transformers/tests/models/perceiver/test_tokenization_perceiver.py

# Copyright 2021 The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import re
import shutil
import tempfile
import unittest
from functools import cached_property

from transformers import BatchEncoding, PerceiverTokenizer

from ...test_tokenization_common import TokenizerTesterMixin


class PerceiverTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
    from_pretrained_id = "deepmind/language-perceiver"
    tokenizer_class = PerceiverTokenizer
    test_rust_tokenizer = False

    @classmethod
    def setUpClass(cls):
        super().setUpClass()
        tokenizer = PerceiverTokenizer()
        tokenizer.save_pretrained(cls.tmpdirname)

    @cached_property
    def perceiver_tokenizer(self):
        return PerceiverTokenizer.from_pretrained("deepmind/language-perceiver")

    @classmethod
    def get_tokenizer(cls, pretrained_name=None, **kwargs) -> PerceiverTokenizer:
        pretrained_name = pretrained_name or cls.tmpdirname
        return cls.tokenizer_class.from_pretrained(pretrained_name, **kwargs)

    def get_clean_sequence(self, tokenizer, with_prefix_space=False, max_length=20, min_length=5) -> tuple[str, list]:
        # XXX The default common tokenizer tests assume that every ID is decodable on its own.
        # This assumption is invalid for Perceiver because single bytes might not be
        # valid utf-8 (byte 128 for instance).
        # Here we're overriding the smallest possible method to provide
        # a clean sequence without making the same assumption.

        toks = []
        for i in range(len(tokenizer)):
            try:
                tok = tokenizer.decode([i], clean_up_tokenization_spaces=False)
            except UnicodeDecodeError:
                pass
            toks.append((i, tok))

        toks = list(filter(lambda t: re.match(r"^[ a-zA-Z]+$", t[1]), toks))
        toks = list(filter(lambda t: [t[0]] == tokenizer.encode(t[1], add_special_tokens=False), toks))
        if max_length is not None and len(toks) > max_length:
            toks = toks[:max_length]
        if min_length is not None and len(toks) < min_length and len(toks) > 0:
            while len(toks) < min_length:
                toks = toks + toks
        # toks_str = [t[1] for t in toks]
        toks_ids = [t[0] for t in toks]

        # Ensure consistency
        output_txt = tokenizer.decode(toks_ids, clean_up_tokenization_spaces=False)
        if " " not in output_txt and len(toks_ids) > 1:
            output_txt = (
                tokenizer.decode([toks_ids[0]], clean_up_tokenization_spaces=False)
                + " "
                + tokenizer.decode(toks_ids[1:], clean_up_tokenization_spaces=False)
            )
        if with_prefix_space:
            output_txt = " " + output_txt
        output_ids = tokenizer.encode(output_txt, add_special_tokens=False)
        return output_txt, output_ids

    def test_multibytes_char(self):
        tokenizer = self.perceiver_tokenizer
        src_text = "Unicode €."
        encoded = tokenizer(src_text)
        encoded_ids = [4, 91, 116, 111, 105, 117, 106, 107, 38, 232, 136, 178, 52, 5]
        self.assertEqual(encoded["input_ids"], encoded_ids)

        # decoding
        decoded = tokenizer.decode(encoded_ids)
        self.assertEqual(decoded, "[CLS]Unicode €.[SEP]")

        encoded = tokenizer("e è é ê ë")
        encoded_ids = [4, 107, 38, 201, 174, 38, 201, 175, 38, 201, 176, 38, 201, 177, 5]
        self.assertEqual(encoded["input_ids"], encoded_ids)
        # decoding
        decoded = tokenizer.decode(encoded_ids)
        self.assertEqual(decoded, "[CLS]e è é ê ë[SEP]")

        # encode/decode, but with `encode` instead of `__call__`
        self.assertEqual(tokenizer.decode(tokenizer.encode("e è é ê ë")), "[CLS]e è é ê ë[SEP]")

    def test_prepare_batch_integration(self):
        tokenizer = self.perceiver_tokenizer
        src_text = ["A long paragraph for summarization.", "Another paragraph for summarization."]
        expected_src_tokens = [4, 71, 38, 114, 117, 116, 109, 38, 118, 103, 120, 103, 109, 120, 103, 118, 110, 38, 108, 117, 120, 38, 121, 123, 115, 115, 103, 120, 111, 128, 103, 122, 111, 117, 116, 52, 5, 0]  # fmt: skip
        batch = tokenizer(src_text, padding=True, return_tensors="pt")
        self.assertIsInstance(batch, BatchEncoding)

