model.py

import math
from dataclasses import dataclass
from typing import Dict

import helpers
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from distance_gat_fc import DistanceGatFC
from torch.nn import functional as F
from torch.profiler import ProfilerActivity, profile, record_function
from tqdm import tqdm


class TimeEncode(torch.nn.Module):
    def __init__(self, expand_dim):
        super(TimeEncode, self).__init__()
        time_dim = expand_dim
        self.basis_freq = torch.nn.Parameter(
            (torch.from_numpy(1 / 10 ** np.linspace(0, 9, time_dim))).float()
        )
        self.phase = torch.nn.Parameter(torch.zeros(time_dim).float())
     
    def forward(self, ts):
        # ts: [N, L]
        batch_size = ts.size(0)
        seq_len = ts.size(1)

        ts = ts.view(batch_size, seq_len, 1)  # [N, L, 1]
        map_ts = ts * self.basis_freq.view(1, 1, -1)  # [N, L, time_dim]
        map_ts += self.phase.view(1, 1, -1)

        harmonic = torch.cos(map_ts)
        return harmonic  # self.dense(harmonic)


class LayerNorm(nn.Module):
    def __init__(self, ndim, bias):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(ndim))
        self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None

    def forward(self, input):
        return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)


class CausalSelfAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        assert config["n_embd"] % config["n_head"] == 0
        # key, query, value projections for all heads, but in a batch
        self.c_attn = nn.Linear(
            config["n_embd"], 3 * config["n_embd"], bias=config["bias"]
        )
        # output projection
        self.c_proj = nn.Linear(config["n_embd"], config["n_embd"], bias=config["bias"])
        self.attn_dropout = nn.Dropout(config["dropout"])
        self.resid_dropout = nn.Dropout(config["dropout"])
        self.n_head = config["n_head"]
        self.n_embd = config["n_embd"]
        self.dropout = config["dropout"]
        # support only in PyTorch >= 2.0
        self.flash = hasattr(torch.nn.functional, "scaled_dot_product_attention")
        self.flash = False
        if not self.flash:
            print(
                "WARNING: using slow attention. Flash Attention requires PyTorch >= 2.0"
            )
            self.register_buffer(
                "bias",
                torch.tril(torch.ones(config["block_size"], config["block_size"])).view(
                    1, 1, config["block_size"], config["block_size"]
                ),
            )

        self.register_buffer(
            "k_cache",
            torch.empty(
                config["batch_size"],
                config["block_size"],
                config["n_embd"],
            ),
            persistent=False,
        )

        self.register_buffer(
            "v_cache",
            torch.empty(
                config["batch_size"],
                config["block_size"],
                config["n_embd"],
            ),
            persistent=False,
        )

    @torch.no_grad()
    def infer_next(self, start_pos: int, x: torch.Tensor):
        B, T, C = x.size()
        q, k, v = self.c_attn(x).split(self.n_embd, dim=2)

        kc = self.get_buffer("k_cache")
        vc = self.get_buffer("v_cache")

        kc[:B, start_pos : start_pos + T, :] = k
        vc[:B, start_pos : start_pos + T, :] = v

        q: torch.Tensor = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)
        k: torch.Tensor = (
            kc[:B, : start_pos + T]
            .view(B, -1, self.n_head, C // self.n_head)
            .transpose(1, 2)
        )
        v: torch.Tensor = (
            vc[:B, : start_pos + T]
            .view(B, -1, self.n_head, C // self.n_head)
            .transpose(1, 2)
        )

        assert not self.flash
        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
        assert T == 1
        if T != 1:
            att = att.masked_fill(self.bias[:, :, start_pos:start_pos+T, :start_pos+T] == 0, float("-inf"))
        att = F.softmax(att, dim=-1)
        att = self.attn_dropout(att)
        y = att @ v  # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)

        y = y.transpose(1, 2).contiguous().view(B, T, C)
        y = self.resid_dropout(self.c_proj(y))
        return y

    def forward(self, x):
        B, T, C = x.size()
        q, k, v = self.c_attn(x).split(self.n_embd, dim=2)
        k = k.view(B, T, self.n_head, C // self.n_head).transpose(
            1, 2
        )  # (B, nh, T, hs)
        q = q.view(B, T, self.n_head, C // self.n_head).transpose(
            1, 2
        )  # (B, nh, T, hs)
        v = v.view(B, T, self.n_head, C // self.n_head).transpose(
            1, 2
        )  # (B, nh, T, hs)

