Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_cmnrl.py 4.63 KB
Newer Older
Rayyyyy's avatar
First add  
Rayyyyy committed
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 
from contextlib import nullcontext
from typing import List
import pytest
from sentence_transformers import SentenceTransformer, InputExample, losses
import tqdm
from transformers import set_seed
import torch
from torch.optim import Adam


@pytest.mark.parametrize(
    ["train_samples_mnrl", "train_samples_cmnrl", "same_grad", "scaler", "precision"],
    [
        (
            [
                InputExample(texts=[q, p, n])
                for q, p, n in zip(
                    ["aaa", "bbb", "ccc", "ddd", "eee"],
                    ["aas", "bbs", "ccs", "dds", "ees"],
                    ["xxx", "yyy", "zzz", "kkk", "fff"],
                )
            ],
            [
                InputExample(texts=[q, p, n])
                for q, p, n in zip(
                    ["aaa", "bbb", "ccc", "ddd", "eee"],
                    ["aas", "bbs", "ccs", "dds", "ees"],
                    ["xxx", "yyy", "zzz", "kkk", "fff"],
                )
            ],
            True,
            1.0,
            1e-6,
        ),
        (
            [
                InputExample(texts=[q, p, n])
                for q, p, n in zip(
                    ["adsa", "czx", "dsada"],
                    ["b", "fas", "xcz"],
                    ["c", "yyy", "asdas"],
                )
            ],
            [
                InputExample(texts=[q, p, n])
                for q, p, n in zip(
                    ["aaa", "bbb", "ccc", "ddd", "eee"],
                    ["aas", "bbs", "ccs", "dds", "ees"],
                    ["xxx", "yyy", "zzz", "kkk", "fff"],
                )
            ],
            False,
            1.0,
            1e-6,
        ),
        (
            [
                InputExample(texts=[q, p, n])
                for q, p, n in zip(
                    ["aaa", "bbb", "ccc", "ddd", "eee"],
                    ["aas", "bbs", "ccs", "dds", "ees"],
                    ["xxx", "yyy", "zzz", "kkk", "fff"],
                )
            ],
            [
                InputExample(texts=[q, p, n])
                for q, p, n in zip(
                    ["aaa", "bbb", "ccc", "ddd", "eee"],
                    ["aas", "bbs", "ccs", "dds", "ees"],
                    ["xxx", "yyy", "zzz", "kkk", "fff"],
                )
            ],
            True,
            1000.0,
            1e-3,
        ),
    ],
)
def test_cmnrl_same_grad(
    train_samples_mnrl: List[InputExample],
    train_samples_cmnrl: List[InputExample],
    same_grad: bool,
    scaler: float,
    precision: float,
):
    # Given:
    sbert = SentenceTransformer("distilbert-base-uncased")
    sbert.to("cpu")
    optimizer = Adam(sbert.parameters())
    # train_samples_mnrl
    # train_samples_cmnrl
    # same_grad
    # scaler  # This simulates AMP scenarios
    # precision

    # When:
    # First run with MNRL
    set_seed(42)
    optimizer.zero_grad()
    loss_mnrl = losses.MultipleNegativesRankingLoss(sbert)
    loss_mnrl_value: torch.Tensor = loss_mnrl.forward(*sbert.smart_batching_collate(train_samples_mnrl)) * scaler
    loss_mnrl_value.backward()
    grad_expected = {name: p.grad.clone() for name, p in loss_mnrl.named_parameters() if p.grad is not None}

    # Then run with this cached version:
    set_seed(42)
    optimizer.zero_grad()
    loss_cmnrl = losses.CachedMultipleNegativesRankingLoss(sbert, mini_batch_size=2)
    loss_cmnrl_value = loss_cmnrl.forward(*sbert.smart_batching_collate(train_samples_cmnrl)) * scaler
    loss_cmnrl_value.backward()
    grad = {name: p.grad.clone() for name, p in loss_cmnrl.named_parameters() if p.grad is not None}

    # Then:
    if same_grad:
        assert pytest.approx(loss_mnrl_value.item()) == loss_cmnrl_value.item()
    else:
        assert pytest.approx(loss_mnrl_value.item()) != loss_cmnrl_value.item()

    nclose = 0
    for name in tqdm.tqdm(grad_expected):
        nclose += torch.allclose(grad[name], grad_expected[name], precision, precision)

    if same_grad:
        assert nclose == len(grad_expected)
    else:
        assert nclose != len(grad_expected)


@pytest.mark.parametrize("use_rand_context", [True, False])
def test_rand_context_working(use_rand_context: bool):
    # Given:
    from sentence_transformers.losses.CachedMultipleNegativesRankingLoss import (
        RandContext,
    )

    a = torch.Tensor(1)
    b = torch.Tensor(1)
    random_state = RandContext(a, b) if use_rand_context else nullcontext()
    expected = torch.rand(1000)
    precision = 1e-6

    # When:
    with random_state:
        # Then:
        if use_rand_context:
            assert torch.allclose(torch.rand(1000), expected, precision, precision)
        else:
            assert not torch.allclose(torch.rand(1000), expected, precision, precision)