使用 JAX 和 Flax 微調 RecurrentGemma

前往 ai.google.dev 查看 在 Google Colab 中執行 在 Vertex AI 中開啟 前往 GitHub 查看原始碼

本教學課程說明如何使用 Google DeepMind 的 recurrentgemma 程式庫、JAX (高效能數值運算程式庫)、Flax (用於編寫 JAX 的類神經網路程式庫，以及用於編寫 JAX 的 MTA 的 MTA1} (一種用於編寫 JAX 的 MTA1) 的公用程式程式庫，微調 RecurrentGemma 2B Instruct 模型雖然這個筆記本並未直接使用 Flax，但要使用 Flax 建立 Gemma。

recurrentgemma 程式庫是使用 JAX、Flax、Orbax (用於檢查點等訓練公用程式的 JAX 程式庫) 和 SentencePiece (符記化工具/去符記化程式庫) 編寫。

這個筆記本可以在 Google Colab 和 T4 GPU 上執行 (依序前往「編輯」 >「筆記本設定」 >「硬體加速器」下方選取「T4 GPU」)。

設定

以下各節說明準備筆記本使用 RecurrentGemma 模型的步驟，包括模型存取、取得 API 金鑰，以及設定筆記本執行階段。

設定 Gemma 的 Kaggle 存取權

如要完成本教學課程，請先按照類似 Gemma 設定的設定指示操作，但請留意以下例外情況：

  • 透過 kaggle.com 存取 RecurrentGemma (而非 Gemma)。
  • 選取資源充足的 Colab 執行階段來執行 RecurrentGemma 模型。
  • 產生並設定 Kaggle 使用者名稱和 API 金鑰。

完成 RecurrentGemma 設定後，請繼續前往下一節，設定 Colab 環境的環境變數。

設定環境變數

設定 KAGGLE_USERNAME 和 KAGGLE_KEY 的環境變數。出現「要授予存取權？」訊息時，請同意提供密鑰存取權。

import os
from google.colab import userdata # `userdata` is a Colab API.

os.environ["KAGGLE_USERNAME"] = userdata.get('KAGGLE_USERNAME')
os.environ["KAGGLE_KEY"] = userdata.get('KAGGLE_KEY')

安裝 recurrentgemma 程式庫

目前免費的 Colab 硬體加速功能不足，無法執行這個筆記本。如果使用 Colab Pay As You Go 或 Colab Pro，請依序點選「編輯」 >「筆記本設定」 > 選取「A100 GPU」 >「儲存」來啟用硬體加速功能。

接下來，您需要從 github.com/google-deepmind/recurrentgemma 安裝 Google DeepMind recurrentgemma 程式庫。如果收到有關「pip 依附元件解析器」的錯誤，通常可以忽略。

pip install -q git+https://github.com/google-deepmind/recurrentgemma.git

匯入程式庫

這個筆記本使用 Flax (適用於類神經網路)、核心 JAX、SentencePiece (用於權杖化)、Chex (編寫可靠 JAX 程式碼的公用程式程式庫)、Optax (梯度處理和最佳化程式庫)，以及 TensorFlow 資料集。

import pathlib
from typing import Any, Mapping, Iterator
import enum
import functools

import chex
import jax
import jax.numpy as jnp
import optax

import tensorflow as tf
import tensorflow_datasets as tfds

import sentencepiece as spm

from recurrentgemma import jax as recurrentgemma

載入 RecurrentGemma 模型

  1. 使用 kagglehub.model_download 載入 RecurrentGemma 模型，該模型會使用三個引數：
  • handle：Kaggle 的模型控制代碼
  • path：(選用字串) 本機路徑
  • force_download：(選用布林值) 強制重新下載模型
RECURRENTGEMMA_VARIANT = '2b-it' # @param ['2b', '2b-it'] {type:"string"}

import kagglehub

RECURRENTGEMMA_PATH = kagglehub.model_download(f'google/recurrentgemma/flax/{RECURRENTGEMMA_VARIANT}')

