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
"""Tests for utils."""
17

18
import functools
19
from absl.testing import parameterized
20
import numpy as np
21
import tensorflow.compat.v1 as tf
22
from protein_lm import domains
23
from protein_lm import models
24
from protein_lm import utils
25

26
lm_cfg = dict(
27
    batch_size=1, num_layers=2, num_heads=2, emb_dim=32, mlp_dim=32, qkv_dim=32)
28
lm_cls = functools.partial(models.FlaxLM, **lm_cfg)
29

30

31
class UtilsTest(tf.test.TestCase, parameterized.TestCase):
32

33
  def test_count_params(self):
34
    domain = domains.FixedLengthDiscreteDomain(length=4, vocab_size=2)
35
    lm = lm_cls(domain=domain)
36
    count = utils.param_count(lm)
37
    self.assertEqual(13059, count)
38

39
    # Check these methods run.
40
    utils.param_pprint(lm)
41
    sizes = utils.param_reduce(lm, log=True)
42
    self.assertIsInstance(sizes, dict)
43

44
  @parameterized.parameters((5, 5), (5, 1), (5, 2), (5, 6), (5, 12))
45
  def test_batch_apply(self, batch_size, num_inputs):
46
    def fn(inputs):
47
      return np.power(inputs + 1, 2)
48

49
    def batch_fn(batched_inputs):
50
      if len(batched_inputs) != batch_size:
51
        raise ValueError('fn() called with a batch that is '
52
                         'the wrong size (%d vs. %d).' % (len(batched_inputs),
53
                                                          batch_size))
54
      return fn(batched_inputs)
55
    inputs = np.stack([np.arange(num_inputs), -np.arange(num_inputs)], axis=1)
56
    unbatched_output = fn(inputs)
57
    batched_output = utils.batch_apply(batch_fn, inputs, batch_size)
58
    np.testing.assert_array_equal(unbatched_output, batched_output)
59

60
  def test_get_normalized_matrix(self):
61
    """Tests that the normalized matrix is computed correctly."""
62
    domain = domains.FixedLengthDiscreteDomain(
63
        vocab=domains.Vocabulary(tokens=['A', 'B', 'C']),
64
        length=2)
65
    freq_dict = {'A': {'A': 5, 'B': 3, 'C': 1},
66
                 'B': {'A': 3, 'B': 5, 'C': 1},
67
                 'C': {'A': 1, 'B': 1, 'C': 1}}
68
    matrix = utils.get_normalized_matrix(domain, freq_dict)
69
    expected_matrix = [[1, 0.5, 0], [0.5, 1, 0,], [0, 0, 0]]
70
    self.assertAllEqual(matrix, expected_matrix)
71

72
  def test_soft_accuracy(self):
73
    """Tests that soft accuracy is computed correctly."""
74
    domain = domains.FixedLengthDiscreteDomain(
75
        vocab=domains.Vocabulary(tokens=['A', 'B', 'C']),
76
        length=2)
77
    targets = np.array([[0, 1]])
78
    logits = np.log([[[0.9, 0.1], [0.6, 0.4]]])
79
    freq_dict = {'A': {'A': 5, 'B': 3, 'C': 1},
80
                 'B': {'A': 3, 'B': 5, 'C': 1},
81
                 'C': {'A': 1, 'B': 1, 'C': 1}}
82
    accuracy, denominator = utils.compute_weighted_soft_accuracy(
83
        logits, targets,
84
        weights=None,
85
        matrix=utils.get_normalized_matrix(domain, freq_dict))
86
    self.assertEqual(accuracy / denominator, 0.75)
87

88

89
if __name__ == '__main__':
90
  tf.test.main()
91