        # causal self-attention (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
        if self.flash:
            y = torch.nn.functional.scaled_dot_product_attention(
                q, k, v, attn_mask=None, dropout_p=self.dropout, is_causal=True
            )
        else:
            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
            att = att.masked_fill(self.bias[:, :, :T, :T] == 0, float("-inf"))
            att = F.softmax(att, dim=-1)
            att = self.attn_dropout(att)
            y = att @ v  # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
        y = y.transpose(1, 2).contiguous().view(B, T, C)

        # output projection
        y = self.resid_dropout(self.c_proj(y))
        return y


class MLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.c_fc = nn.Linear(
            config["n_embd"], 4 * config["n_embd"], bias=config["bias"]
        )
        self.gelu = nn.GELU()
        self.c_proj = nn.Linear(
            4 * config["n_embd"], config["n_embd"], bias=config["bias"]
        )
        self.dropout = nn.Dropout(config["dropout"])

    def forward(self, x):
        x = self.c_fc(x)
        x = self.gelu(x)
        x = self.c_proj(x)
        x = self.dropout(x)
        return x


class Block(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.ln_1 = LayerNorm(config["n_embd"], bias=config["bias"])
        self.attn = CausalSelfAttention(config)
        self.ln_2 = LayerNorm(config["n_embd"], bias=config["bias"])
        self.mlp = MLP(config)

    @torch.no_grad()
    def infer_next(self, start_pos: int, x: torch.Tensor):
        x = x + self.attn.infer_next(start_pos, self.ln_1(x))
        x = x + self.mlp(self.ln_2(x))
        return x

    def forward(self, x):
        x = x + self.attn(self.ln_1(x))
        x = x + self.mlp(self.ln_2(x))
        return x


@dataclass
class MyTransformerConfig:
    gnn_config: Dict
    data_feature: Dict
    tf_config: Dict
    seed: int = 0
    data: str = ""
    datapath: str = ""
    vocab_size: int = 0
    epochs: int = 50
    batch_size: int = 32
    device: torch.device = torch.device("cuda:0")

    def to_dict(self):
        return {k: v for k, v in self.__dict__.items()}


class MyTransformer(nn.Module):
    def __init__(self, config):
        super().__init__()
        assert config.vocab_size is not None
        assert config.tf_config["block_size"] is not None
        self.config = config
        self.device = config.device
        self.loss = nn.MSELoss(reduction="none")

        config.tf_config["batch_size"] = config.batch_size

        self.transformer = nn.ModuleDict(
            dict(
                wte=nn.Embedding(
                    config.vocab_size,
                    config.tf_config["n_embd"] - config.tf_config["t_embd"],
                ),
                tme=TimeEncode(config.tf_config["t_embd"]),
                wpe=nn.Embedding(
                    config.tf_config["block_size"], config.tf_config["n_embd"]
                ),
                drop=nn.Dropout(config.tf_config["dropout"]),
                h=nn.ModuleList(
                    [
                        Block(config.tf_config)
                        for _ in range(config.tf_config["n_layer"])
                    ]
                ),
                ln_f=LayerNorm(
                    config.tf_config["n_embd"], bias=config.tf_config["bias"]
                ),
            )
        )
        self.lm_head = nn.Linear(
            config.tf_config["n_embd"], config.vocab_size, bias=False
        )
        self.time_pred = nn.Linear(config.tf_config["n_embd"], 1, bias=False)
        self.dist_layer = nn.BatchNorm1d(config.vocab_size)
        self.time_layer = nn.BatchNorm1d(config.vocab_size)
        # self.transformer.wte.weight = self.lm_head.weight
        