Downloading from https://www.kaggle.com/api/v1/models/google/recurrentgemma/flax/2b-it/1/download...
100%|██████████| 3.85G/3.85G [00:50<00:00, 81.5MB/s]
Extracting model files...

print('RECURRENTGEMMA_VARIANT:', RECURRENTGEMMA_VARIANT)

RECURRENTGEMMA_VARIANT: 2b-it
  1. 檢查模型權重的位置和符記化工具，然後設定路徑變數。符記化工具目錄會存放您下載模型的主要目錄，而模型權重則位於子目錄。例如：
  • tokenizer.model 檔案會在 /LOCAL/PATH/TO/recurrentgemma/flax/2b-it/1 中)。
  • 模型查核點位於 /LOCAL/PATH/TO/recurrentgemma/flax/2b-it/1/2b-it)。
CKPT_PATH = os.path.join(RECURRENTGEMMA_PATH, RECURRENTGEMMA_VARIANT)
TOKENIZER_PATH = os.path.join(RECURRENTGEMMA_PATH, 'tokenizer.model')
print('CKPT_PATH:', CKPT_PATH)
print('TOKENIZER_PATH:', TOKENIZER_PATH)

CKPT_PATH: /root/.cache/kagglehub/models/google/recurrentgemma/flax/2b-it/1/2b-it
TOKENIZER_PATH: /root/.cache/kagglehub/models/google/recurrentgemma/flax/2b-it/1/tokenizer.model

載入並準備 MTNT 資料集和 Gemma 權杖化工具

您將使用 MTNT (Noisy Text 機器翻譯) 資料集，您可以透過 TensorFlow 資料集取得該資料集。

下載 MTNT 資料集的「English-to-French」資料集部分，然後取樣兩個範例。資料集中的每個樣本都包含兩個項目：src：原始英文語句，dst：對應的法文翻譯。

ds = tfds.load("mtnt/en-fr", split="train")

ds = ds.take(2)
ds = ds.as_numpy_iterator()

for idx, example in enumerate(ds):
  print(f'Example {idx}:')
  for key, val in example.items():
    print(f'{key}: {val}')
  print()

Downloading and preparing dataset 35.08 MiB (download: 35.08 MiB, generated: 11.33 MiB, total: 46.41 MiB) to /root/tensorflow_datasets/mtnt/en-fr/1.0.0...
Dl Completed...: 0 url [00:00, ? url/s]
Dl Size...: 0 MiB [00:00, ? MiB/s]
Extraction completed...: 0 file [00:00, ? file/s]
Generating splits...:   0%|          | 0/3 [00:00<?, ? splits/s]
Generating train examples...:   0%|          | 0/35692 [00:00<?, ? examples/s]
Shuffling /root/tensorflow_datasets/mtnt/en-fr/1.0.0.incompleteJLH33K/mtnt-train.tfrecord*...:   0%|          …
Generating test examples...:   0%|          | 0/1020 [00:00<?, ? examples/s]
Shuffling /root/tensorflow_datasets/mtnt/en-fr/1.0.0.incompleteJLH33K/mtnt-test.tfrecord*...:   0%|          |…
Generating valid examples...:   0%|          | 0/811 [00:00<?, ? examples/s]
Shuffling /root/tensorflow_datasets/mtnt/en-fr/1.0.0.incompleteJLH33K/mtnt-valid.tfrecord*...:   0%|          …
Dataset mtnt downloaded and prepared to /root/tensorflow_datasets/mtnt/en-fr/1.0.0. Subsequent calls will reuse this data.
Example 0:
dst: b'Le groupe de " toutes les \xc3\xa9toiles potentielles de la conf\xc3\xa9rence de l\'Est mais qui ne s\'en sortent pas dans le groupe de l\'Ouest ".'
src: b'The group of \xe2\x80\x9ceastern conference potential all stars but not making it in the West\xe2\x80\x9d group.'

