File size: 16,103 Bytes
228af26
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
"""
🌪️ STORM ORACLE — Tornado Super-Predictor (training-ready, no placeholders)

- RadarPatternExtractor: multi-scale CNN + spatial attention pooling
- AtmosphericConditionEncoder: per-variable MLPs -> tokens -> attention -> fused vector
- Heads: probability (sigmoid), EF (logits), location (reg), timing (reg), uncertainty (sigmoid)
- Calibration: single temperature parameter (learnable/fittable after training)
- ContinuousLearner: online fine-tuning with replay buffer and EMA weights
"""

from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple

import torch
import torch.nn as nn
import torch.nn.functional as F


# ----------------------------- Types ---------------------------------

@dataclass
class TornadoPredictionBatch:
    """All outputs are BATCH TENSORS (no Python scalars)."""
    tornado_probability: torch.Tensor        # (B,)
    ef_scale_probs: torch.Tensor             # (B,6)
    most_likely_ef_scale: torch.Tensor       # (B,)
    location_offset: torch.Tensor            # (B,2)
    timing_predictions: torch.Tensor         # (B,3)
    uncertainty_scores: torch.Tensor         # (B,4) in [0,1]
    radar_signatures: torch.Tensor           # (B,3) [hook, meso, couplet]
    atmospheric_indicators: torch.Tensor     # (B,3) [cape, shear_norm, instability]
    logits: Optional[torch.Tensor] = None    # (B,) pre-sigmoid (for calibration/loss)


# ---------------------- Building blocks --------------------------------

class SpatialAttentionPool(nn.Module):
    """
    Turns a 2D feature map (B,C,H,W) into (B,C) using a learned query and MHA over H*W tokens.
    """
    def __init__(self, channels: int, num_heads: int = 8):
        super().__init__()
        self.channels = channels
        self.pos_embed = nn.Parameter(torch.randn(1, channels, 1))  # simple scalar per-channel bias over tokens
        self.query = nn.Parameter(torch.randn(1, 1, channels))      # learned global query token
        self.attn = nn.MultiheadAttention(embed_dim=channels, num_heads=num_heads, batch_first=True)
        self.ln = nn.LayerNorm(channels)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # x: (B,C,H,W) -> tokens: (B, H*W, C)
        B, C, H, W = x.shape
        tokens = x.view(B, C, H * W).transpose(1, 2)  # (B, HW, C)
        tokens = self.ln(tokens + self.pos_embed.expand(B, C, 1).transpose(1, 2))  # broadcast mild bias
        q = self.query.expand(B, -1, -1)  # (B,1,C)
        pooled, _ = self.attn(q, tokens, tokens)  # (B,1,C)
        return pooled.squeeze(1)  # (B,C)


class RadarPatternExtractor(nn.Module):
    """
    Advanced radar pattern extraction with spatial attention pooling.
    Accepts variable input_channels (e.g., 3×T for T time steps).
    """
    def __init__(self, input_channels: int = 3):
        super().__init__()
        self.conv1 = nn.Conv2d(input_channels, 64, kernel_size=7, padding=3)
        self.conv2 = nn.Conv2d(64, 128, kernel_size=5, padding=2)
        self.conv3 = nn.Conv2d(128, 256, kernel_size=3, padding=1)
        self.conv4 = nn.Conv2d(256, 512, kernel_size=3, padding=1)

        self.bn4 = nn.BatchNorm2d(512)

        # Specialized detectors
        self.hook_echo_detector = nn.Conv2d(512, 64, kernel_size=3, padding=1)
        self.mesocyclone_detector = nn.Conv2d(512, 64, kernel_size=5, padding=2)
        self.velocity_couplet_detector = nn.Conv2d(512, 64, kernel_size=3, padding=1)

