使用 SASRec 进行序列检索

!pip install -q keras-rs

import os

os.environ["KERAS_BACKEND"] = "jax"  # `"tensorflow"`/`"torch"`

import collections
import os

import keras
import keras_hub
import numpy as np
import pandas as pd
import tensorflow as tf  # Needed only for the dataset
from keras import ops

import keras_rs

DATA_DIR = "./raw/data/"

# MovieLens-specific variables
MOVIELENS_1M_URL = "https://files.grouplens.org/datasets/movielens/ml-1m.zip"
MOVIELENS_ZIP_HASH = "a6898adb50b9ca05aa231689da44c217cb524e7ebd39d264c56e2832f2c54e20"

RATINGS_FILE_NAME = "ratings.dat"
MOVIES_FILE_NAME = "movies.dat"

# Data processing args
MAX_CONTEXT_LENGTH = 200
MIN_SEQUENCE_LENGTH = 3
PAD_ITEM_ID = 0

RATINGS_DATA_COLUMNS = ["UserID", "MovieID", "Rating", "Timestamp"]
MOVIES_DATA_COLUMNS = ["MovieID", "Title", "Genres"]
MIN_RATING = 2

# Training/model args picked from SASRec paper
BATCH_SIZE = 128
NUM_EPOCHS = 10
LEARNING_RATE = 0.001

NUM_LAYERS = 2
NUM_HEADS = 1
HIDDEN_DIM = 50
DROPOUT = 0.2

# Download the MovieLens dataset.
if not os.path.exists(DATA_DIR):
    os.makedirs(DATA_DIR)

path_to_zip = keras.utils.get_file(
    fname="ml-1m.zip",
    origin=MOVIELENS_1M_URL,
    file_hash=MOVIELENS_ZIP_HASH,
    hash_algorithm="sha256",
    extract=True,
    cache_dir=DATA_DIR,
)
movielens_extracted_dir = os.path.join(
    os.path.dirname(path_to_zip),
    "ml-1m_extracted",
    "ml-1m",
)


# Read the dataset.
def read_data(data_directory, min_rating=None):
    """Read movielens ratings.dat and movies.dat file
    into dataframe.
    """

    ratings_df = pd.read_csv(
        os.path.join(data_directory, RATINGS_FILE_NAME),
        sep="::",
        names=RATINGS_DATA_COLUMNS,
        encoding="unicode_escape",
    )
    ratings_df["Timestamp"] = ratings_df["Timestamp"].apply(int)

    # Remove movies with `rating < min_rating`.
    if min_rating is not None:
        ratings_df = ratings_df[ratings_df["Rating"] >= min_rating]

    movies_df = pd.read_csv(
        os.path.join(data_directory, MOVIES_FILE_NAME),
        sep="::",
        names=MOVIES_DATA_COLUMNS,
        encoding="unicode_escape",
    )
    return ratings_df, movies_df


ratings_df, movies_df = read_data(
    data_directory=movielens_extracted_dir, min_rating=MIN_RATING
)

# Need to know #movies so as to define embedding layers.
movies_count = movies_df["MovieID"].max()

def get_movie_sequence_per_user(ratings_df):
    """Get movieID sequences for every user."""
    sequences = collections.defaultdict(list)

    for user_id, movie_id, rating, timestamp in ratings_df.values:
        sequences[user_id].append(
            {
                "movie_id": movie_id,
                "timestamp": timestamp,
                "rating": rating,
            }
        )

    # Sort movie sequences by timestamp for every user.
    for user_id, context in sequences.items():
        context.sort(key=lambda x: x["timestamp"])
        sequences[user_id] = context

    return sequences


sequences = get_movie_sequence_per_user(ratings_df)

def format_data(sequences):
    examples = {
        "sequence": [],
        "negative_sequence": [],
    }

    for user_id in sequences:
        sequence = [int(d["movie_id"]) for d in sequences[user_id]]

        # Get negative sequence.
        def random_negative_item_id(low, high, positive_lst):
            sampled = np.random.randint(low=low, high=high)
            while sampled in positive_lst:
                sampled = np.random.randint(low=low, high=high)
            return sampled

        negative_sequence = [
            random_negative_item_id(1, movies_count + 1, sequence)
            for _ in range(len(sequence))
        ]

        examples["sequence"].append(np.array(sequence))
        examples["negative_sequence"].append(np.array(negative_sequence))

    examples["sequence"] = tf.ragged.constant(examples["sequence"])
    examples["negative_sequence"] = tf.ragged.constant(examples["negative_sequence"])

    return examples


examples = format_data(sequences)
ds = tf.data.Dataset.from_tensor_slices(examples).batch(BATCH_SIZE)

def _preprocess(example, train=False):
    sequence = example["sequence"]
    negative_sequence = example["negative_sequence"]

    if train:
        sequence = example["sequence"][..., :-1]
        negative_sequence = example["negative_sequence"][..., :-1]

    batch_size = tf.shape(sequence)[0]

    if not train:
        # Loss computed only on last token.
        sample_weight = tf.zeros_like(sequence, dtype="float32")[..., :-1]
        sample_weight = tf.concat(
            [sample_weight, tf.ones((batch_size, 1), dtype="float32")], axis=1
        )