Example 1:
dst: b"Kameron est-elle un peu aigrie de son manque de temps \xc3\xa0 l'\xc3\xa9cran ?"
src: b'Is Kameron a Little Salty About Her Lack of Air Time?'

載入使用 sentencepiece.SentencePieceProcessor 建構的 Gemma 權杖化工具：

vocab = spm.SentencePieceProcessor()
vocab.Load(TOKENIZER_PATH)

True

自訂 SentencePieceProcessor 用於英文到法文的翻譯工作。由於會微調 RecurrentGemma (Griffin) 模型的英文部分，因此您必須進行調整，例如：

  • 輸入前置字元：在每個輸入值中加入共同的前置字元，代表翻譯工作。舉例來說，您可以在提示中使用 Translate this into French: [INPUT_SENTENCE] 等前置字串。

  • 翻譯開始後置字串：在每個提示結尾加上後置字串，可指示 Gemma 模型開始轉譯程序的確切時間。這樣工作應該就會有一行。

  • 語言模型符記：RecurrentGemma (Griffin) 模型預期每個序列的開頭都會有「序列的開頭」符記。同樣地，您必須在每個訓練範例的結尾處新增「序列結尾」符記。

在 SentencePieceProcessor 周圍建構自訂包裝函式，如下所示：

class GriffinTokenizer:
  """A custom wrapper around a SentencePieceProcessor."""

  def __init__(self, spm_processor: spm.SentencePieceProcessor):
    self._spm_processor = spm_processor

  @property
  def pad_id(self) -> int:
    """Fast access to the pad ID."""
    return self._spm_processor.pad_id()

  def tokenize(
      self,
      example: str | bytes,
      prefix: str = '',
      suffix: str = '',
      add_eos: bool = True,
  ) -> jax.Array:
    """
    A tokenization function.

    Args:
      example: Input string to tokenize.
      prefix:  Prefix to add to the input string.
      suffix:  Suffix to add to the input string.
      add_eos: If True, add an end of sentence token at the end of the output
               sequence.
    Returns:
      Tokens corresponding to the input string.
    """
    int_list = [self._spm_processor.bos_id()]
    int_list.extend(self._spm_processor.EncodeAsIds(prefix + example + suffix))
    if add_eos:
      int_list.append(self._spm_processor.eos_id())

    return jnp.array(int_list, dtype=jnp.int32)

  def tokenize_tf_op(
      self,
      str_tensor: tf.Tensor,
      prefix: str = '',
      suffix: str = '',
      add_eos: bool = True,
  ) -> tf.Tensor:
    """A TensforFlow operator for the `tokenize` function."""
    encoded = tf.numpy_function(
        self.tokenize,
        [str_tensor, prefix, suffix, add_eos],
        tf.int32)
    encoded.set_shape([None])
    return encoded

  def to_string(self, tokens: jax.Array) -> str:
    """Convert an array of tokens to a string."""
    return self._spm_processor.EncodeIds(tokens.tolist())

如要試用這項功能，請將新的自訂 GriffinTokenizer 執行個體化，然後套用至 MTNT 資料集的小樣本：

def tokenize_source(tokenizer, example: tf.Tensor):
  return tokenizer.tokenize_tf_op(
      example,
      prefix='Translate this into French:\n',
      suffix='\n',
      add_eos=False
  )
def tokenize_destination(tokenizer, example: tf.Tensor):
  return tokenizer.tokenize_tf_op(example, add_eos=True)

tokenizer = GriffinTokenizer(vocab)

ds = tfds.load("mtnt/en-fr",split="train")
ds = ds.take(2)
ds = ds.map(lambda x: {
    'src': tokenize_source(tokenizer, x['src']),
    'dst': tokenize_destination(tokenizer, x['dst'])
  })
ds = ds.as_numpy_iterator()

for idx, example in enumerate(ds):
  print(f'Example {idx}:')
  for key, val in example.items():
    print(f'{key}: {val}')
  print()

