File size: 4,838 Bytes
2366e36
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) OpenMMLab. All rights reserved.
import math

import pytest
import torch

from mmocr.models.textrecog.decoders import (ABILanguageDecoder,
                                             ABIVisionDecoder, BaseDecoder,
                                             NRTRDecoder, ParallelSARDecoder,
                                             ParallelSARDecoderWithBS,
                                             SequentialSARDecoder)
from mmocr.models.textrecog.decoders.sar_decoder_with_bs import DecodeNode


def _create_dummy_input():
    feat = torch.rand(1, 512, 4, 40)
    out_enc = torch.rand(1, 512)
    tgt_dict = {'padded_targets': torch.LongTensor([[1, 1, 1, 1, 36]])}
    img_metas = [{'valid_ratio': 1.0}]

    return feat, out_enc, tgt_dict, img_metas


def test_base_decoder():
    decoder = BaseDecoder()
    with pytest.raises(NotImplementedError):
        decoder.forward_train(None, None, None, None)
    with pytest.raises(NotImplementedError):
        decoder.forward_test(None, None, None)


def test_parallel_sar_decoder():
    # test parallel sar decoder
    decoder = ParallelSARDecoder(num_classes=37, padding_idx=36, max_seq_len=5)
    decoder.init_weights()
    decoder.train()

    feat, out_enc, tgt_dict, img_metas = _create_dummy_input()
    with pytest.raises(AssertionError):
        decoder(feat, out_enc, tgt_dict, [], True)
    with pytest.raises(AssertionError):
        decoder(feat, out_enc, tgt_dict, img_metas * 2, True)

    out_train = decoder(feat, out_enc, tgt_dict, img_metas, True)
    assert out_train.shape == torch.Size([1, 5, 36])

    out_test = decoder(feat, out_enc, tgt_dict, img_metas, False)
    assert out_test.shape == torch.Size([1, 5, 36])


def test_sequential_sar_decoder():
    # test parallel sar decoder
    decoder = SequentialSARDecoder(
        num_classes=37, padding_idx=36, max_seq_len=5)
    decoder.init_weights()
    decoder.train()

    feat, out_enc, tgt_dict, img_metas = _create_dummy_input()
    with pytest.raises(AssertionError):
        decoder(feat, out_enc, tgt_dict, [])
    with pytest.raises(AssertionError):
        decoder(feat, out_enc, tgt_dict, img_metas * 2)

    out_train = decoder(feat, out_enc, tgt_dict, img_metas, True)
    assert out_train.shape == torch.Size([1, 5, 36])

    out_test = decoder(feat, out_enc, tgt_dict, img_metas, False)
    assert out_test.shape == torch.Size([1, 5, 36])


def test_parallel_sar_decoder_with_beam_search():
    with pytest.raises(AssertionError):
        ParallelSARDecoderWithBS(beam_width='beam')
    with pytest.raises(AssertionError):
        ParallelSARDecoderWithBS(beam_width=0)

    feat, out_enc, tgt_dict, img_metas = _create_dummy_input()
    decoder = ParallelSARDecoderWithBS(
        beam_width=1, num_classes=37, padding_idx=36, max_seq_len=5)
    decoder.init_weights()
    decoder.train()
    with pytest.raises(AssertionError):
        decoder(feat, out_enc, tgt_dict, [])
    with pytest.raises(AssertionError):
        decoder(feat, out_enc, tgt_dict, img_metas * 2)

    out_test = decoder(feat, out_enc, tgt_dict, img_metas, train_mode=False)
    assert out_test.shape == torch.Size([1, 5, 36])

    # test decodenode
    with pytest.raises(AssertionError):
        DecodeNode(1, 1)
    with pytest.raises(AssertionError):
        DecodeNode([1, 2], ['4', '3'])
    with pytest.raises(AssertionError):
        DecodeNode([1, 2], [0.5])
    decode_node = DecodeNode([1, 2], [0.7, 0.8])
    assert math.isclose(decode_node.eval(), 1.5)


def test_transformer_decoder():
    decoder = NRTRDecoder(num_classes=37, padding_idx=36, max_seq_len=5)
    decoder.init_weights()
    decoder.train()

    out_enc = torch.rand(1, 25, 512)
    tgt_dict = {'padded_targets': torch.LongTensor([[1, 1, 1, 1, 36]])}
    img_metas = [{'valid_ratio': 1.0}]
    tgt_dict['padded_targets'] = tgt_dict['padded_targets']

    out_train = decoder(None, out_enc, tgt_dict, img_metas, True)
    assert out_train.shape == torch.Size([1, 5, 36])

    out_test = decoder(None, out_enc, tgt_dict, img_metas, False)
    assert out_test.shape == torch.Size([1, 5, 36])


def test_abi_language_decoder():
    decoder = ABILanguageDecoder(max_seq_len=25)
    logits = torch.randn(2, 25, 90)
    result = decoder(
        feat=None, out_enc=logits, targets_dict=None, img_metas=None)
    assert result['feature'].shape == torch.Size([2, 25, 512])
    assert result['logits'].shape == torch.Size([2, 25, 90])


def test_abi_vision_decoder():
    model = ABIVisionDecoder(
        in_channels=128, num_channels=16, max_seq_len=10, use_result=None)
    x = torch.randn(2, 128, 8, 32)
    result = model(x, None)
    assert result['feature'].shape == torch.Size([2, 10, 128])
    assert result['logits'].shape == torch.Size([2, 10, 90])
    assert result['attn_scores'].shape == torch.Size([2, 10, 8, 32])