        self.gat = DistanceGatFC(
            config=config.gnn_config,
            data_feature=config.data_feature,
            device=config.device,
        )
        self.apply(self._init_weights)
        for pn, p in self.named_parameters():
            if pn.endswith("c_proj.weight"):
                torch.nn.init.normal_(
                    p, mean=0.0, std=0.02 / math.sqrt(2 * config.tf_config["n_layer"])
                )
        print("number of parameters: %.2fM" % (self.get_num_params() / 1e6,))

    def get_num_params(self, non_embedding=True):
        n_params = sum(p.numel() for p in self.parameters())
        if non_embedding:
            n_params -= self.transformer.wpe.weight.numel()
        return n_params

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
            if module.bias is not None:
                torch.nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)

    def get_acc_topk(self, preds, targets):
        acc_K = [1, 5, 10, 20]
        result = {}
        totalMRR = []
        for K in acc_K:
            result[K] = 0

        seq_len_l = []
        for i in range(len(preds)):
            max_len = self.config.tf_config["block_size"] - 1
            seq_len = max_len - len(torch.where(targets[i] == -1)[0])
            seq_len_l.append(seq_len)

            for j in range(seq_len):
                pred, target = preds[i][j], targets[i][j].item()
                sortedPred = torch.topk(pred, len(pred))[1].tolist()
                truthIndex = sortedPred.index(target) + 1
                avgPrec = 1 / truthIndex
                totalMRR.append(avgPrec)

                sorted_indexs = {}
                for K in acc_K:
                    sorted_indexs[K] = sortedPred[:K]
                    if target in sorted_indexs[K]:
                        result[K] += 1

        result["num_of_test"] = sum(seq_len_l)
        result["mrr"] = np.sum(totalMRR)
        result["mrr_num"] = len(totalMRR)
        return result

    @torch.no_grad()
    def infer_next(
        self,
        start_pos: int,
        idx,
        tim_real,
        adj_batch,
        dist_geo_batch_x,
        dist_geo_batch_dest,
        gnn_emb,
    ):
        assert gnn_emb is not None
        tok_emb = gnn_emb[idx]  # [batch, len-1, n_embd]
        tim_emb = self.transformer.tme(tim_real)  # [batch, len-1, t_embd]

        device = idx.device
        T = idx.size(1)
        assert (
            start_pos + T <= self.config.tf_config["block_size"]
        ), f"Cannot forward sequence of length {start_pos + T}, block size is only {self.config.tf_config['block_size']}"
        pos = torch.arange(
            start_pos, start_pos + T, dtype=torch.long, device=device
        ).unsqueeze(0)
        pos_emb = self.transformer.wpe(pos)  # [1, len-1, n_embd + t_embd]

        x = self.transformer.drop(torch.concat((tok_emb, tim_emb), dim=-1) + pos_emb)
        for block in self.transformer.h:
            x = block.infer_next(start_pos, x)
        x = self.transformer.ln_f(x)

        N, L, C = dist_geo_batch_x.shape
        weight_dis = (
            self.dist_layer((dist_geo_batch_x + dist_geo_batch_dest).view(-1, C))
            .view(N, L, C)
            .sigmoid()
        )  # two type distance activation(batch, len-1, n)

        weight_tim = (
            self.time_layer(dist_geo_batch_x.reshape(-1, C)).reshape(N, L, C).sigmoid()
        )  # last distance activation (batch, len-1, n)

        logits_weighted = (
            self.lm_head(x[:, [-1], :]) * weight_dis[:, [-1], :]
        )  # (batch, 1, n) 
        logits_masked = torch.where(
            adj_batch[:, [-1], :] == 0,
            torch.tensor(float("-inf")).to(self.device),
            logits_weighted,
        )  # (batch, 1, n)