        # Attention pooling to summarize (B,512,H',W') -> (B,512)
        self.pool = SpatialAttentionPool(512, num_heads=8)

        # Combine base + specialists -> 512 + 64*3 = 704 -> project to 1024
        self.proj = nn.Sequential(
            nn.Linear(512 + 64 * 3, 1024),
            nn.ReLU(),
            nn.Dropout(0.5),
        )

    def forward(self, radar_data: torch.Tensor) -> Dict[str, torch.Tensor]:
        # radar_data: (B,C,H,W)
        x = F.relu(self.conv1(radar_data)); x = F.max_pool2d(x, 2)
        x = F.relu(self.conv2(x));          x = F.max_pool2d(x, 2)
        x = F.relu(self.conv3(x));          x = F.max_pool2d(x, 2)
        x = F.relu(self.conv4(x));          x = self.bn4(x)

        hook = F.relu(self.hook_echo_detector(x))
        meso = F.relu(self.mesocyclone_detector(x))
        vel  = F.relu(self.velocity_couplet_detector(x))

        base_vec = self.pool(x)                             # (B,512)
        hook_vec = hook.mean(dim=(2, 3))                    # (B,64)
        meso_vec = meso.mean(dim=(2, 3))                    # (B,64)
        vel_vec  = vel.mean(dim=(2, 3))                     # (B,64)

        fused = torch.cat([base_vec, hook_vec, meso_vec, vel_vec], dim=1)  # (B,704)
        combined = self.proj(fused)  # (B,1024)

        strengths = torch.stack([
            hook_vec.mean(dim=1),      # (B,)
            meso_vec.mean(dim=1),      # (B,)
            vel_vec.mean(dim=1),       # (B,)
        ], dim=1)  # (B,3)

        return {
            "combined_features": combined,
            "signature_strengths": strengths,  # hook, meso, velocity couplet
        }


class AtmosphericConditionEncoder(nn.Module):
    """
    Encode environmental parameters using per-variable MLPs, then treat them as tokens and apply MHA.
    """
    def __init__(self):
        super().__init__()
        self.enc_cape        = nn.Linear(1, 32)
        self.enc_shear       = nn.Linear(4, 64)  # 0–1, 0–3, 0–6, deep
        self.enc_helicity    = nn.Linear(2, 32)  # 0–1, 0–3
        self.enc_temp        = nn.Linear(3, 32)  # sfc, 850, 500
        self.enc_dewpoint    = nn.Linear(2, 32)  # sfc, 850
        self.enc_pressure    = nn.Linear(1, 16)

        # we will embed each of the 6 groups to dim=64 and self-attend
        self.to_64 = nn.ModuleDict({
            "cape":     nn.Linear(32, 64),
            "shear":    nn.Identity(),          # already 64
            "helicity": nn.Linear(32, 64),
            "temp":     nn.Linear(32, 64),
            "dewpoint": nn.Linear(32, 64),
            "pressure": nn.Linear(16, 64),
        })
        self.ln = nn.LayerNorm(64)
        self.attn = nn.MultiheadAttention(embed_dim=64, num_heads=4, batch_first=True)

        self.fuse = nn.Sequential(
            nn.Linear(64 * 6, 256),
            nn.ReLU(),
            nn.Dropout(0.3),
        )

    def forward(self, atmo: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
        def ensure_2d(t: torch.Tensor, d: int) -> torch.Tensor:
            # make (B,d)
            t = t if t.ndim == 2 else t.view(-1, d)
            return t

        cape = ensure_2d(atmo.get("cape",        torch.zeros(1, 1, device=next(self.parameters()).device)), 1)
        shear= ensure_2d(atmo.get("wind_shear",  torch.zeros(1, 4, device=next(self.parameters()).device)), 4)
        hel  = ensure_2d(atmo.get("helicity",    torch.zeros(1, 2, device=next(self.parameters()).device)), 2)
        temp = ensure_2d(atmo.get("temperature", torch.zeros(1, 3, device=next(self.parameters()).device)), 3)
        dew  = ensure_2d(atmo.get("dewpoint",    torch.zeros(1, 2, device=next(self.parameters()).device)), 2)
        pres = ensure_2d(atmo.get("pressure",    torch.zeros(1, 1, device=next(self.parameters()).device)), 1)

        cape_e = F.relu(self.enc_cape(cape))           # (B,32)
        shear_e= F.relu(self.enc_shear(shear))         # (B,64)
        hel_e  = F.relu(self.enc_helicity(hel))        # (B,32)
        temp_e = F.relu(self.enc_temp(temp))           # (B,32)
        dew_e  = F.relu(self.enc_dewpoint(dew))        # (B,32)
        pres_e = F.relu(self.enc_pressure(pres))       # (B,16)

        tokens = torch.stack([
            self.ln(self.to_64["cape"](cape_e)),
            self.ln(self.to_64["shear"](shear_e)),
            self.ln(self.to_64["helicity"](hel_e)),
            self.ln(self.to_64["temp"](temp_e)),
            self.ln(self.to_64["dewpoint"](dew_e)),
            self.ln(self.to_64["pressure"](pres_e)),
        ], dim=1)  # (B, 6, 64)

        attn_out, _ = self.attn(tokens, tokens, tokens)  # (B,6,64)
        fused = self.fuse(attn_out.reshape(attn_out.size(0), -1))  # (B,256)