Example 0:
src: [     2  49688    736   1280   6987 235292    108    651   2778    576
   1080 104745  11982   5736    832   8995    901    780   3547    665
    575    573   4589 235369   2778 235265    108]
dst: [     2   2025  29653    581    664  16298   1437  55563  41435   7840
    581    683 111452    581    533 235303   9776   4108   2459    679
    485 235303    479   6728    579   1806   2499    709  29653    581
    533 235303 101323  16054      1]

Example 1:
src: [     2  49688    736   1280   6987 235292    108   2437  87150    477
    476  11709 230461   8045   3636  40268    576   4252   4897 235336
    108]
dst: [     2 213606    477   1455 235290   3510    748   8268 191017   2809
    581   2032  69972    581  11495   1305    533 235303  65978   1654
      1]

建構整個 MTNT 資料集的資料載入器：

@chex.dataclass(frozen=True)
class TrainingInput:
  # Input tokens provided to the model.
  input_tokens: jax.Array

  # A mask that determines which tokens contribute to the target loss
  # calculation.
  target_mask: jax.Array

class DatasetSplit(enum.Enum):
  TRAIN = 'train'
  VALIDATION = 'valid'


class MTNTDatasetBuilder:
  """A data loader for the MTNT dataset."""

  N_ITEMS = {DatasetSplit.TRAIN: 35_692, DatasetSplit.VALIDATION: 811}

  BUFFER_SIZE_SHUFFLE = 10_000
  TRANSLATION_PREFIX = 'Translate this into French:\n'
  TRANSLATION_SUFFIX = '\n'

  def __init__(self,
               tokenizer : GriffinTokenizer,
               max_seq_len: int):
    """A constructor.

    Args:
      tokenizer: The tokenizer to use.
      max_seq_len: The size of each sequence in a given batch.
    """
    self._tokenizer = tokenizer
    self._base_data = {
        DatasetSplit.TRAIN: tfds.load("mtnt/en-fr",split="train"),
        DatasetSplit.VALIDATION: tfds.load("mtnt/en-fr",split="valid"),
    }
    self._max_seq_len = max_seq_len

  def _tokenize_source(self, example: tf.Tensor):
    """A tokenization function for the source."""
    return self._tokenizer.tokenize_tf_op(
        example, prefix=self.TRANSLATION_PREFIX, suffix=self.TRANSLATION_SUFFIX,
        add_eos=False
    )

  def _tokenize_destination(self, example: tf.Tensor):
    """A tokenization function for the French translation."""
    return self._tokenizer.tokenize_tf_op(example, add_eos=True)

  def _pad_up_to_max_len(self,
                         input_tensor: tf.Tensor,
                         pad_value: int | bool,
                         ) -> tf.Tensor:
    """Pad the given tensor up to sequence length of a batch."""
    seq_len = tf.shape(input_tensor)[0]
    to_pad = tf.maximum(self._max_seq_len - seq_len, 0)
    return tf.pad(
        input_tensor, [[0, to_pad]], mode='CONSTANT', constant_values=pad_value,
    )

  def _to_training_input(
      self,
      src_tokens: jax.Array,
      dst_tokens: jax.Array,
  ) -> TrainingInput:
    """Build a training input from a tuple of source and destination tokens."""

    # The input sequence fed to the model is simply the concatenation of the
    # source and the destination.
    tokens = tf.concat([src_tokens, dst_tokens], axis=0)

    # You want to prevent the model from updating based on the source (input)
    # tokens. To achieve this, add a target mask to each input.
    q_mask = tf.zeros_like(src_tokens, dtype=tf.bool)
    a_mask = tf.ones_like(dst_tokens, dtype=tf.bool)
    mask = tf.concat([q_mask, a_mask], axis=0)

    # If the output tokens sequence is smaller than the target sequence size,
    # then pad it with pad tokens.
    tokens = self._pad_up_to_max_len(tokens, self._tokenizer.pad_id)

    # You don't want to perform the backward on the pad tokens.
    mask = self._pad_up_to_max_len(mask, False)

    return TrainingInput(input_tokens=tokens, target_mask=mask)


  def get_train_dataset(self, batch_size: int, num_epochs: int):
    """Build the training dataset."""