        # duration prediction
        dur_hat = self.time_pred(x[:, [-1], :]).squeeze(-1)  # (batch, 1)
        probs = logits_masked.softmax(dim=-1)
        delta_dis = (probs * weight_tim[:, [-1], :]).sum(dim=-1)  # (batch, 1)
        dur_pred = dur_hat * delta_dis  # (batch, 1)
        dur_pred = torch.clamp(dur_pred, min=0.0)
        loss = None
        return logits_masked, dur_pred, loss

    def forward(
        self,
        idx,
        targets=None,
        tim_real=None,
        adj_batch=None,
        dist_geo_batch_x=None,
        dist_geo_batch_dest=None,
        gnn_emb=None,
    ):
        """
        idx (input): (batch, len-1)
        targets (output): (batch, len-1)
        dest (destination): (batch, len-1)
        tim_real:
        adj_batch: (batch, len-1, n)
        dist_geo_batch: (batch, len-1, n)
        """

        if gnn_emb == None:
            assert targets != None
            gnn_emb = self.gat.compute_gnn()  # [loc_num, n_embd] 
        tok_emb = gnn_emb[idx]  # [batch, len-1, n_embd]

        assert targets is not None
        if targets is not None:
            tim_emb = self.transformer.tme(tim_real[:, :-1])  # [batch, len-1, t_embd]
        else:
            tim_emb = self.transformer.tme(tim_real)  # [batch, len-1, t_embd]

        device = idx.device
        t = idx.size()[1]
        assert (
            t <= self.config.tf_config["block_size"]
        ), f"Cannot forward sequence of length {t}, block size is only {self.config.tf_config['block_size']}"
        pos = torch.arange(0, t, dtype=torch.long, device=device).unsqueeze(0)
        pos_emb = self.transformer.wpe(pos)  # [1, len-1, n_embd + t_embd]

        x = self.transformer.drop(torch.concat((tok_emb, tim_emb), dim=-1) + pos_emb)
        for block in self.transformer.h:
            x = block(x)
        x = self.transformer.ln_f(x)

        import torch.utils.checkpoint as ckpt_utils

        def geo_func(x):
            N, L, C = dist_geo_batch_x.shape
            len_mask = tim_real[:, 1:] == -1  # (batch, len-1)

            weight_dis = (
                self.dist_layer(
                    (dist_geo_batch_x + dist_geo_batch_dest).view(-1, C)
                )
                .view(N, L, C)
                .sigmoid()
            )  # two type distance activation(batch, len-1, n)

            weight_tim = (
                self.time_layer(dist_geo_batch_x.view(-1, C))
                .view(N, L, C)
                .sigmoid()
            )  # last distance activation (batch, len-1, n)

            # pretraining, target is the correct next step rid
            assert targets is not None
            if targets is not None:
                logits_weighted = (
                    self.lm_head(x) * weight_dis
                )  # distance activation (batch, len-1, n)
                logits_masked = torch.where(
                    adj_batch == 0,
                    torch.tensor(float("-inf")).to(self.device),
                    logits_weighted,
                )  # adj mask (batch, len-1, n)
                loss = F.cross_entropy(
                    logits_masked.view(-1, logits_masked.size(-1)),
                    targets.reshape(-1),
                    ignore_index=-1,
                )
                dur_pred = None

                # duration prediction
                dur_hat = self.time_pred(x).squeeze(-1)  # (batch, len-1)
                probs = logits_masked.softmax(
                    dim=-1
                ) 
                prob_mask = torch.where(
                    adj_batch == 0, torch.tensor(float(0.0)).to(self.device), probs
                )
                delta_dis = (prob_mask * weight_tim).sum(
                    dim=-1
                )  # expectation duration for all locations (batch, len-1)
                dur_pred = dur_hat * delta_dis  # to activate (batch, len-1)
                
                dur_real = tim_real[:, 1:] - tim_real[:, :-1]  # (batch, len-1)
                dur_real[len_mask] = 0
                dur_pred[len_mask] = 0
                loss_dur = self.loss(dur_pred, dur_real)  # (batch, len-1)
                loss_dur_masked = (
                    loss_dur * (~len_mask).float()
                ).sum()  # gives \sigma_euclidean over unmasked elements
                loss += loss_dur_masked / (~len_mask).sum()

