Function test_textdecoder_multiple_rows

tests/test_utils.py:82–111

(monkeypatch)

def test_textdecoder_multiple_rows(monkeypatch):
    chars_dict = {"<PAD>": 0, "<EOS>": 1, "A": 2, "B": 3, "C": 4, "D": 5}
    decoder = TextDecoder(chars_dict=chars_dict, decode_mode=DecodeMode.Normal)

    # 對 decode 方法進行 monkeypatch，不將輸入轉換為 NumPy 陣列，以處理不同長度的序列
    def patched_decode(self, encode):
        if self.decode_mode == DecodeMode.CTC:
            masks = [(row != np.roll(row, 1)) & (row != 0) for row in encode]
        elif self.decode_mode in [DecodeMode.Default, DecodeMode.Normal]:
            masks = []
            for row in encode:
                eos_index = np.where(row == self.chars_dict["<EOS>"])[0]
                if eos_index.size > 0:
                    mask = np.zeros_like(row, dtype=bool)
                    mask[:eos_index[0]] = True
                else:
                    mask = np.ones_like(row, dtype=bool)
                mask = mask & (row != self.chars_dict["<PAD>"])
                masks.append(mask)
        chars_list = [''.join([self.chars[idx] for idx in row[m]])
                      for row, m in zip(encode, masks)]
        return chars_list
    monkeypatch.setattr(TextDecoder, "decode", patched_decode)

    row1 = np.array([2, 3, 4, 1, 0], dtype=np.int32)  # "ABC"
    # 預期輸出：tokens 為 [5, 2] => "DA"
    row2 = np.array([5, 2, 1, 0], dtype=np.int32)
    result = decoder.decode([row1, row2])
    # 將預期結果調整為 ["ABC", "DA"]，符合 decode 方法邏輯
    assert result == ["ABC", "DA"]

# 測試 decode_mode 的型別轉換功能