        # easy indicators for explanations/QA
        shear_mag = torch.linalg.vector_norm(shear, dim=-1)  # (B,)
        instab = cape.squeeze(-1) * shear_mag                # (B,)

        return {
            "atmospheric_features": fused,                # (B,256)
            "cape_score": cape.squeeze(-1),               # (B,)
            "shear_magnitude": shear_mag,                 # (B,)
            "instability_index": instab,                  # (B,)
        }


# -------------------------- Main model --------------------------------

class TornadoSuperPredictor(nn.Module):
    def __init__(self, in_channels: int = 3):
        super().__init__()
        self.radar_extractor = RadarPatternExtractor(input_channels=in_channels)
        self.atmo_encoder    = AtmosphericConditionEncoder()

        fused_dim = 1024 + 256

        self.prob_head = nn.Sequential(
            nn.Linear(fused_dim, 512), nn.ReLU(), nn.Dropout(0.4),
            nn.Linear(512, 256), nn.ReLU(),
            nn.Linear(256, 1)
        )
        self.ef_head = nn.Sequential(
            nn.Linear(fused_dim, 512), nn.ReLU(), nn.Dropout(0.4),
            nn.Linear(512, 6)
        )
        self.loc_head = nn.Sequential(
            nn.Linear(fused_dim, 512), nn.ReLU(), nn.Dropout(0.4),
            nn.Linear(512, 2)
        )
        self.time_head = nn.Sequential(
            nn.Linear(fused_dim, 512), nn.ReLU(), nn.Dropout(0.4),
            nn.Linear(512, 3)
        )
        self.unc_head = nn.Sequential(
            nn.Linear(fused_dim, 256), nn.ReLU(),
            nn.Linear(256, 4)
        )

        # temperature parameter for calibration (start at 1.0)
        self.register_parameter("log_temperature", nn.Parameter(torch.zeros(())))

        self._init_weights()

    def _init_weights(self):
        for m in self.modules():
            if isinstance(m, (nn.Linear, nn.Conv2d)):
                if isinstance(m, nn.Linear):
                    nn.init.xavier_uniform_(m.weight)
                else:
                    nn.init.kaiming_uniform_(m.weight, mode="fan_out", nonlinearity="relu")
                if m.bias is not None:
                    nn.init.zeros_(m.bias)

    @property
    def temperature(self) -> torch.Tensor:
        return torch.exp(self.log_temperature)  # positive

    def forward(self, radar_x: torch.Tensor, atmo: Dict[str, torch.Tensor]) -> TornadoPredictionBatch:
        # radar_x: (B,C,H,W), atmo: dict of (B,dim)
        r = self.radar_extractor(radar_x)
        a = self.atmo_encoder(atmo)

        fused = torch.cat([r["combined_features"], a["atmospheric_features"]], dim=1)  # (B,1280)

        logits = self.prob_head(fused).squeeze(-1)             # (B,)
        logits = logits / self.temperature.clamp_min(1e-6)     # calibrated logits
        probs  = torch.sigmoid(logits)                         # (B,)

        ef_logits = self.ef_head(fused)                        # (B,6)
        ef_probs  = F.softmax(ef_logits, dim=-1)
        ef_idx    = ef_probs.argmax(dim=-1)

        loc = self.loc_head(fused)                             # (B,2)
        tim = self.time_head(fused)                            # (B,3)
        unc = torch.sigmoid(self.unc_head(fused))              # (B,4) in [0,1]

        return TornadoPredictionBatch(
            tornado_probability=probs,
            ef_scale_probs=ef_probs,
            most_likely_ef_scale=ef_idx,
            location_offset=loc,
            timing_predictions=tim,
            uncertainty_scores=unc,
            radar_signatures=r["signature_strengths"],
            atmospheric_indicators=torch.stack([
                a["cape_score"], a["shear_magnitude"], a["instability_index"]
            ], dim=1),
            logits=logits,
        )


