mmocr/tests/test_models/test_ocr_encoder.py

import pytest
import torch

from mmocr.models.textrecog.encoders import BaseEncoder, SAREncoder, TFEncoder


def test_sar_encoder():
    with pytest.raises(AssertionError):
        SAREncoder(enc_bi_rnn='bi')
    with pytest.raises(AssertionError):
        SAREncoder(enc_do_rnn=2)
    with pytest.raises(AssertionError):
        SAREncoder(enc_gru='gru')
    with pytest.raises(AssertionError):
        SAREncoder(d_model=512.5)
    with pytest.raises(AssertionError):
        SAREncoder(d_enc=200.5)
    with pytest.raises(AssertionError):
        SAREncoder(mask='mask')

    encoder = SAREncoder()
    encoder.init_weights()
    encoder.train()

    feat = torch.randn(1, 512, 4, 40)
    img_metas = [{'valid_ratio': 1.0}]
    with pytest.raises(AssertionError):
        encoder(feat, img_metas * 2)
    out_enc = encoder(feat, img_metas)

    assert out_enc.shape == torch.Size([1, 512])


def test_transformer_encoder():
    tf_encoder = TFEncoder()
    tf_encoder.init_weights()
    tf_encoder.train()

    feat = torch.randn(1, 512, 1, 25)
    out_enc = tf_encoder(feat)
    print('hello', out_enc.size())
    assert out_enc.shape == torch.Size([1, 512, 1, 25])


def test_base_encoder():
    encoder = BaseEncoder()
    encoder.init_weights()
    encoder.train()

    feat = torch.randn(1, 256, 4, 40)
    out_enc = encoder(feat)
    assert out_enc.shape == torch.Size([1, 256, 4, 40])