    # Truncate/pad sequence. +1 to account for truncation later.
    sequence = sequence.to_tensor(
        shape=[batch_size, MAX_CONTEXT_LENGTH + 1], default_value=PAD_ITEM_ID
    )
    negative_sequence = negative_sequence.to_tensor(
        shape=[batch_size, MAX_CONTEXT_LENGTH + 1], default_value=PAD_ITEM_ID
    )
    if train:
        sample_weight = tf.cast(sequence != PAD_ITEM_ID, dtype="float32")
    else:
        sample_weight = sample_weight.to_tensor(
            shape=[batch_size, MAX_CONTEXT_LENGTH + 1], default_value=0
        )

    example = (
        {
            # last token does not have a next token
            "item_ids": sequence[..., :-1],
            # padding mask for controlling attention mask
            "padding_mask": (sequence != PAD_ITEM_ID)[..., :-1],
        },
        {
            "positive_sequence": sequence[
                ..., 1:
            ],  # 0th token's label will be 1st token, and so on
            "negative_sequence": negative_sequence[..., 1:],
        },
        sample_weight[..., 1:],  # loss will not be computed on pad tokens
    )
    return example


def preprocess_train(examples):
    return _preprocess(examples, train=True)


def preprocess_val(examples):
    return _preprocess(examples, train=False)


train_ds = ds.map(preprocess_train)
val_ds = ds.map(preprocess_val)

for batch in train_ds.take(1):
    print(batch)

for batch in val_ds.take(1):
    print(batch)

 loss = - (y_true * np.log(y_pred) + (1 - y_true) * np.log(1 - y_pred))

loss = -np.log(positive_logits)

class SasRec(keras.Model):
    def __init__(
        self,
        vocabulary_size,
        num_layers,
        num_heads,
        hidden_dim,
        dropout=0.0,
        max_sequence_length=100,
        dtype=None,
        **kwargs,
    ):
        super().__init__(dtype=dtype, **kwargs)

        # ======== Layers ========

        # === Embeddings ===
        self.item_embedding = keras_hub.layers.ReversibleEmbedding(
            input_dim=vocabulary_size,
            output_dim=hidden_dim,
            embeddings_initializer="glorot_uniform",
            embeddings_regularizer=keras.regularizers.l2(0.001),
            dtype=dtype,
            name="item_embedding",
        )
        self.position_embedding = keras_hub.layers.PositionEmbedding(
            initializer="glorot_uniform",
            sequence_length=max_sequence_length,
            dtype=dtype,
            name="position_embedding",
        )
        self.embeddings_add = keras.layers.Add(
            dtype=dtype,
            name="embeddings_add",
        )
        self.embeddings_dropout = keras.layers.Dropout(
            dropout,
            dtype=dtype,
            name="embeddings_dropout",
        )

        # === Decoder layers ===
        self.transformer_layers = []
        for i in range(num_layers):
            self.transformer_layers.append(
                keras_hub.layers.TransformerDecoder(
                    intermediate_dim=hidden_dim,
                    num_heads=num_heads,
                    dropout=dropout,
                    layer_norm_epsilon=1e-05,
                    # SASRec uses ReLU, although GeLU might be a better option
                    activation="relu",
                    kernel_initializer="glorot_uniform",
                    normalize_first=True,
                    dtype=dtype,
                    name=f"transformer_layer_{i}",
                )
            )

        # === Final layer norm ===
        self.layer_norm = keras.layers.LayerNormalization(
            axis=-1,
            epsilon=1e-8,
            dtype=dtype,
            name="layer_norm",
        )

        # === Retrieval ===
        # The layer that performs the retrieval.
        self.retrieval = keras_rs.layers.BruteForceRetrieval(k=10, return_scores=False)

        # === Loss ===
        self.loss_fn = keras.losses.BinaryCrossentropy(from_logits=True, reduction=None)

        # === Attributes ===
        self.vocabulary_size = vocabulary_size
        self.num_layers = num_layers
        self.num_heads = num_heads
        self.hidden_dim = hidden_dim
        self.dropout = dropout
        self.max_sequence_length = max_sequence_length

    def _get_last_non_padding_token(self, tensor, padding_mask):
        valid_token_mask = ops.logical_not(padding_mask)
        seq_lengths = ops.sum(ops.cast(valid_token_mask, "int32"), axis=1)
        last_token_indices = ops.maximum(seq_lengths - 1, 0)

        indices = ops.expand_dims(last_token_indices, axis=(-2, -1))
        gathered_tokens = ops.take_along_axis(tensor, indices, axis=1)
        last_token_embedding = ops.squeeze(gathered_tokens, axis=1)

        return last_token_embedding

    def build(self, input_shape):
        embedding_shape = list(input_shape) + [self.hidden_dim]

