File size: 12,128 Bytes
e27c6bd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
383
384
385
386
387
"""

Vision Transformer (ViT) for Palette Feature Extraction



Implements a standard ViT with Samsung TRM best practices:

- RMS Normalization

- SwiGLU activation

- Truncated normal initialization

- Spatial feature preservation

"""

import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from typing import Tuple


# ============================================================================
# Helper functions (local copies)
# NOTE: These are intentionally local copies, NOT imported from transformer_layers.py.
# transformer_layers.py uses different parameter names (variance_epsilon vs eps,
# lower/upper vs a/b), CastedLinear instead of nn.Linear, and different SwiGLU
# expansion defaults. Callers here rely on the local signatures.
# ============================================================================

def rms_norm(hidden_states: torch.Tensor, eps: float = 1e-5) -> torch.Tensor:
    """

    RMS Normalization (more stable than LayerNorm)



    Args:

        hidden_states: Input tensor

        eps: Epsilon for numerical stability



    Returns:

        Normalized tensor

    """
    input_dtype = hidden_states.dtype
    hidden_states = hidden_states.to(torch.float32)

    variance = hidden_states.pow(2).mean(-1, keepdim=True)
    hidden_states = hidden_states * torch.rsqrt(variance + eps)

    return hidden_states.to(input_dtype)


def trunc_normal_init_(tensor: torch.Tensor, std: float = 1.0, a: float = -2, b: float = 2):
    """

    Truncated normal initialization (better than uniform)



    Args:

        tensor: Tensor to initialize

        std: Standard deviation

        a: Lower truncation bound (in std units)

        b: Upper truncation bound (in std units)



    Returns:

        Initialized tensor

    """
    with torch.no_grad():
        tensor.normal_(0, std)
        tensor.clamp_(min=a*std, max=b*std)
    return tensor


# ============================================================================
# SwiGLU Activation
# ============================================================================

class SwiGLU(nn.Module):
    """

    SwiGLU activation (Gated Linear Unit with Swish/SiLU)



    Superior to ReLU for expressiveness.

    Used in modern LLMs (LLaMA, PaLM, etc.)

    """

    def __init__(self, hidden_size: int, expansion: float = 2.0):
        super().__init__()

        # Compute intermediate dimension (round to multiple of 256 for efficiency)
        inter = int(expansion * hidden_size * 2 / 3)
        inter = ((inter + 255) // 256) * 256

        self.gate_up_proj = nn.Linear(hidden_size, inter * 2, bias=False)
        self.down_proj = nn.Linear(inter, hidden_size, bias=False)

    def forward(self, x):
        gate, up = self.gate_up_proj(x).chunk(2, dim=-1)
        return self.down_proj(F.silu(gate) * up)


# ============================================================================
# Multi-Head Self-Attention
# ============================================================================

class MultiHeadSelfAttention(nn.Module):
    """Multi-head self-attention for ViT"""

    def __init__(self, hidden_dim: int, num_heads: int = 8, dropout: float = 0.1, rms_eps: float = 1e-5):
        super().__init__()
        assert hidden_dim % num_heads == 0, "hidden_dim must be divisible by num_heads"

        self.hidden_dim = hidden_dim
        self.num_heads = num_heads
        self.head_dim = hidden_dim // num_heads
        self.rms_eps = rms_eps

        # Projections
        self.qkv_proj = nn.Linear(hidden_dim, hidden_dim * 3, bias=False)
        self.out_proj = nn.Linear(hidden_dim, hidden_dim, bias=False)

        self.dropout = nn.Dropout(dropout)
        self.scale = self.head_dim ** -0.5

        # Initialize with truncated normal
        self._init_weights()

    def _init_weights(self):
        """Initialize weights with truncated normal"""
        for module in [self.qkv_proj, self.out_proj]:
            std = 1.0 / math.sqrt(module.in_features)
            trunc_normal_init_(module.weight, std=std)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """

        Args:

            x: (B, N, D) input sequence



        Returns:

            (B, N, D) output sequence

        """
        B, N, D = x.shape

        # Project to Q, K, V
        qkv = self.qkv_proj(x)  # (B, N, 3*D)
        qkv = qkv.reshape(B, N, 3, self.num_heads, self.head_dim)
        qkv = qkv.permute(2, 0, 3, 1, 4)  # (3, B, H, N, d)
        Q, K, V = qkv[0], qkv[1], qkv[2]

        # Attention
        scores = torch.matmul(Q, K.transpose(-2, -1)) * self.scale
        attn_weights = F.softmax(scores, dim=-1)
        attn_weights = self.dropout(attn_weights)

        context = torch.matmul(attn_weights, V)

        # Merge heads
        context = context.transpose(1, 2).contiguous().view(B, N, D)
        output = self.out_proj(context)

        return output


# ============================================================================
# Transformer Block
# ============================================================================

class TransformerBlock(nn.Module):
    """

    Standard transformer block with RMS norm and SwiGLU

    """

    def __init__(

        self,

        hidden_dim: int,

        num_heads: int = 8,

        dropout: float = 0.1,

        swiglu_expansion: float = 2.0,

        rms_eps: float = 1e-5

    ):
        super().__init__()

        self.hidden_dim = hidden_dim
        self.rms_eps = rms_eps

        # Self-attention
        self.attention = MultiHeadSelfAttention(hidden_dim, num_heads, dropout, rms_eps)

        # Feed-forward with SwiGLU
        self.ffn = SwiGLU(hidden_dim, swiglu_expansion)

        self.dropout = nn.Dropout(dropout)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """

        Args:

            x: (B, N, D) input sequence



        Returns:

            (B, N, D) output sequence

        """
        # Attention with residual + RMS norm
        x_norm = rms_norm(x, eps=self.rms_eps)
        attn_out = self.attention(x_norm)
        x = x + self.dropout(attn_out)

        # FFN with residual + RMS norm
        x_norm = rms_norm(x, eps=self.rms_eps)
        ffn_out = self.ffn(x_norm)
        x = x + self.dropout(ffn_out)

        return x


# ============================================================================
# Vision Transformer
# ============================================================================

class VisionTransformer(nn.Module):
    """

    Vision Transformer for palette feature extraction



    Takes embedded palettes (B, H, W, D) and outputs spatial features (B, H, W, D)



    Architecture:

    - Patchify input (reduce spatial dimensions)

    - Apply transformer layers

    - Unpatchify back to original spatial dimensions



    Best practices from Samsung TRM:

    - RMS normalization

    - SwiGLU activation

    - Truncated normal initialization

    """

    def __init__(

        self,

        hidden_dim: int = 768,

        num_layers: int = 6,

        num_heads: int = 8,

        patch_size: int = 4,

        dropout: float = 0.1,

        rms_eps: float = 1e-5

    ):
        super().__init__()

        self.hidden_dim = hidden_dim
        self.num_layers = num_layers
        self.num_heads = num_heads
        self.patch_size = patch_size
        self.rms_eps = rms_eps

        # Patch embedding (reduce spatial dimensions)
        self.patch_embed = nn.Conv2d(
            hidden_dim, hidden_dim,
            kernel_size=patch_size,
            stride=patch_size,
            bias=False
        )

        # Transformer blocks
        self.blocks = nn.ModuleList([
            TransformerBlock(hidden_dim, num_heads, dropout, rms_eps=rms_eps)
            for _ in range(num_layers)
        ])

        # Unpatch (restore spatial dimensions)
        self.unpatch = nn.ConvTranspose2d(
            hidden_dim, hidden_dim,
            kernel_size=patch_size,
            stride=patch_size,
            bias=False
        )

        # Final normalization
        self.final_norm = lambda x: rms_norm(x, eps=rms_eps)

        # Initialize weights
        self._init_weights()

    def _init_weights(self):
        """Initialize all weights with truncated normal"""
        for module in self.modules():
            if isinstance(module, (nn.Linear, nn.Conv2d, nn.ConvTranspose2d)):
                std = 1.0 / math.sqrt(module.weight.shape[1] if len(module.weight.shape) > 1 else module.weight.shape[0])
                trunc_normal_init_(module.weight, std=std)
                if module.bias is not None:
                    module.bias.data.zero_()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """

        Extract spatial features from embedded palettes



        Args:

            x: (B, H, W, D) embedded palette



        Returns:

            (B, H, W, D) spatial features

        """
        B, H, W, D = x.shape

        # Rearrange for Conv2d: (B, H, W, D) → (B, D, H, W)
        x = x.permute(0, 3, 1, 2)

        # 1. Patchify: (B, D, H, W) → (B, D, H/P, W/P)
        x_patches = self.patch_embed(x)
        B, D, H_p, W_p = x_patches.shape

        # 2. Flatten patches: (B, D, H_p, W_p) → (B, N, D) where N = H_p * W_p
        x_seq = x_patches.flatten(2).transpose(1, 2)  # (B, N, D)

        # 3. Apply transformer blocks
        for block in self.blocks:
            x_seq = block(x_seq)

        # 4. Reshape back to patches: (B, N, D) → (B, D, H_p, W_p)
        x_patches = x_seq.transpose(1, 2).reshape(B, D, H_p, W_p)

        # 5. Unpatchify: (B, D, H_p, W_p) → (B, D, H, W)
        x_out = self.unpatch(x_patches)

        # 6. Final normalization
        # Normalize along feature dimension (D)
        x_out_norm = x_out.permute(0, 2, 3, 1)  # (B, H, W, D)
        x_out_norm = self.final_norm(x_out_norm)

        return x_out_norm


# ============================================================================
# Palette Embedding + ViT Pipeline
# ============================================================================

class PaletteFeatureExtractor(nn.Module):
    """

    Complete pipeline: Palette embedding → ViT → Features



    Combines:

    1. Token embedding (palette indices → continuous vectors)

    2. ViT feature extraction (spatial transformations)



    Input: (B, H, W) LongTensor palette indices

    Output: (B, H, W, D) FloatTensor features

    """

    def __init__(

        self,

        palette_size: int = 4096,

        hidden_dim: int = 768,

        num_layers: int = 6,

        num_heads: int = 8,

        patch_size: int = 4,

        dropout: float = 0.1

    ):
        super().__init__()

        self.palette_size = palette_size
        self.hidden_dim = hidden_dim

        # Token embedding
        self.palette_embed = nn.Embedding(palette_size, hidden_dim)

        # ViT
        self.vit = VisionTransformer(
            hidden_dim=hidden_dim,
            num_layers=num_layers,
            num_heads=num_heads,
            patch_size=patch_size,
            dropout=dropout
        )

        # Initialize embeddings
        self._init_embeddings()

    def _init_embeddings(self):
        """Initialize embedding with truncated normal"""
        std = 1.0 / math.sqrt(self.hidden_dim)
        trunc_normal_init_(self.palette_embed.weight, std=std)

    def forward(self, palette: torch.Tensor) -> torch.Tensor:
        """

        Extract features from palette



        Args:

            palette: (B, H, W) LongTensor palette indices



        Returns:

            (B, H, W, D) FloatTensor features

        """
        # Embed palette tokens
        x = self.palette_embed(palette)  # (B, H, W, D)

        # Extract features with ViT
        features = self.vit(x)  # (B, H, W, D)

        return features