            # generate seq based on last idx
            else:
                logits_weighted = (
                    self.lm_head(x[:, [-1], :]) * weight_dis[:, [-1], :]
                )  # (batch, 1, n)
                logits_masked = torch.where(
                    adj_batch[:, [-1], :] == 0,
                    torch.tensor(float("-inf")).to(self.device),
                    logits_weighted,
                )  # (batch, 1, n)
                # duration prediction
                dur_hat = self.time_pred(x[:, [-1], :]).squeeze(-1)  # (batch, 1)
                probs = logits_masked.softmax(dim=-1)
                prob_mask = torch.where(
                    adj_batch[:, [-1], :] == 0,
                    torch.tensor(float(0.0)).to(self.device),
                    probs,
                )
                delta_dis = (probs * weight_tim[:, [-1], :]).sum(
                    dim=-1
                )  # (batch, 1)
                dur_pred = dur_hat * delta_dis  # (batch, 1)
                loss = None
            return logits_masked, dur_pred, loss

        return ckpt_utils.checkpoint(geo_func, x)
    

    @torch.no_grad()
    def generate(self, args, dist_geo, adj, num_samples, temperature=1.0, top_k=None):
        """
        Take a conditioning sequence of indices idx (LongTensor of shape (b,t)) and complete
        the sequence max_new_tokens times, feeding the predictions back into the model each time.
        Most likely you'll want to make sure to be in model.eval() mode of operation for this.
        """

        od_and_probs_float = helpers.read_od_pair_distribution(
            args.data
        )  # tensor(n, 3)
        od_ids = torch.multinomial(
            od_and_probs_float[:, -1], num_samples, replacement=True
        )
        origins = (
            od_and_probs_float[od_ids, 0].int().reshape(-1, 1).to(self.device).long()
        )
        destinations = (
            od_and_probs_float[od_ids, 1].int().reshape(-1, 1).to(self.device).long()
        )

        start_t_probs = helpers.read_start_t_probs(args.data)  
        assert len(start_t_probs) == 2880
        start_ts = (
            torch.multinomial(start_t_probs, num_samples, replacement=True)
            .reshape(-1, 1)
            .to(self.device)
        )

        gen_seq_len = self.config.tf_config["block_size"] - 1
        gen_seq_batch = 32
        gnn_emb = self.gat.compute_gnn()
        pred_traj, pred_tim, pred = [], [], []
        batch_num = int(np.ceil(num_samples / gen_seq_batch))

        adj_tensor = torch.from_numpy(adj)
        dist_geo_tensor = torch.from_numpy(dist_geo)
        for i in tqdm(range(batch_num)):
            sid_bch = i * gen_seq_batch
            eid_bch = min(sid_bch + gen_seq_batch, num_samples)
            des = destinations[sid_bch:eid_bch]

            bs = eid_bch - sid_bch
            idx_cond = torch.zeros(bs, gen_seq_len+1).type_as(origins)
            tim = torch.zeros(bs, gen_seq_len+1, dtype=torch.float32).to(self.device)

            idx_cond[:,0:1] = origins[sid_bch:eid_bch]
            tim[:,0:1] = start_ts[sid_bch:eid_bch]

            adj_batch = torch.zeros(bs, gen_seq_len+1, *adj_tensor.shape[1:],
                                    dtype=adj_tensor.dtype, device=self.device)
            dist_geo_batch_x = torch.zeros(bs, gen_seq_len+1, *dist_geo_tensor.shape[1:],
                                           dtype=dist_geo_tensor.dtype, device=self.device)
            dist_geo_batch_dest = torch.zeros(bs, 1, *dist_geo_tensor.shape[1:],
                                              dtype=dist_geo_tensor.dtype, device=self.device)
            adj_batch[:,0,:].copy_(adj_tensor[idx_cond[:,0].cpu()])
            dist_geo_batch_x[:,0,:].copy_(dist_geo_tensor[idx_cond[:,0].cpu()])
            dist_geo_batch_dest[:,0,:].copy_(dist_geo_tensor[des[:,0].cpu()])