    # Tokenize each sample.
    ds = self._base_data[DatasetSplit.TRAIN].map(
        lambda x : (self._tokenize_source(x['src']),
                    self._tokenize_destination(x['dst']))
    )

    # Convert them to training inputs.
    ds = ds.map(lambda x, y: self._to_training_input(x, y))

    # Remove the samples which are too long.
    ds = ds.filter(lambda x: tf.shape(x.input_tokens)[0] <= self._max_seq_len)

    # Shuffle the dataset.
    ds = ds.shuffle(buffer_size=self.BUFFER_SIZE_SHUFFLE)

    # Repeat if necessary.
    ds = ds.repeat(num_epochs)

    # Build batches.
    ds = ds.batch(batch_size, drop_remainder=True)
    return ds

  def get_validation_dataset(self, batch_size: int):
    """Build the validation dataset."""

    # Same as the training dataset, but no shuffling and no repetition
    ds = self._base_data[DatasetSplit.VALIDATION].map(
        lambda x : (self._tokenize_source(x['src']),
                    self._tokenize_destination(x['dst']))
    )
    ds = ds.map(lambda x, y: self._to_training_input(x, y))
    ds = ds.filter(lambda x: tf.shape(x.input_tokens)[0] <= self._max_seq_len)
    ds = ds.batch(batch_size, drop_remainder=True)
    return ds

再次將自訂 GriffinTokenizer 執行個體化，然後套用至 MTNT 資料集，然後取樣兩個範例，以試用 MTNTDatasetBuilder：

dataset_builder = MTNTDatasetBuilder(tokenizer, max_seq_len=20)
ds = dataset_builder.get_train_dataset(3, 1)
ds = ds.take(2)
ds = ds.as_numpy_iterator()

for idx, example in enumerate(ds):
  print(f'Example {idx}:')
  for key, val in example.items():
    print(f'{key}: {val}')
  print()

WARNING:tensorflow:Mapping types may not work well with tf.nest. Prefer using MutableMapping for <class '__main__.TrainingInput'>
WARNING:tensorflow:Mapping types may not work well with tf.nest. Prefer using MutableMapping for <class '__main__.TrainingInput'>
WARNING:tensorflow:Mapping types may not work well with tf.nest. Prefer using MutableMapping for <class '__main__.TrainingInput'>
Example 0:
input_tokens: [[     2  49688    736   1280   6987 235292    108  12583    665 235265
     108      2   6151  94975   1320   6238 235265      1      0      0]
 [     2  49688    736   1280   6987 235292    108   4899  29960  11270
  108282 235265    108      2   4899  79025  11270 108282      1      0]
 [     2  49688    736   1280   6987 235292    108  26620 235265    108
       2  26620 235265      1      0      0      0      0      0      0]]
target_mask: [[False False False False False False False False False False False  True
   True  True  True  True  True  True False False]
 [False False False False False False False False False False False False
  False  True  True  True  True  True  True False]
 [False False False False False False False False False False  True  True
   True  True False False False False False False]]

Example 1:
input_tokens: [[     2  49688    736   1280   6987 235292    108    527   5174   1683
  235336    108      2 206790    581  20726    482   2208   1654      1]
 [     2  49688    736   1280   6987 235292    108  28484 235256 235336
     108      2 120500  13832   1654      1      0      0      0      0]
 [     2  49688    736   1280   6987 235292    108 235324 235304   2705
  235265    108      2 235324 235304  19963 235265      1      0      0]]
target_mask: [[False False False False False False False False False False False False
   True  True  True  True  True  True  True  True]
 [False False False False False False False False False False False  True
   True  True  True  True False False False False]
 [False False False False False False False False False False False False
   True  True  True  True  True  True False False]]

設定模型

您必須先設定 Gemma 模型，才能開始微調。

使用 recurrentgemma.jax.utils.load_parameters 方法載入 RecurrentGemma (Griffin) 模型查核點：

params =  recurrentgemma.load_parameters(CKPT_PATH, "single_device")

