File size: 8,905 Bytes
fbaf630
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
aba9dd8
fbaf630
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
# src/syntax_pred/model.py — removed sigma & yulie, robust auto-loading, backward-compatible **kwargs
from __future__ import annotations
from typing import Any, Dict, List, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.models.video as tvmv
import lightning.pytorch as pl

class SyntaxLightningModule(pl.LightningModule):
    """
    R3D-18 backbone с настраиваемой "головой" для последовательностей.
    На выходе тензор (B, 2): [логит классификации, регрессия = log(1 + score)].
    """
    def __init__(
        self,
        num_classes: int,
        lr: float,
        variant: str,
        weight_decay: float = 0.0,
        max_epochs: Optional[int] = None,
        weight_path: Optional[str] = None,     # загрузка только бэкбона (опционально)
        pl_weight_path: Optional[str] = None,  # загрузка целого модуля (ckpt/pt)
        rnn_hidden_div: int = 4,
        rnn_dropout: float = 0.2,
        bert_nhead: int = 4,
        bert_layers: int = 1,
        bert_ff_div: int = 4,
        bert_dropout: float = 0.2,
        precision: str = "bf16",
        **_ignore,  # игнорирование лишних/неиспользуемых аргументов в вызове
    ) -> None:
        super().__init__()
        self.save_hyperparameters()
        self.num_classes = num_classes
        self.variant = variant
        self.lr = lr
        self.weight_decay = weight_decay
        self.max_epochs = max_epochs

        # --- Backbone (R3D-18)
        self.model = tvmv.r3d_18(weights=tvmv.R3D_18_Weights.DEFAULT)
        in_features = self.model.fc.in_features
        self.model.fc = nn.Linear(in_features, 2, bias=True)

        if weight_path is not None:
            self._load_backbone_weight(weight_path, self.model)

        # Для последовательных голов используем признаки до FC
        if self.variant != "mean_out":
            self.model.fc = nn.Identity()

        # --- Heads
        if self.variant == "mean_out":
            pass
        elif self.variant in ("gru_mean", "gru_last"):
            self.rnn = nn.GRU(in_features, in_features // rnn_hidden_div, batch_first=True)
            self.dropout = nn.Dropout(rnn_dropout)
            self.fc = nn.Linear(in_features // rnn_hidden_div, num_classes)
        elif self.variant in ("lstm_mean", "lstm_last"):
            self.lstm = nn.LSTM(
                input_size=in_features,
                hidden_size=in_features // rnn_hidden_div,
                proj_size=num_classes,
                batch_first=True,
            )
        elif self.variant == "mean":
            self.fc = nn.Linear(in_features, num_classes)
        elif self.variant in ("bert_mean", "bert_cls", "bert_cls2"):
            enc_layer = nn.TransformerEncoderLayer(
                d_model=in_features,
                nhead=bert_nhead,
                batch_first=True,
                dim_feedforward=in_features // bert_ff_div,
                dropout=bert_dropout,
            )
            self.encoder = nn.TransformerEncoder(enc_layer, num_layers=bert_layers)
            self.dropout = nn.Dropout(bert_dropout)
            self.fc = nn.Linear(in_features, num_classes)
            if self.variant == "bert_cls2":
                self.cls = nn.Parameter(torch.randn(1, 1, in_features))
        else:
            raise ValueError(f"Unknown variant: {self.variant}")

        # Загрузка полного state_dict (поддерживает форматы ckpt/pt)
        if pl_weight_path is not None:
            self._load_full_module(pl_weight_path)

        # Лоссы
        self.loss_clf = nn.BCEWithLogitsLoss(reduction="none")
        self.loss_reg = nn.MSELoss(reduction="none")

        # Кэши для валидационных метрик
        self.y_val: List[int] = []
        self.p_val: List[float] = []
        self.r_val: List[int] = []
        self.ty_val: List[float] = []
        self.tp_val: List[float] = []

