, 3.0])[:, None], (4, 6))
)
with self.subTest("context_size=2"):
cacher = weight_fns.SharedRNNCacher(
vocab_size=3,
context_size=2,
rnn_size=4,
rnn_embedding_size=6,
rnn_cell=FakeRNNCell(features=4),
)
cacher.embed.weight = params
torch.testing.assert_close(
cacher(),
torch.tensor(
[
# Start.
[pad, pad, pad, start],
# Unigrams.
[pad, pad, start, 1],
[pad, pad, start, 2],
[pad, pad, start, 3],
# Bigrams.
[pad, start, 1, 1],
[pad, start, 1, 2],
[pad, start, 1, 3],
[pad, start, 2, 1],
[pad, start, 2, 2],
[pad, start, 2, 3],
[pad, start, 3, 1],
[pad, start, 3, 2],
[pad, start, 3, 3],
]
),
)
with self.subTest("context_size=0"):
cacher = weight_fns.SharedRNNCacher(
vocab_size=3,
context_size=0,
rnn_size=4,
rnn_embedding_size=6,
rnn_cell=FakeRNNCell(features=4),
)
cacher.embed.weight = params
torch.testing.assert_close(
cacher(),
torch.tensor([[pad, pad, pad, start]]),
)
if __name__ == "__main__":
absltest.main()