File size: 11,160 Bytes
53becf5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Pure-PyTorch engineering reproduction of the Clay v1.5 model specification."""

import math

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


def fourier_encode(values, dim, max_frequency=10000.0):
    """Encode scalar metadata while preserving exactly ``dim`` output features."""
    if dim < 1:
        return values.new_zeros(*values.shape, 0)
    pairs = (dim + 1) // 2
    frequencies = torch.exp(
        torch.linspace(0, math.log(max_frequency), pairs, device=values.device, dtype=values.dtype)
    )
    angles = values.unsqueeze(-1) * frequencies
    return torch.cat((angles.sin(), angles.cos()), dim=-1)[..., :dim]


def position_encoding_2d(height, width, dim, gsd, device, dtype):
    if dim % 4:
        raise ValueError("spatial position dimension must be divisible by four")
    y, x = torch.meshgrid(
        torch.arange(height, device=device, dtype=dtype),
        torch.arange(width, device=device, dtype=dtype),
        indexing="ij",
    )
    scale = torch.as_tensor(gsd, device=device, dtype=dtype) / 10.0
    quarter = dim // 4
    frequencies = torch.exp(
        torch.arange(quarter, device=device, dtype=dtype) * (-math.log(10000.0) / max(quarter, 1))
    )
    x_angles = x.reshape(-1, 1) * scale * frequencies
    y_angles = y.reshape(-1, 1) * scale * frequencies
    return torch.cat((x_angles.sin(), x_angles.cos(), y_angles.sin(), y_angles.cos()), dim=-1)


def patchify(pixels, patch_size):
    batch, channels, height, width = pixels.shape
    if height % patch_size or width % patch_size:
        raise ValueError("image dimensions must be divisible by patch_size")
    return pixels.reshape(
        batch, channels, height // patch_size, patch_size, width // patch_size, patch_size
    ).permute(0, 2, 4, 1, 3, 5).reshape(batch, -1, channels * patch_size * patch_size)


def unpatchify(patches, channels, height, width, patch_size):
    batch = patches.shape[0]
    return patches.reshape(
        batch, height // patch_size, width // patch_size, channels, patch_size, patch_size
    ).permute(0, 3, 1, 4, 2, 5).reshape(batch, channels, height, width)


class Transformer(nn.Module):
    def __init__(self, dim, depth, heads, mlp_ratio=4):
        super().__init__()
        layer = nn.TransformerEncoderLayer(
            dim, heads, int(dim * mlp_ratio), activation="gelu", batch_first=True, norm_first=True
        )
        self.layers = nn.TransformerEncoder(layer, depth)
        self.norm = nn.LayerNorm(dim)

    def forward(self, values):
        return self.norm(self.layers(values))


class DynamicEmbedding(nn.Module):
    """Create sensor-agnostic patches by conditioning per-band kernels on wavelength."""

    def __init__(self, patch_size, embed_dim, wave_dim, wave_latents):
        super().__init__()
        self.patch_size = patch_size
        self.wave_dim = wave_dim
        self.wave_mlp = nn.Sequential(nn.Linear(wave_dim, wave_dim), nn.GELU(), nn.Linear(wave_dim, wave_dim))
        self.latents = nn.Parameter(torch.randn(wave_latents, wave_dim) * 0.02)
        self.cross_attention = nn.MultiheadAttention(wave_dim, 4, batch_first=True)
        self.kernel = nn.Linear(wave_dim, patch_size * patch_size * embed_dim)
        self.bias = nn.Parameter(torch.zeros(embed_dim))
        self.embed_dim = embed_dim

