src/models/two_tower.py

import numpy as np
import pandas as pd
import pickle
from pathlib import Path

import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader


# ── Dataset ──────────────────────────────────────────────────────────────────

class InteractionDataset(Dataset):
    """
    Feeds (user_features, item_features, label) triplets to the model.
    Positive label  = 1 (observed interaction)
    Negative label  = 0 (randomly sampled unobserved pair)
    We use 1:1 negative sampling ratio.
    """

    def __init__(self, train_df, user_feats_norm, item_feats_norm, neg_ratio=1):
        self.user_feat_cols = [c for c in user_feats_norm.columns if c != "user_idx"]
        self.item_feat_cols = [c for c in item_feats_norm.columns if c != "item_idx"]

        self.user_feat_map = (
            user_feats_norm.set_index("user_idx")[self.user_feat_cols]
            .astype(np.float32).to_dict("index")
        )
        self.item_feat_map = (
            item_feats_norm.set_index("item_idx")[self.item_feat_cols]
            .astype(np.float32).to_dict("index")
        )

        self.all_items  = list(item_feats_norm["item_idx"].values)
        self.neg_ratio  = neg_ratio

        # Build positive pairs
        self.positives  = train_df[["user_idx","item_idx"]].values.tolist()

        # Build user history for negative sampling
        self.user_history = (
            train_df.groupby("user_idx")["item_idx"].apply(set).to_dict()
        )

        # Pre-sample negatives for speed
        self.negatives = self._sample_negatives()

        self.pairs  = (
            [(u, i, 1.0) for u, i in self.positives] +
            [(u, i, 0.0) for u, i in self.negatives]
        )

    def _sample_negatives(self):
        negs = []
        for user_idx, item_idx in self.positives:
            seen = self.user_history.get(user_idx, set())
            for _ in range(self.neg_ratio):
                neg = np.random.choice(self.all_items)
                while neg in seen:
                    neg = np.random.choice(self.all_items)
                negs.append((user_idx, neg))
        return negs

    def __len__(self):
        return len(self.pairs)

    def __getitem__(self, idx):
        user_idx, item_idx, label = self.pairs[idx]

        u_feats = list(self.user_feat_map.get(
            user_idx, {c: 0.0 for c in self.user_feat_cols}
        ).values())
        i_feats = list(self.item_feat_map.get(
            item_idx, {c: 0.0 for c in self.item_feat_cols}
        ).values())

        return (
            torch.tensor(u_feats, dtype=torch.float32),
            torch.tensor(i_feats, dtype=torch.float32),
            torch.tensor(label,   dtype=torch.float32),
        )


# ── Model ─────────────────────────────────────────────────────────────────────

class TwoTowerModel(nn.Module):
    """
    Two-Tower (Dual Encoder) architecture.

    User tower  : user features → dense embedding
    Item tower  : item features → dense embedding
    Score       : cosine similarity of the two embeddings

    Why two towers?
    At serving time, item embeddings are pre-computed offline.
    Only the user tower runs at request time → fast inference.
    This is how YouTube, Pinterest, and most large-scale recommenders work.

    Architecture:
    Input → Linear → BatchNorm → ReLU → Dropout
          → Linear → BatchNorm → ReLU → Dropout
          → Linear → L2-normalised embedding
    """

    def __init__(self, user_dim: int, item_dim: int, embedding_dim: int = 64,
                 hidden_dim: int = 128, dropout: float = 0.2):
        super().__init__()
        self.embedding_dim = embedding_dim

        self.user_tower = nn.Sequential(
            nn.Linear(user_dim,    hidden_dim),
            nn.BatchNorm1d(hidden_dim),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim,  hidden_dim),
            nn.BatchNorm1d(hidden_dim),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim,  embedding_dim),
        )

        self.item_tower = nn.Sequential(
            nn.Linear(item_dim,    hidden_dim),
            nn.BatchNorm1d(hidden_dim),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim,  hidden_dim),
            nn.BatchNorm1d(hidden_dim),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim,  embedding_dim),
        )

    def forward(self, user_feats, item_feats):
        u_emb = self.user_tower(user_feats)
        i_emb = self.item_tower(item_feats)
        # L2 normalise → cosine similarity = dot product
        u_emb = nn.functional.normalize(u_emb, dim=1)
        i_emb = nn.functional.normalize(i_emb, dim=1)
        score = (u_emb * i_emb).sum(dim=1)
        return score

    def get_user_embedding(self, user_feats):
        with torch.no_grad():
            emb = self.user_tower(user_feats)
            return nn.functional.normalize(emb, dim=1)

    def get_item_embedding(self, item_feats):
        with torch.no_grad():
            emb = self.item_tower(item_feats)
            return nn.functional.normalize(emb, dim=1)