    # ---------------- weights -----------------
    @staticmethod
    def _strip_prefix(key: str) -> str:
        for pref in ("model.", "backbone.", "module.", "net."):
            if key.startswith(pref):
                return key[len(pref):]
        return key

    def _load_backbone_weight(self, weight_path: str, model: nn.Module) -> None:
        ckpt = torch.load(weight_path, map_location="cpu")
        if isinstance(ckpt, dict) and "state_dict" in ckpt:
            ckpt = ckpt["state_dict"]
        new_sd = {
            self._strip_prefix(k).replace("fc.", ""): v
            for k, v in ckpt.items()
            if k.startswith(("model.", "backbone.", "module."))
        }
        model.load_state_dict(new_sd, strict=False)

    def _load_block(self, sd: Dict[str, torch.Tensor], module: nn.Module, prefix: str) -> None:
        filtered = {k.replace(f"{prefix}.", ""): v for k, v in sd.items() if k.startswith(prefix)}
        missing, unexpected = module.load_state_dict(filtered, strict=False)
        if missing:
            print(f"[load] Missing for {prefix}: {missing}")
        if unexpected:
            print(f"[load] Unexpected for {prefix}: {unexpected}")

    def _load_full_module(self, path: str) -> None:
        raw = torch.load(path, map_location="cpu")
        sd: Dict[str, torch.Tensor] = raw["state_dict"] if (isinstance(raw, dict) and "state_dict" in raw) else raw
        self._load_block(sd, self.model, "model")
        if self.variant == "mean_out":
            pass
        elif self.variant in ("gru_mean", "gru_last"):
            self._load_block(sd, self.rnn, "rnn")
            self._load_block(sd, self.fc, "fc")
        elif self.variant in ("lstm_mean", "lstm_last"):
            self._load_block(sd, self.lstm, "lstm")
        elif self.variant == "mean":
            self._load_block(sd, self.fc, "fc")
        elif self.variant in ("bert_mean", "bert_cls", "bert_cls2"):
            self._load_block(sd, self.encoder, "encoder")
            self._load_block(sd, self.fc, "fc")
            if self.variant == "bert_cls2" and "cls" in sd:
                with torch.no_grad():
                    self.cls.copy_(sd["cls"])

    # ---------------- forward -----------------
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Подача батча последовательностей формы (B,S,C,T,H,W),
        где S — число клипов внутри исследования для одной артерии.
        """
        b, s = x.shape[:2]
        x = x.flatten(0, 1)   # (B*S, C, T, H, W)
        x = self.model(x)     # (B*S, 2) или (B*S, F) при fc=Identity
        x = x.unflatten(0, (b, s))  # (B, S, ...)

        if self.variant == "mean_out":
            x = x.mean(dim=1)
        elif self.variant in ("gru_mean", "gru_last"):
            all_outs, last = self.rnn(x)
            x = all_outs.mean(dim=1) if self.variant == "gru_mean" else last[0]
            x = self.dropout(x)
            x = self.fc(x)
        elif self.variant in ("lstm_mean", "lstm_last"):
            all_outs, (last_out, _) = self.lstm(x)
            x = all_outs.mean(dim=1) if self.variant == "lstm_mean" else last_out
        elif self.variant == "mean":
            x = x.mean(dim=1)
            x = self.fc(x)
        elif self.variant in ("bert_mean", "bert_cls", "bert_cls2"):
            if self.variant == "bert_cls":
                x = F.pad(x, (0, 0, 1, 0), value=0.0)  # добавление CLS токена
            elif self.variant == "bert_cls2":
                bs = x.size(0)
                x = torch.cat([self.cls.expand(bs, -1, -1), x], dim=1)
            x = self.encoder(x)
            x = x.mean(dim=1) if self.variant == "bert_mean" else x[:, 0, :]
            x = self.dropout(x)
            x = self.fc(x)
        else:
            raise ValueError(self.variant)
        return x  # (B, 2)

    # --------------- training (kept for completeness) ---------------
    def training_step(self, batch, batch_idx):
        x, y, target, _ = batch
        y_hat = self(x)
        yp_clf = y_hat[:, 0:1]
        yp_reg = y_hat[:, 1:]

        weights_clf = torch.where(y > 0, 1.0, 0.2)
        clf_loss = (self.loss_clf(yp_clf, y) * weights_clf).mean()

        reg_loss = self.loss_reg(yp_reg, target).mean()
        return clf_loss + 0.5 * reg_loss

    def configure_optimizers(self):
        params = [p for p in self.parameters() if p.requires_grad]
        opt = torch.optim.Adam(params, lr=self.lr, weight_decay=self.weight_decay)
        if self.max_epochs:
            sch = torch.optim.lr_scheduler.OneCycleLR(opt, max_lr=self.lr, total_steps=self.max_epochs)
            return [opt], [sch]
        return opt