    def forward(self, pixels, wavelengths):
        batch, channels, height, width = pixels.shape
        if wavelengths.ndim == 1:
            wavelengths = wavelengths[None].expand(batch, -1)
        if wavelengths.shape != (batch, channels):
            raise ValueError(f"wavelengths must have shape {(batch, channels)}, got {tuple(wavelengths.shape)}")
        wave_features = fourier_encode(wavelengths / 1000.0, self.wave_dim)
        wave_features = self.wave_mlp(wave_features)
        queries = self.latents[None].expand(batch, -1, -1)
        context = self.cross_attention(queries, wave_features, wave_features, need_weights=False)[0].mean(dim=1)
        conditioned = wave_features + context[:, None]
        kernels = self.kernel(conditioned).reshape(
            batch, channels, self.embed_dim, self.patch_size, self.patch_size
        )
        patches = []
        for index in range(batch):
            patches.append(F.conv2d(pixels[index:index + 1], kernels[index].permute(1, 0, 2, 3),
                                    stride=self.patch_size) + self.bias[None, :, None, None])
        return torch.cat(patches).flatten(2).transpose(1, 2), conditioned


class DynamicDecoder(nn.Module):
    def __init__(self, patch_size, decoder_dim, wave_dim):
        super().__init__()
        self.patch_size = patch_size
        self.wave_mlp = nn.Sequential(nn.Linear(wave_dim, decoder_dim), nn.GELU(), nn.Linear(decoder_dim, decoder_dim))
        self.output = nn.Linear(decoder_dim, patch_size * patch_size)

    def forward(self, tokens, wave_features):
        wave_context = self.wave_mlp(wave_features)
        joint = tokens[:, :, None, :] + wave_context[:, None, :, :]
        return self.output(joint).permute(0, 1, 2, 3).flatten(2)


class ClayFoundation(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = dict(config)
        self.patch_size = int(config["patch_size"])
        self.mask_ratio = float(config["mask_ratio"])
        enc_dim, dec_dim = int(config["encoder_dim"]), int(config["decoder_dim"])
        if enc_dim < 12 or (enc_dim - 8) % 4:
            raise ValueError("encoder_dim - 8 must be positive and divisible by four")
        if dec_dim < 12 or (dec_dim - 8) % 4:
            raise ValueError("decoder_dim - 8 must be positive and divisible by four")
        self.dynamic_embedding = DynamicEmbedding(
            self.patch_size, enc_dim, int(config["wave_dim"]), int(config["wave_latents"])
        )
        self.cls_token = nn.Parameter(torch.randn(1, 1, enc_dim) * 0.02)
        self.encoder = Transformer(enc_dim, int(config["encoder_depth"]), int(config["encoder_heads"]))
        self.encoder_to_decoder = nn.Linear(enc_dim, dec_dim)
        self.mask_token = nn.Parameter(torch.randn(1, 1, dec_dim) * 0.02)
        self.decoder = Transformer(dec_dim, int(config["decoder_depth"]), int(config["decoder_heads"]))
        self.wave_to_decoder = nn.Linear(int(config["wave_dim"]), int(config["wave_dim"]))
        self.dynamic_decoder = DynamicDecoder(self.patch_size, dec_dim, int(config["wave_dim"]))
        self.representation_head = nn.Linear(enc_dim, int(config["teacher_dim"]))
        self.norm_pix_loss = bool(config.get("norm_pix_loss", False))

    @staticmethod
    def _metadata_encoding(time, latlon, dim):
        values = torch.cat((time, latlon), dim=1)
        widths = [dim // 4] * 4
        for index in range(dim % 4):
            widths[index] += 1
        return torch.cat([fourier_encode(values[:, index], widths[index]) for index in range(4)], dim=1)

    def _add_encoding(self, tokens, time, latlon, gsd):
        batch, length, dim = tokens.shape
        grid = int(math.sqrt(length))
        if grid * grid != length:
            raise ValueError("Clay reproduction requires a square patch grid")
        spatial = position_encoding_2d(grid, grid, dim - 8, gsd, tokens.device, tokens.dtype)
        metadata = self._metadata_encoding(time, latlon, 8)
        encoding = torch.cat((spatial[None].expand(batch, -1, -1), metadata[:, None].expand(-1, length, -1)), dim=-1)
        return tokens + encoding