# ── Trainer ───────────────────────────────────────────────────────────────────

class TwoTowerRecommender:
    """
    Wraps TwoTowerModel with fit / recommend interface
    matching PopularityRecommender and ALSRecommender.
    """

    def __init__(self, embedding_dim=64, hidden_dim=128,
                 dropout=0.2, lr=1e-3, epochs=10, batch_size=512):
        self.embedding_dim = embedding_dim
        self.hidden_dim    = hidden_dim
        self.dropout       = dropout
        self.lr            = lr
        self.epochs        = epochs
        self.batch_size    = batch_size
        self.model         = None
        self.device        = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.is_fitted     = False

        # Stored after fit for recommend()
        self.item_embeddings  = None
        self.user_feat_cols   = None
        self.item_feat_cols   = None
        self.user_feat_map    = None
        self.item_feats_norm  = None
        self.user_history     = {}

    def fit(self, train_df, user_feats_norm, item_feats_norm):
        print(f"Training on: {self.device}")

        dataset = InteractionDataset(train_df, user_feats_norm, item_feats_norm)
        loader  = DataLoader(dataset, batch_size=self.batch_size,
                             shuffle=True, num_workers=0)

        self.user_feat_cols  = dataset.user_feat_cols
        self.item_feat_cols  = dataset.item_feat_cols
        self.user_feat_map   = dataset.user_feat_map
        self.item_feats_norm = item_feats_norm
        self.user_history    = dataset.user_history

        user_dim = len(self.user_feat_cols)
        item_dim = len(self.item_feat_cols)

        self.model = TwoTowerModel(
            user_dim      = user_dim,
            item_dim      = item_dim,
            embedding_dim = self.embedding_dim,
            hidden_dim    = self.hidden_dim,
            dropout       = self.dropout,
        ).to(self.device)

        optimizer = torch.optim.Adam(self.model.parameters(), lr=self.lr)
        criterion = nn.BCEWithLogitsLoss()
        scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=3, gamma=0.5)

        self.model.train()
        for epoch in range(self.epochs):
            total_loss = 0.0
            for u_feats, i_feats, labels in loader:
                u_feats = u_feats.to(self.device)
                i_feats = i_feats.to(self.device)
                labels  = labels.to(self.device)

                optimizer.zero_grad()
                scores = self.model(u_feats, i_feats)
                loss   = criterion(scores, labels)
                loss.backward()
                optimizer.step()
                total_loss += loss.item()

            scheduler.step()
            avg_loss = total_loss / len(loader)
            print(f"  Epoch {epoch+1:02d}/{self.epochs} — loss: {avg_loss:.4f}")

        # Pre-compute all item embeddings for fast inference
        self._precompute_item_embeddings()
        self.is_fitted = True
        print("Training complete.")
        return self

    def _precompute_item_embeddings(self):
        """Compute and cache all item embeddings once after training."""
        self.model.eval()
        item_feat_matrix = (
            self.item_feats_norm
            .set_index("item_idx")[self.item_feat_cols]
            .astype(np.float32)
        )
        tensor = torch.tensor(
            item_feat_matrix.values, dtype=torch.float32
        ).to(self.device)

        with torch.no_grad():
            embs = self.model.get_item_embedding(tensor)

        self.item_embeddings = embs.cpu().numpy()
        self.item_indices    = item_feat_matrix.index.values
        print(f"Pre-computed {len(self.item_indices):,} item embeddings")

    def recommend(self, user_idx: int, k: int = 10) -> np.ndarray:
        if not self.is_fitted:
            raise RuntimeError("Call fit() first")

        self.model.eval()
        u_feats_dict = self.user_feat_map.get(
            user_idx, {c: 0.0 for c in self.user_feat_cols}
        )
        u_tensor = torch.tensor(
            list(u_feats_dict.values()), dtype=torch.float32
        ).unsqueeze(0).to(self.device)

        with torch.no_grad():
            u_emb = self.model.get_user_embedding(u_tensor).cpu().numpy()

        # Dot product with all item embeddings (cosine sim since normalised)
        scores = self.item_embeddings @ u_emb.T
        scores = scores.flatten()

        # Filter seen items
        seen = self.user_history.get(user_idx, set())
        for i, item_idx in enumerate(self.item_indices):
            if item_idx in seen:
                scores[i] = -np.inf

        top_k_local = np.argsort(scores)[::-1][:k]
        return self.item_indices[top_k_local]

    def recommend_batch(self, user_indices, k: int = 10) -> dict:
        return {u: self.recommend(u, k) for u in user_indices}

    def save(self, path):
        path = Path(path)
        torch.save(self.model.state_dict(), path.with_suffix(".pt"))
        tmp_model = self.model
        self.model = None
        with open(path, "wb") as f:
            pickle.dump(self, f)
        self.model = tmp_model
        print(f"Saved TwoTowerRecommender to {path}")

    @staticmethod
    def load(path, user_dim, item_dim):
        path = Path(path)
        with open(path, "rb") as f:
            obj = pickle.load(f)
        obj.model = TwoTowerModel(
            user_dim=user_dim, item_dim=item_dim,
            embedding_dim=obj.embedding_dim, hidden_dim=obj.hidden_dim,
            dropout=obj.dropout,
        )
        obj.model.load_state_dict(torch.load(path.with_suffix(".pt"), map_location="cpu"))
        obj.model.eval()
        return obj