        result = list(batch.input_ids.numpy()[0])

        self.assertListEqual(expected_src_tokens, result)

        self.assertEqual((2, 38), batch.input_ids.shape)
        self.assertEqual((2, 38), batch.attention_mask.shape)

    def test_empty_target_text(self):
        tokenizer = self.perceiver_tokenizer
        src_text = ["A long paragraph for summarization.", "Another paragraph for summarization."]
        batch = tokenizer(src_text, padding=True, return_tensors="pt")
        # check if input_ids are returned and no decoder_input_ids
        self.assertIn("input_ids", batch)
        self.assertIn("attention_mask", batch)
        self.assertNotIn("decoder_input_ids", batch)
        self.assertNotIn("decoder_attention_mask", batch)

    def test_max_length_integration(self):
        tokenizer = self.perceiver_tokenizer
        tgt_text = [
            "Summary of the text.",
            "Another summary.",
        ]
        targets = tokenizer(
            text_target=tgt_text, max_length=32, padding="max_length", truncation=True, return_tensors="pt"
        )
        self.assertEqual(32, targets["input_ids"].shape[1])

    # cannot use default save_and_load_tokenizer test method because tokenizer has no vocab
    def test_save_and_load_tokenizer(self):
        # safety check on max_len default value so we are sure the test works
        tokenizers = self.get_tokenizers()
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                self.assertNotEqual(tokenizer.model_max_length, 42)

        # Now let's start the test
        tokenizers = self.get_tokenizers()
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                # Isolate this from the other tests because we save additional tokens/etc
                tmpdirname = tempfile.mkdtemp()

                sample_text = " He is very happy, UNwant\u00e9d,running"
                before_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
                tokenizer.save_pretrained(tmpdirname)

                after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
                after_tokens = after_tokenizer.encode(sample_text, add_special_tokens=False)
                self.assertListEqual(before_tokens, after_tokens)

                shutil.rmtree(tmpdirname)

        tokenizers = self.get_tokenizers(model_max_length=42)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                # Isolate this from the other tests because we save additional tokens/etc
                tmpdirname = tempfile.mkdtemp()

                sample_text = " He is very happy, UNwant\u00e9d,running"
                tokenizer.add_tokens(["bim", "bambam"])
                extra_special_tokens = tokenizer.extra_special_tokens
                extra_special_tokens.append("new_extra_special_token")
                tokenizer.add_special_tokens(
                    {"extra_special_tokens": extra_special_tokens}, replace_extra_special_tokens=False
                )
                before_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
                tokenizer.save_pretrained(tmpdirname)

                after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
                after_tokens = after_tokenizer.encode(sample_text, add_special_tokens=False)
                self.assertListEqual(before_tokens, after_tokens)
                self.assertIn("new_extra_special_token", after_tokenizer.extra_special_tokens)
                self.assertEqual(after_tokenizer.model_max_length, 42)

                tokenizer = tokenizer.__class__.from_pretrained(tmpdirname, model_max_length=43)
                self.assertEqual(tokenizer.model_max_length, 43)

                shutil.rmtree(tmpdirname)

    def test_decode_invalid_byte_id(self):
        tokenizer = self.perceiver_tokenizer
        self.assertEqual(tokenizer.decode([178]), "<EFBFBD>")

    @unittest.skip(reason="tokenizer does not have vocabulary")
    def test_get_vocab(self):
        pass

    @unittest.skip(reason="inputs cannot be pretokenized")
    def test_pretokenized_inputs(self):
        # inputs cannot be pretokenized since ids depend on whole input string and not just on single characters
        pass

    @unittest.skip(reason="vocab does not exist")
    def test_conversion_reversible(self):
        pass

    def test_convert_tokens_to_string_format(self):
        # The default common tokenizer tests uses invalid tokens for Perceiver that can only accept one-character
        # strings and special added tokens as tokens
        tokenizers = self.get_tokenizers(fast=True, do_lower_case=True)
        for tokenizer in tokenizers:
            with self.subTest(f"{tokenizer.__class__.__name__}"):
                tokens = ["[CLS]", "t", "h", "i", "s", " ", "i", "s", " ", "a", " ", "t", "e", "s", "t", "[SEP]"]
                string = tokenizer.convert_tokens_to_string(tokens)

                self.assertIsInstance(string, str)