            for t in range(gen_seq_len): 
                # with profile(
                #     activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
                #     record_shapes=True,
                # ) as prof:
                    # with record_function("generate self.forward"):
                logits, dur_pred, _ = self.infer_next(
                    start_pos=t,
                    idx=idx_cond[:,t:t+1],
                    tim_real=torch.round(tim[:, t:t+1]).long(),
                    adj_batch=adj_batch[:, :t+1, :],
                    dist_geo_batch_x=dist_geo_batch_x[:, :t+1, :],
                    dist_geo_batch_dest=dist_geo_batch_dest,
                    gnn_emb=gnn_emb,
                )  # (batch, 1, n)

                logits = logits[:, -1, :] / temperature
                if top_k is not None:
                    v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
                    logits[logits < v[:, [-1]]] = -float("Inf")
                probs = F.softmax(logits, dim=-1)
                fill_matrix = torch.full_like(
                    probs, 1 / self.config.tf_config["block_size"]
                )
                probs = torch.where(torch.isnan(probs), fill_matrix, probs)
                idx_next = torch.multinomial(probs, num_samples=1)

                idx_next = idx_next.view(-1)
                idx_cond[:,t+1].copy_(idx_next)

                tmp = tim[:, t].clone()
                mask = (tmp >= 1440)
                tmp[mask] -= 1440
                tmp = (tmp + dur_pred.view(-1)) % 1440
                tmp[mask] += 1440
                tim[:,t+1].copy_(tmp)

                adj_batch[:,t+1].copy_(adj_tensor[idx_next.cpu()])
                dist_geo_batch_x[:,t+1].copy_(dist_geo_tensor[idx_next.cpu()])

            pred_traj.extend(idx_cond.tolist())
            pred_tim.extend(torch.round(tim).long().tolist())

        df = pd.read_csv(f"./data/{args.data}/roadmap.rel")
        oid = df["origin_id"].tolist()
        did = df["destination_id"].tolist()
        # 这个stop_points的意思应该是那种进去了就出不来的道路
        stop_points = list(set(did) - set(oid))

        destinations = destinations.reshape(-1).cpu().tolist()
        map_manager = helpers.MapManager(args.data)
        to_des_cnt = 0
        for i in range(len(pred_traj)):
            include_stop = len(set(pred_traj[i]).intersection(stop_points)) != 0
            if destinations[i] in pred_traj[i] and (
                not include_stop
            ):  
                to_des_cnt += 1
                dest_pos = pred_traj[i].index(destinations[i])
                pred.append(
                    [
                        pred_traj[i][: dest_pos + 1],
                        pred_tim[i][: dest_pos + 1],
                        destinations[i],
                    ]
                )
            # 不会执行
            elif destinations[i] in pred_traj[i] and include_stop: 
                stp = list(set(pred_traj[i]).intersection(stop_points))[0]
                stop_pos = pred_traj[i].index(stp)
                dest_pos = pred_traj[i].index(destinations[i])
                if (
                    dest_pos <= stop_pos
                ):  
                    to_des_cnt += 1
                    pred.append(
                        [
                            pred_traj[i][: dest_pos + 1],
                            pred_tim[i][: dest_pos + 1],
                            destinations[i],
                        ]
                    )
                elif stop_pos + 1 >= map_manager.min_len: 
                    pred.append(
                        [
                            pred_traj[i][: stop_pos + 1],
                            pred_tim[i][: stop_pos + 1],
                            destinations[i],
                        ]
                    )
            elif (
                not include_stop
            ):  # Known: des not in traj                              
                length = self.config.tf_config["block_size"]
                pred.append(
                    [pred_traj[i][:length], pred_tim[i][:length], destinations[i]]
                )
            # 不会执行
            else:
                stp = list(set(pred_traj[i]).intersection(stop_points))[0]  
                stop_pos = pred_traj[i].index(stp)
                if stop_pos + 1 >= map_manager.min_len:
                    pred.append(
                        [
                            pred_traj[i][: stop_pos + 1],
                            pred_tim[i][: stop_pos + 1],
                            destinations[i],
                        ]
                    )
        return pred