    def encode(self, pixels, time, latlon, gsd, wavelengths):
        patches, _ = self.dynamic_embedding(pixels, wavelengths)
        patches = self._add_encoding(patches, time, latlon, gsd)
        cls = self.cls_token.expand(len(pixels), -1, -1)
        encoded = self.encoder(torch.cat((cls, patches), dim=1))
        return encoded[:, 0], encoded[:, 1:]

    def forward(self, pixels, time, latlon, gsd, wavelengths, teacher_target=None, mask_ratio=None):
        ratio = self.mask_ratio if mask_ratio is None else float(mask_ratio)
        patches, wave_features = self.dynamic_embedding(pixels, wavelengths)
        patches = self._add_encoding(patches, time, latlon, gsd)
        batch, length, _ = patches.shape
        keep = max(1, length - int(length * ratio))
        noise = torch.rand(batch, length, device=pixels.device)
        ordering = noise.argsort(dim=1)
        unmasked_indices, masked_indices = ordering[:, :keep], ordering[:, keep:]
        gather = unmasked_indices[:, :, None].expand(-1, -1, patches.shape[-1])
        visible = patches.gather(1, gather)
        encoded = self.encoder(torch.cat((self.cls_token.expand(batch, -1, -1), visible), dim=1))
        embedding = encoded[:, 0]

        decoded_visible = self.encoder_to_decoder(encoded[:, 1:])
        decoder_tokens = self.mask_token.expand(batch, length, -1).clone()
        decoder_tokens.scatter_(1, unmasked_indices[:, :, None].expand(-1, -1, decoded_visible.shape[-1]), decoded_visible)
        grid = int(math.sqrt(length))
        spatial = position_encoding_2d(grid, grid, decoder_tokens.shape[-1] - 8, gsd,
                                       decoder_tokens.device, decoder_tokens.dtype)
        metadata = self._metadata_encoding(time, latlon, 8)
        decoder_tokens = decoder_tokens + torch.cat((spatial[None].expand(batch, -1, -1),
                                                       metadata[:, None].expand(-1, length, -1)), dim=-1)
        decoded = self.decoder(decoder_tokens)
        predicted_patches = self.dynamic_decoder(decoded, self.wave_to_decoder(wave_features))
        target_patches = patchify(pixels, self.patch_size)
        if self.norm_pix_loss:
            mean = target_patches.mean(dim=-1, keepdim=True)
            variance = target_patches.var(dim=-1, keepdim=True)
            target_patches = (target_patches - mean) / (variance + 1e-6).sqrt()
        mask = torch.zeros(batch, length, device=pixels.device)
        mask.scatter_(1, masked_indices, 1.0)
        patch_loss = (predicted_patches - target_patches).abs().mean(dim=-1)
        reconstruction_loss = (patch_loss * mask).sum() / mask.sum().clamp_min(1)
        projected = F.normalize(self.representation_head(embedding), dim=1)
        if teacher_target is None:
            representation_loss = embedding.new_zeros(())
        else:
            representation_loss = 1.0 - (projected * F.normalize(teacher_target, dim=1)).sum(dim=1).mean()
        reconstruction = unpatchify(predicted_patches, pixels.shape[1], pixels.shape[2], pixels.shape[3], self.patch_size)
        return {
            "embedding": embedding,
            "projected_embedding": projected,
            "reconstruction": reconstruction,
            "mask": mask,
            "reconstruction_loss": reconstruction_loss,
            "representation_loss": representation_loss,
        }


def compute_loss(outputs, reconstruction_weight=0.95, representation_weight=0.05):
    total = reconstruction_weight * outputs["reconstruction_loss"] + representation_weight * outputs["representation_loss"]
    return total, {
        "reconstruction": outputs["reconstruction_loss"],
        "representation": outputs["representation_loss"],
        "total": total,
    }