如要從 RecurrentGemma 模型查核點自動載入正確的設定，請使用 recurrentgemma.GriffinConfig.from_flax_params_or_variables：

config = recurrentgemma.GriffinConfig.from_flax_params_or_variables(params)

使用 recurrentgemma.jax.Griffin 將 Griffin 模型例項化：

model = recurrentgemma.Griffin(config)

在 RecurrentGemma 模型查核點/權重和符記化器上方建立sampler recurrentgemma.jax.Sampler，檢查模型是否可以進行翻譯：

sampler = recurrentgemma.Sampler(model=model, vocab=vocab, params=params)

微調模型

在本節中，您將進行以下作業：

  • 使用 gemma.transformer.Transformer 類別建立正向傳遞和損失函式。
  • 建立符記的位置和注意力遮罩向量
  • 使用 Flax 建構訓練步數函式。
  • 建立驗證步驟而不使用反向傳遞。
  • 建立訓練迴圈。
  • 微調 Gemma 模型。

使用 recurrentgemma.jax.griffin.Griffin 類別定義正向傳遞和損失函式。RecurrentGemma Griffin 繼承自 flax.linen.Module，並提供兩種基本方法：

  • init：初始化模型的參數。
  • apply：使用指定的一組參數執行模型的 __call__ 函式。

由於您使用的是預先訓練的 Gemma 權重，因此不需要使用 init 函式。

def forward_and_loss_fn(
    params,
    *,
    model: recurrentgemma.Griffin,
    input_tokens: jax.Array,            # Shape [B, L]
    input_mask: jax.Array,              # Shape [B, L]
    positions: jax.Array,               # Shape [B, L]
) -> jax.Array:
  """Forward pass and loss function.

  Args:
    params: model's input parameters.
    model: Griffin model to call.
    input_tokens: input tokens sequence, shape [B, L].
    input_mask: tokens to ignore when computing the loss, shape [B, L].
    positions: relative position of each token, shape [B, L].

  Returns:
    Softmax cross-entropy loss for the next-token prediction task.
  """
  batch_size = input_tokens.shape[0]
  # Forward pass on the input data.
  # No attention cache is needed here.
  # Exclude the last step as it does not appear in the targets.
  logits, _ = model.apply(
        {"params": params},
        tokens=input_tokens[:, :-1],
        segment_pos=positions[:, :-1],
        cache=None,
    )

  # Similarly, the first token cannot be predicteds.
  target_tokens = input_tokens[:, 1:]
  target_mask = input_mask[:, 1:]

  # Convert the target labels into one-hot encoded vectors.
  one_hot = jax.nn.one_hot(target_tokens, logits.shape[-1])

  # Don't update on unwanted tokens.
  one_hot = one_hot * target_mask.astype(one_hot.dtype)[...,None]

  # Normalization factor.
  norm_factor = batch_size * (jnp.sum(target_mask) + 1e-8)

  # Return the negative log-likelihood loss (NLL) function.
  return -jnp.sum(jax.nn.log_softmax(logits) * one_hot) / norm_factor

建構 train_step 函式，執行反向傳遞，並據此更新模型的參數，其中：

Params = Mapping[str, Any]

def get_positions(example: jax.Array, pad_id : int) -> jax.Array:
  """Builds the position vector from the given tokens."""
  pad_mask = example != pad_id
  positions = jnp.cumsum(pad_mask, axis=-1)
  # Subtract one for all positions from the first valid one as they are
  # 0-indexed
  positions = positions - (positions >= 1)
  return positions

@functools.partial(
    jax.jit,
    static_argnames=['model', 'optimizer'],
    donate_argnames=['params', 'opt_state'],
)
def train_step(
    model: recurrentgemma.Griffin,
    params: Params,
    optimizer: optax.GradientTransformation,
    opt_state: optax.OptState,
    pad_id: int,
    example: TrainingInput,
) -> tuple[jax.Array, Params, optax.OptState]:
  """The train step.

