src/translate_imdb_flax.py

import functools
import json
import logging
import os
import pprint
from typing import Tuple

from datasets import get_dataset_config_names, load_dataset, get_dataset_split_names
import jax
import numpy as np
from flax import jax_utils
from flax.jax_utils import pad_shard_unpad
from transformers import AutoTokenizer, FlaxAutoModelForSeq2SeqLM
import pandas as pd

logger = logging.getLogger(__name__)

DATASET_NAME = "imdb"
OUTPUT_DIR = "./imdb_dutch"
MODEL_370 = "yhavinga/ul2-large-en-nl"
# BATCH_SIZE = 64
BATCH_SIZE = 32
# BATCH_SIZE = 2
MODEL_MAX_LENGTH = 370
MAX_WORDS = int(MODEL_MAX_LENGTH / 3)
END_MARKS = (".", "?", "!", '"', "'", "\n")


class FlaxModel:
    def __init__(self, model_name: str, tokenizer_name: str, tokenizer_args={}):
        """
        Initializes the FlaxModel with the specified model and tokenizer names, as well as tokenizer arguments.
        """
        self.model = FlaxAutoModelForSeq2SeqLM.from_pretrained(
            model_name, use_auth_token=True
        )
        self.model.params = self.model.to_fp32(self.model.params, mask=None)
        self.tokenizer_args = {
            # "model_max_length": self.model.config.max_length,
            **tokenizer_args,
        }
        self.tokenizer = AutoTokenizer.from_pretrained(
            tokenizer_name, use_auth_token=True, **self.tokenizer_args
        )
        # if not (
        #     self.model.config.max_length
        #     == self.tokenizer.model_max_length
        #     == self.tokenizer_args.get("model_max_length")
        # ):
        #     print(
        #         f"Warning: model max length {self.model.config.max_length} != tokenizer max length {self.tokenizer.model_max_length} != tokenizer_args max length {tokenizer_args.get('model_max_length')}"
        #     )
        #     raise ValueError("Model and tokenizer max_length should be equal")

        self.params = jax_utils.replicate(self.model.params)

        kwargs = {
            "max_length": self.tokenizer.model_max_length,
            "length_penalty": 1.0,
            "num_beams": 4,
            "early_stopping": True,
        }

        def shard(xs):
            local_device_count = jax.local_device_count()
            return jax.tree_map(
                lambda x: x.reshape((local_device_count, -1) + x.shape[1:]), xs
            )

        def generate_step(params, batch):
            self.model.params = params
            output_ids = self.model.generate(
                batch["input_ids"], attention_mask=batch["attention_mask"], **kwargs
            )
            return output_ids.sequences

        self.p_generate_step = jax.pmap(generate_step, "batch")

    @functools.lru_cache()
    def translate_batch(self, texts: Tuple[str]):
        overflowed = False
        texts = list(texts)
        if self.model.config.prefix:
            texts = [self.model.config.prefix + x for x in texts]
        texts = [x.replace("\n", "<n>").replace("<br />", "<n>") for x in texts]
        inputs = self.tokenizer(
            texts,
            max_length=self.tokenizer_args.get("model_max_length"),
            truncation=True,
            padding="max_length",
            return_tensors="np",
        )
        if not np.array_equal(
            inputs.data["input_ids"][:, self.tokenizer.model_max_length - 1],
            np.zeros(BATCH_SIZE),
        ):
            overflowed = True
            return BATCH_SIZE * [""], overflowed

        batch = inputs.data
        print(f"Batch inputs shape is {batch['input_ids'].shape}")
        translated = pad_shard_unpad(self.p_generate_step)(self.params, batch)
        predictions = jax.device_get(
            translated.reshape(-1, self.tokenizer.model_max_length)
        )
        if not np.array_equal(
            predictions[:, self.tokenizer.model_max_length - 1],
            np.zeros(BATCH_SIZE),
        ):
            overflowed = True

        output = [
            self.tokenizer.decode(t, skip_special_tokens=False) for t in predictions
        ]
        # If there is <extra_id in the output, remove it and everything after it
        output = [
            x.replace("<pad>", "").replace("</s>", "").split("<extra_id")[0]
            for x in output
        ]
        output = [x.replace("<n>", "<br />").strip() for x in output]
        return output, overflowed


def split_text(text):
    text_parts = []
    current_part = ""

    def split_on_end_marks(text):
        sentences = []
        current_sentence = ""
        for char in text:
            if char in END_MARKS:
                sentences.append(current_sentence + char)
                current_sentence = ""
            else:
                current_sentence += char

        # Add the final sentence if it wasn't ended by an end of line mark
        if current_sentence:
            sentences.append(current_sentence)
        return sentences

    text_lines = split_on_end_marks(text)

    for line in text_lines:
        # If adding the line to the current part would not exceed MAX_WORDS words, add it to the current part
        if len((current_part + line).split()) <= MAX_WORDS:
            current_part += line
        # If adding the line to the current part would exceed 200 characters, add the current part to the list and reset the current part
        else:
            if len(current_part) > 0:
                text_parts.append(current_part)
            while len(line.split()) > MAX_WORDS:
                # print(f"Line {line} is longer than MAX_WORDS words")
                current_part = " ".join(line.split()[:MAX_WORDS])
                text_parts.append(current_part + " ")
                line = " ".join(line.split()[MAX_WORDS:])
            current_part = line
    # Add the final part to the list
    text_parts.append(current_part)
    text_parts[-1] = text_parts[-1].rstrip()
    return text_parts


def test_split_text():
    # Test with single line that is less than MAX_WORDS words
    text = " ".join([f"n{i}" for i in range(MAX_WORDS - 20)])
    a = list(text)
    a[150] = END_MARKS[0]
    text = "".join(a)
    text_parts = split_text(text)
    assert text_parts == [text]

