File size: 9,533 Bytes
b73936d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import math
import logging
from itertools import chain

import torch
import torch.nn as nn
from torch.utils.checkpoint import checkpoint

from timm.models.layers import DropPath, trunc_normal_
import torch.fft

from .transformer_ls import AttentionLS

_logger = logging.getLogger(__name__)


class Mlp(nn.Module):
    def __init__(
        self,
        in_features,
        hidden_features=None,
        out_features=None,
        act_layer=nn.GELU,
        drop=0.0,
    ):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x


class SpectralGatingNetwork(nn.Module):
    def __init__(self, dim, h=14, w=8):
        super().__init__()
        self.complex_weight = nn.Parameter(torch.randn(h, w, dim, 2) * 0.02)
        self.w = w
        self.h = h

    def forward(self, x, spatial_size=None):
        B, N, C = x.shape  # torch.Size([1, 262144, 1024])
        if spatial_size is None:
            a = b = int(math.sqrt(N))  # a=b=512
        else:
            a, b = spatial_size

        x = x.view(B, a, b, C)  # torch.Size([1, 512, 512, 1024])

        # FROM HERE USED TO BE AUTOCAST to float32
        dtype = x.dtype
        x = x.to(torch.float32)
        x = torch.fft.rfft2(
            x, dim=(1, 2), norm="ortho"
        )  # torch.Size([1, 512, 257, 1024])
        weight = torch.view_as_complex(
            self.complex_weight.to(torch.float32)
        )  # torch.Size([512, 257, 1024])
        x = x * weight
        x = torch.fft.irfft2(
            x, s=(a, b), dim=(1, 2), norm="ortho"
        )  # torch.Size([1, 512, 512, 1024])
        x = x.to(dtype)

        x = x.reshape(B, N, C)  # torch.Size([1, 262144, 1024])
        # UP TO HERE USED TO BE AUTOCAST to float32

        return x


class BlockSpectralGating(nn.Module):
    def __init__(
        self,
        dim,
        mlp_ratio=4.0,
        drop=0.0,
        drop_path=0.0,
        act_layer=nn.GELU,
        norm_layer=nn.LayerNorm,
        h=14,
        w=8,
    ):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.filter = SpectralGatingNetwork(dim, h=h, w=w)
        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(
            in_features=dim,
            hidden_features=mlp_hidden_dim,
            act_layer=act_layer,
            drop=drop,
        )

    def forward(self, x, *args):
        x = x + self.drop_path(self.mlp(self.norm2(self.filter(self.norm1(x)))))
        return x


class BlockAttention(nn.Module):
    def __init__(
        self,
        dim,
        num_heads: int = 8,
        mlp_ratio=4.0,
        drop=0.0,
        drop_path=0.0,
        w=2,
        dp_rank=2,
        act_layer=nn.GELU,
        norm_layer=nn.LayerNorm,
        rpe=False,
        adaLN=False,
        nglo=0,
    ):
        """
        num_heads: Attention heads. 4 for tiny, 8 for small and 12 for base
        """
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(
            in_features=dim,
            hidden_features=mlp_hidden_dim,
            act_layer=act_layer,
            drop=drop,
        )
        self.attn = AttentionLS(
            dim=dim,
            num_heads=num_heads,
            w=w,
            dp_rank=dp_rank,
            nglo=nglo,
            rpe=rpe,
        )

        if adaLN:
            self.adaLN_modulation = nn.Sequential(
                nn.Linear(dim, dim, bias=True),
                act_layer(),
                nn.Linear(dim, 6 * dim, bias=True),
            )
        else:
            self.adaLN_modulation = None

    def forward(self, x: torch.Tensor, c: torch.Tensor) -> torch.Tensor:
        if self.adaLN_modulation is not None:
            (
                shift_mha,
                scale_mha,
                gate_mha,
                shift_mlp,
                scale_mlp,
                gate_mlp,
            ) = self.adaLN_modulation(c).chunk(6, dim=2)
        else:
            shift_mha, scale_mha, gate_mha, shift_mlp, scale_mlp, gate_mlp = 6 * (1.0,)