        # Model
        self.item_embedding.build(input_shape)
        self.position_embedding.build(embedding_shape)

        self.embeddings_add.build((embedding_shape, embedding_shape))
        self.embeddings_dropout.build(embedding_shape)

        for transformer_layer in self.transformer_layers:
            transformer_layer.build(decoder_sequence_shape=embedding_shape)

        self.layer_norm.build(embedding_shape)

        # Retrieval
        self.retrieval.candidate_embeddings = self.item_embedding.embeddings
        self.retrieval.build(input_shape)

        # Chain to super
        super().build(input_shape)

    def call(self, inputs, training=False):
        item_ids, padding_mask = inputs["item_ids"], inputs["padding_mask"]

        x = self.item_embedding(item_ids)
        position_embedding = self.position_embedding(x)
        x = self.embeddings_add((x, position_embedding))
        x = self.embeddings_dropout(x)

        for transformer_layer in self.transformer_layers:
            x = transformer_layer(x, decoder_padding_mask=padding_mask)

        item_sequence_embedding = self.layer_norm(x)
        result = {"item_sequence_embedding": item_sequence_embedding}

        # At inference, perform top-k retrieval.
        if not training:
            # need to extract last non-padding token.
            last_item_embedding = self._get_last_non_padding_token(
                item_sequence_embedding, padding_mask
            )
            result["predictions"] = self.retrieval(last_item_embedding)

        return result

    def compute_loss(self, x, y, y_pred, sample_weight, training=False):
        item_sequence_embedding = y_pred["item_sequence_embedding"]
        y_positive_sequence = y["positive_sequence"]
        y_negative_sequence = y["negative_sequence"]

        # Embed positive, negative sequences.
        positive_sequence_embedding = self.item_embedding(y_positive_sequence)
        negative_sequence_embedding = self.item_embedding(y_negative_sequence)

        # Logits
        positive_logits = ops.sum(
            ops.multiply(positive_sequence_embedding, item_sequence_embedding),
            axis=-1,
        )
        negative_logits = ops.sum(
            ops.multiply(negative_sequence_embedding, item_sequence_embedding),
            axis=-1,
        )
        logits = ops.concatenate([positive_logits, negative_logits], axis=1)

        # Labels
        labels = ops.concatenate(
            [
                ops.ones_like(positive_logits),
                ops.zeros_like(negative_logits),
            ],
            axis=1,
        )

        # sample weights
        sample_weight = ops.concatenate(
            [sample_weight, sample_weight],
            axis=1,
        )

        loss = self.loss_fn(
            y_true=ops.expand_dims(labels, axis=-1),
            y_pred=ops.expand_dims(logits, axis=-1),
            sample_weight=sample_weight,
        )
        loss = ops.divide_no_nan(ops.sum(loss), ops.sum(sample_weight))

        return loss

    def compute_output_shape(self, inputs_shape):
        return list(inputs_shape) + [self.hidden_dim]

model = SasRec(
    vocabulary_size=movies_count + 1,
    num_layers=NUM_LAYERS,
    num_heads=NUM_HEADS,
    hidden_dim=HIDDEN_DIM,
    dropout=DROPOUT,
    max_sequence_length=MAX_CONTEXT_LENGTH,
)

# Training
output = model(
    inputs={
        "item_ids": ops.ones((2, MAX_CONTEXT_LENGTH), dtype="int32"),
        "padding_mask": ops.ones((2, MAX_CONTEXT_LENGTH), dtype="bool"),
    },
    training=True,
)
print(output["item_sequence_embedding"].shape)

# Inference
output = model(
    inputs={
        "item_ids": ops.ones((2, MAX_CONTEXT_LENGTH), dtype="int32"),
        "padding_mask": ops.ones((2, MAX_CONTEXT_LENGTH), dtype="bool"),
    },
    training=False,
)
print(output["predictions"].shape)

model.compile(
    optimizer=keras.optimizers.Adam(learning_rate=LEARNING_RATE, beta_2=0.98),
)
model.fit(
    x=train_ds,
    validation_data=val_ds,
    epochs=NUM_EPOCHS,
)

movie_id_to_movie_title = dict(zip(movies_df["MovieID"], movies_df["Title"]))
movie_id_to_movie_title[0] = ""  # Because id 0 is not in the dataset.

for ele in val_ds.unbatch().take(1):
    test_sample = ele[0]
    test_sample["item_ids"] = tf.expand_dims(test_sample["item_ids"], axis=0)
    test_sample["padding_mask"] = tf.expand_dims(test_sample["padding_mask"], axis=0)

movie_sequence = np.array(test_sample["item_ids"])[0]
for movie_id in movie_sequence:
    if movie_id == 0:
        continue
    print(movie_id_to_movie_title[movie_id], end="; ")
print()

predictions = model.predict(test_sample)["predictions"]
predictions = keras.ops.convert_to_numpy(predictions)

for movie_id in predictions[0]:
    print(movie_id_to_movie_title[movie_id])