  Args:
    model: The RecurrentGemma (Griffin) model.
    params: The model's input parameters.
    optimizer: The Optax optimizer to use.
    opt_state: The input optimizer's state.
    pad_id: The ID of the pad token.
    example: The input batch.

  Returns:
    Training loss, updated parameters, updated optimizer state.
  """

  positions = get_positions(example.input_tokens, pad_id)

  # Forward and backward passes.
  train_loss, grads = jax.value_and_grad(forward_and_loss_fn)(
      params,
      model=model,
      input_tokens=example.input_tokens,
      input_mask=example.target_mask,
      positions=positions,
  )
  # Update the parameters.
  updates, opt_state = optimizer.update(grads, opt_state, params)
  params = optax.apply_updates(params, updates)

  return train_loss, params, opt_state

建構 validation_step 函式，但不使用反向傳遞：

@functools.partial(jax.jit, static_argnames=['model'])
def validation_step(
    model: recurrentgemma.Griffin,
    params: Params,
    pad_id: int,
    example: TrainingInput,
) -> jax.Array:
  return forward_and_loss_fn(
      params,
      model=model,
      input_tokens=example.input_tokens,
      input_mask=example.target_mask,
      positions=get_positions(example.input_tokens, pad_id),
  )

定義訓練迴圈：

def train_loop(
    model: recurrentgemma.Griffin,
    params: Params,
    optimizer: optax.GradientTransformation,
    train_ds: Iterator[TrainingInput],
    validation_ds: Iterator[TrainingInput],
    num_steps: int | None = None,
    eval_every_n: int = 20,
):
  opt_state = jax.jit(optimizer.init)(params)

  step_counter = 0
  avg_loss=0

  # The first round of the validation loss.
  n_steps_eval = 0
  eval_loss = 0
  for val_example in validation_ds.as_numpy_iterator():
    eval_loss += validation_step(
        model, params, dataset_builder._tokenizer.pad_id, val_example
    )
    n_steps_eval += 1
  print(f"Start, validation loss: {eval_loss/n_steps_eval}")

  for train_example in train_ds:
    train_loss, params, opt_state = train_step(
        model=model,
        params=params,
        optimizer=optimizer,
        opt_state=opt_state,
        pad_id=dataset_builder._tokenizer.pad_id,
        example=train_example,
    )

    step_counter += 1
    avg_loss += train_loss
    if step_counter % eval_every_n == 0:
      eval_loss = 0

      n_steps_eval = 0
      val_iterator = validation_ds.as_numpy_iterator()
      for val_example in val_iterator:
        eval_loss += validation_step(
            model,
            params,
            dataset_builder._tokenizer.pad_id,
            val_example,
        )
        n_steps_eval +=1
      avg_loss /= eval_every_n
      eval_loss /= n_steps_eval
      print(f"STEP {step_counter} training loss: {avg_loss} - eval loss: {eval_loss}")
      avg_loss=0
    if num_steps is not None and step_counter > num_steps:
      break
  return params

您必須在此選擇 (Optax) 最佳化工具。如果是記憶體較小的裝置，則應使用 SGD，因為 SGD 的記憶體使用量較低。如要取得最佳微調效能，請嘗試 Adam-W。此筆記本中特定工作在本範例中為 2b-it 查核點提供每個最佳化器的最佳超參數。

def griffin_weight_decay_mask(params_like: optax.Params) -> Any:
  # Don't put weight decay on the RGLRU, the embeddings and any biases
  def enable_weight_decay(path: list[Any], _: Any) -> bool:
    # Parameters in the LRU and embedder
    path = [dict_key.key for dict_key in path]
    if 'rg_lru' in path or 'embedder' in path:
      return False
    # All biases and scales
    if path[-1] in ('b', 'scale'):
      return False
    return True

  return jax.tree_util.tree_map_with_path(enable_weight_decay, params_like)

optimizer_choice = "sgd"

if optimizer_choice == "sgd":
  optimizer = optax.sgd(learning_rate=1e-3)
  num_steps = 300
elif optimizer_choice == "adamw":
  optimizer = optax.adamw(
        learning_rate=1e-4,
        b2=0.96,
        eps=1e-8,
        weight_decay=0.1,
        mask=griffin_weight_decay_mask,
    )
  num_steps = 100
else:
  raise ValueError(f"Unknown optimizer: {optimizer_choice}")