        x = x + gate_mha * self.drop_path(
            self.attn(
                self.norm1(x) * scale_mha + shift_mha,
            )
        )
        x = x + gate_mlp * self.drop_path(
            self.mlp(self.norm2(x) * scale_mlp + shift_mlp)
        )

        return x


class SpectFormer(nn.Module):
    def __init__(
        self,
        grid_size: int = 224 // 16,
        embed_dim=768,
        depth=12,
        n_spectral_blocks=4,
        num_heads: int = 8,
        mlp_ratio=4.0,
        uniform_drop=False,
        drop_rate=0.0,
        drop_path_rate=0.0,
        window_size=2,
        dp_rank=2,
        norm_layer=nn.LayerNorm,
        checkpoint_layers: list[int] | None = None,
        rpe=False,
        ensemble: int | None = None,
        nglo: int = 0,
    ):
        """
        Args:
            img_size (int, tuple): input image size
            patch_size (int, tuple): patch size
            embed_dim (int): embedding dimension
            depth (int): depth of transformer
            n_spectral_blocks (int): number of spectral gating blocks
            mlp_ratio (int): ratio of mlp hidden dim to embedding dim
            uniform_drop (bool): true for uniform, false for linearly increasing drop path probability.
            drop_rate (float): dropout rate
            drop_path_rate (float): drop path (stochastic depth) rate
            window_size: window size for long/short attention
            dp_rank: dp rank for long/short attention
            norm_layer: (nn.Module): normalization layer for attention blocks
            checkpoint_layers: indicate which layers to use for checkpointing
            rpe: Use relative position encoding in Long-Short attention blocks.
            ensemble: Integer indicating ensemble size or None for deterministic model.
            nglo: Number of (additional) global tokens.
        """
        super().__init__()
        self.embed_dim = embed_dim
        self.n_spectral_blocks = n_spectral_blocks
        self._checkpoint_layers = checkpoint_layers or []
        self.ensemble = ensemble
        self.nglo = nglo

        h = grid_size
        w = h // 2 + 1

        if uniform_drop:
            _logger.info(f"Using uniform droppath with expect rate {drop_path_rate}.")
            dpr = [drop_path_rate for _ in range(depth)]
        else:
            _logger.info(
                f"Using linear droppath with expect rate {drop_path_rate * 0.5}."
            )
            dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]

        self.blocks_spectral_gating = nn.ModuleList()
        self.blocks_attention = nn.ModuleList()
        for i in range(depth):
            if i < n_spectral_blocks:
                layer = BlockSpectralGating(
                    dim=embed_dim,
                    mlp_ratio=mlp_ratio,
                    drop=drop_rate,
                    drop_path=dpr[i],
                    norm_layer=norm_layer,
                    h=h,
                    w=w,
                )
                self.blocks_spectral_gating.append(layer)
            else:
                layer = BlockAttention(
                    dim=embed_dim,
                    num_heads=num_heads,
                    mlp_ratio=mlp_ratio,
                    drop=drop_rate,
                    drop_path=dpr[i],
                    norm_layer=norm_layer,
                    w=window_size,
                    dp_rank=dp_rank,
                    rpe=rpe,
                    adaLN=True if ensemble is not None else False,
                    nglo=nglo,
                )
                self.blocks_attention.append(layer)

        self.apply(self._init_weights)

    def forward(self, tokens: torch.Tensor) -> torch.Tensor:
        """
        Args:
            tokens: Tensor of shape B, N, C for deterministic of BxE, N, C for ensemble forecast.
        Returns:
            Tensor of same shape as input.
        """
        if self.ensemble:
            BE, N, C = tokens.shape
            noise = torch.randn(
                size=(BE, N, C), dtype=tokens.dtype, device=tokens.device
            )
        else:
            noise = None

        for i, blk in enumerate(
            chain(self.blocks_spectral_gating, self.blocks_attention)
        ):
            if i in self._checkpoint_layers:
                tokens = checkpoint(blk, tokens, noise, use_reentrant=False)
            else:
                tokens = blk(tokens, noise)

        return tokens

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=0.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)