File size: 2,018 Bytes
188f0cf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
"""Common interface for every recommender model.

Trainer and evaluator depend only on this class — concrete models (MF,
TwoTower, future GNNs) plug in without any changes upstream.
"""

from __future__ import annotations

from abc import ABC, abstractmethod

import torch
from torch import Tensor, nn


class BaseRecommender(nn.Module, ABC):
    """Abstract base. Subclasses implement `score` and `score_all_items`."""

    num_users: int
    num_items: int

    def __init__(self, num_users: int, num_items: int) -> None:
        super().__init__()
        if num_users < 1 or num_items < 1:
            raise ValueError("num_users and num_items must be >= 1")
        self.num_users = int(num_users)
        self.num_items = int(num_items)

    @abstractmethod
    def score(self, users: Tensor, items: Tensor) -> Tensor:
        """Score (user, item) pairs.

        Args:
            users: int64 tensor of shape [B] or broadcastable to items.
            items: int64 tensor of shape [B] or [B, K].

        Returns:
            float tensor with shape matching `items`.
        """

    @abstractmethod
    def score_all_items(self, users: Tensor) -> Tensor:
        """Score a batch of users against every item in the catalog.

        Args:
            users: int64 tensor of shape [B].

        Returns:
            float tensor of shape [B, num_items].
        """

    def forward(
        self, users: Tensor, pos_items: Tensor, neg_items: Tensor
    ) -> tuple[Tensor, Tensor]:
        """Shared forward used by BPR training.

        Args:
            users: [B] int64.
            pos_items: [B] int64.
            neg_items: [B, K] int64.

        Returns:
            (pos_scores [B], neg_scores [B, K]).
        """
        pos_scores = self.score(users, pos_items)
        # Broadcast users along the K dim before scoring negatives.
        users_expanded = users.unsqueeze(-1).expand_as(neg_items)
        neg_scores = self.score(users_expanded, neg_items)
        return pos_scores, neg_scores