    # Test with single line that is exactly MAX_WORDS words
    text = " ".join([f"n{i}" for i in range(MAX_WORDS)])
    a = list(text)
    a[10] = END_MARKS[0]
    text = "".join(a)
    text_parts = split_text(text)
    assert text_parts == [text]

    # Test with single line that is more than MAX_WORDS words
    text = " ".join([f"n{i}" for i in range(MAX_WORDS + 1)])
    a = list(text)
    a[150] = END_MARKS[0]
    text = "".join(a)
    text_parts = split_text(text)
    assert text_parts == [text[:151], text[151:]]

    # Test with multiple lines, none of which are more than 200 characters
    text = "\n".join([f"n{i}" for i in range(10)])
    text_parts = split_text(text)
    assert text_parts == [text]

    # Test with 500 words
    text = " ".join([f"n{i}" for i in range(500)])
    a = list(text)
    a[150] = END_MARKS[0]
    a[300] = END_MARKS[0]
    a[550] = END_MARKS[0]
    a[600] = END_MARKS[0]
    a[750] = END_MARKS[0]
    a[900] = END_MARKS[0]
    a[950] = END_MARKS[0]
    a[1000] = END_MARKS[0]
    text = "".join(a)
    text_parts = split_text(text)
    assert all(
        [len(x.split()) <= MAX_WORDS for x in text_parts]
    ), "Not all text parts are less than MAX_WORDS words"
    assert "".join(text_parts) == text, "Text parts concatenated != original text"


test_split_text()


def get_file_lines(filename):
    """
    Get the number of lines in a file, 0 if the file does not exist.
    """
    lines = 0
    if os.path.exists(filename):
        with open(filename) as f:
            with open(filename, "r") as f:
                lines = len(f.readlines())
                print(f"{filename} already has {lines} lines")
    return lines


SEP = "\n"
# SEP="<unk>"


def main():
    os.makedirs(OUTPUT_DIR, exist_ok=True)

    model_370 = FlaxModel(
        MODEL_370, MODEL_370, tokenizer_args={"model_max_length": MODEL_MAX_LENGTH}
    )

    for config in get_dataset_config_names(DATASET_NAME):
        print(f"Processing config {config}")
        ds = load_dataset(DATASET_NAME, config)
        # for split in ["validation"]:
        for split in get_dataset_split_names(DATASET_NAME, config):
            output_file = f"{OUTPUT_DIR}/{DATASET_NAME}_dutch_{config}-{split}.json"
            num_examples = len(ds[split])
            # fn = partial(encode_in_single_text, validation=(split == "validation"))
            # single_text_ds = ds[split].map(fn, num_proc=6).sort("length", reverse=True)
            # # fn = partial(batch_single_text_decode, validation=(split == "validation"))
            # # decoded_ds = single_text_ds.map(fn, num_proc=6)
            #
            lines = get_file_lines(output_file)
            start_batch_index = lines // BATCH_SIZE
            with open(output_file, mode="ab" if lines else "wb") as writer:
                for batch_index in range(start_batch_index, num_examples // BATCH_SIZE):
                    ds_split = ds[split]
                    batch = ds_split[
                        batch_index * BATCH_SIZE : (batch_index + 1) * BATCH_SIZE
                    ]
                    print(
                        f"Translating batch {batch_index} of {num_examples // BATCH_SIZE}"
                    )

                    translated, overflow = model_370.translate_batch(
                        tuple(batch["text"])
                    )
                    translated_batch = [{"text": x} for x in translated]
                    if overflow:
                        batch_text_splitted = [
                            split_text(text) for text in batch["text"]
                        ]
                        max_parts = max(
                            [len(text) for text in batch_text_splitted]
                        )
                        text_translated = [""] * BATCH_SIZE
                        for part_index in range(max_parts):
                            text_parts_i = [
                                text[part_index] if part_index < len(text) else ""
                                for text in batch_text_splitted
                            ]
                            (
                                text_part_translated,
                                overflow,
                            ) = model_370.translate_batch(tuple(text_parts_i))
                            if overflow:
                                print(
                                    f"This shouldn't happen, overflow on a splitted text: {text_parts_i}"
                                )
                            for bi in range(BATCH_SIZE):
                                text_translated[bi] += " " + text_part_translated[bi] if text_parts_i[bi] != "" else ""
                        for bi in range(BATCH_SIZE):
                            translated_batch[bi]["text"] = text_translated[bi].strip()

                    # write each object in the batch as a separate line
                    for bi in range(BATCH_SIZE):
                        example = {
                            "text": translated_batch[bi]["text"],
                            "text_en": batch["text"][bi],
                            "label": batch["label"][bi],
                        }

                        pprint.pprint(example)
                        writer.write(json.dumps(example).encode("utf-8"))
                        writer.write("\n".encode("utf-8"))


if __name__ == "__main__":
    main()