準備訓練和驗證資料集：

# Choose a small sequence length size, so that everything fits in memory.
num_epochs = 1
batch_size = 1
sequence_length = 32

# Make the dataset builder.
tokenizer = GriffinTokenizer(vocab)
dataset_builder= MTNTDatasetBuilder(tokenizer, sequence_length + 1)

# Build the training dataset.
train_ds = dataset_builder.get_train_dataset(
    batch_size=batch_size,
    num_epochs=num_epochs,
).as_numpy_iterator()

# Build the validation dataset, with a limited number of samples for this demo.
validation_ds = dataset_builder.get_validation_dataset(
    batch_size=batch_size,
).take(50)

開始以有限的步驟 (num_steps) 開始微調 RecurrentGemma (Griffin) 模型：

trained_params = train_loop(
    model=model,
    params=params,
    optimizer=optimizer,
    train_ds=train_ds,
    validation_ds=validation_ds,
    num_steps=num_steps,
)

Start, validation loss: 7.894117832183838
/usr/local/lib/python3.10/dist-packages/jax/_src/interpreters/mlir.py:920: UserWarning: Some donated buffers were not usable: ShapedArray(int32[1,33]), ShapedArray(bool[1,33]), ShapedArray(int32[], weak_type=True).
See an explanation at https://jax.readthedocs.io/en/latest/faq.html#buffer_donation.
  warnings.warn("Some donated buffers were not usable:"
STEP 20 training loss: 4.592616081237793 - eval loss: 2.847407102584839
STEP 40 training loss: 2.7537424564361572 - eval loss: 2.9258534908294678
STEP 60 training loss: 2.835618257522583 - eval loss: 2.4382340908050537
STEP 80 training loss: 2.6322107315063477 - eval loss: 2.3696839809417725
STEP 100 training loss: 1.8703256845474243 - eval loss: 2.355681896209717
STEP 120 training loss: 2.7280433177948 - eval loss: 2.4059958457946777
STEP 140 training loss: 2.3047447204589844 - eval loss: 2.083082914352417
STEP 160 training loss: 2.3432137966156006 - eval loss: 2.095074415206909
STEP 180 training loss: 2.1081202030181885 - eval loss: 2.006460189819336
STEP 200 training loss: 2.5359647274017334 - eval loss: 1.9667452573776245
STEP 220 training loss: 2.202195644378662 - eval loss: 1.9440618753433228
STEP 240 training loss: 2.756615400314331 - eval loss: 2.1073737144470215
STEP 260 training loss: 2.5128934383392334 - eval loss: 2.117241859436035
STEP 280 training loss: 2.73045015335083 - eval loss: 1.9159646034240723
STEP 300 training loss: 2.0918595790863037 - eval loss: 1.9742532968521118

每個步數都會導致訓練損失和驗證損失都會下降。

為確保輸入內容與訓練格式相符，請使用前置字串 Translate this into French:\n 並在結尾加上換行字元。這會指示模型開始翻譯。

sampler.params = trained_params
output = sampler(
    ["Translate this into French:\nHello, my name is Morgane.\n"],
    total_generation_steps=100,
)
print(output.text[0])

/usr/local/lib/python3.10/dist-packages/jax/_src/interpreters/mlir.py:920: UserWarning: Some donated buffers were not usable: ShapedArray(int32[1,16]).
See an explanation at https://jax.readthedocs.io/en/latest/faq.html#buffer_donation.
  warnings.warn("Some donated buffers were not usable:"
Mais je m'appelle Morgane.

瞭解詳情