# --------------------- Continuous learning wrapper --------------------

class ContinuousLearner(nn.Module):
    """
    Light wrapper that adds:
      - optimizer + (optional) pos_weight or focal loss
      - EMA weights for stable inference during online updates
      - small replay buffer to avoid catastrophic forgetting
    """
    def __init__(
        self,
        model: TornadoSuperPredictor,
        lr: float = 1e-4,
        wd: float = 1e-4,
        use_focal: bool = False,
        pos_weight: Optional[float] = None,
        ema_decay: float = 0.999,
        replay_capacity: int = 2048,
        device: Optional[torch.device] = None,
    ):
        super().__init__()
        self.model = model
        self.device = device or next(model.parameters()).device
        self.opt = torch.optim.AdamW(self.model.parameters(), lr=lr, weight_decay=wd)
        self.use_focal = use_focal
        self.pos_weight = None if pos_weight is None else torch.tensor(pos_weight, device=self.device)
        self.ema_decay = ema_decay

        # EMA weights
        self.shadow = {k: v.detach().clone() for k, v in self.model.state_dict().items()}
        self.replay_capacity = replay_capacity
        self._replay = []  # list of tuples (radar_x, atmo_dict, y)

    def _bce_loss(self, logits: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
        if self.pos_weight is not None:
            return F.binary_cross_entropy_with_logits(logits, y.float(), pos_weight=self.pos_weight)
        return F.binary_cross_entropy_with_logits(logits, y.float())

    def _focal_loss(self, logits: torch.Tensor, y: torch.Tensor, gamma: float = 2.0, alpha: float = 0.5) -> torch.Tensor:
        p = torch.sigmoid(logits)
        pt = p * y + (1 - p) * (1 - y)
        w = (1 - pt).pow(gamma)
        at = alpha * y + (1 - alpha) * (1 - y)
        loss = -(y * torch.log(p.clamp_min(1e-9)) + (1 - y) * torch.log((1 - p).clamp_min(1e-9))) * w * at
        return loss.mean()

    @torch.no_grad()
    def _update_ema(self):
        for k, v in self.model.state_dict().items():
            self.shadow[k].mul_(self.ema_decay).add_(v, alpha=(1.0 - self.ema_decay))

    def train_step(self, radar_x: torch.Tensor, atmo: Dict[str, torch.Tensor], y: torch.Tensor) -> Dict[str, float]:
        self.model.train()
        out = self.model(radar_x, atmo)   # contains logits & probs

        if self.use_focal:
            loss = self._focal_loss(out.logits, y)
        else:
            loss = self._bce_loss(out.logits, y)

        self.opt.zero_grad(set_to_none=True)
        loss.backward()
        nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
        self.opt.step()
        self._update_ema()

        # push to replay
        if self.replay_capacity > 0:
            with torch.no_grad():
                if len(self._replay) >= self.replay_capacity:
                    self._replay.pop(0)
                # store small detached copy (avoid GPU memory blowup)
                self._replay.append((
                    radar_x.detach().cpu(),
                    {k: v.detach().cpu() for k, v in atmo.items()},
                    y.detach().cpu()
                ))

        with torch.no_grad():
            prob = out.tornado_probability.mean().item()
        return {"loss": float(loss.item()), "avg_prob": prob}

    @torch.no_grad()
    def ema_state_dict(self) -> Dict[str, torch.Tensor]:
        return {k: v.clone() for k, v in self.shadow.items()}

    @torch.no_grad()
    def load_ema_weights(self):
        self.model.load_state_dict(self.ema_state_dict())

    def replay_step(self, batch_size: int = 16) -> Optional[Dict[str, float]]:
        if not self._replay:
            return None
        import random
        idxs = random.sample(range(len(self._replay)), k=min(batch_size, len(self._replay)))
        xs = torch.cat([self._replay[i][0] for i in idxs], dim=0).to(self.device)
        ys = torch.cat([self._replay[i][2] for i in idxs], dim=0).to(self.device)
        atmo = {}
        # stack dict fields
        keys = list(self._replay[idxs[0]][1].keys())
        for k in keys:
            atmo[k] = torch.cat([self._replay[i][1][k] for i in idxs], dim=0).to(self.device)
        return self.train_step(xs, atmo, ys)