File size: 13,541 Bytes
56238f0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn

from functools import lru_cache
from src.models.layers.attention_op import attention
from src.models.layers.rope import apply_rotary_emb, precompute_freqs_cis_ex2d as precompute_freqs_cis_2d
from src.models.layers.time_embed import TimestepEmbedder as TimestepEmbedder
from src.models.layers.patch_embed import Embed as Embed
from src.models.layers.swiglu import SwiGLU as FeedForward
from src.models.layers.rmsnorm import RMSNorm as Norm

def modulate(x, shift, scale):
    return x * (1 + scale) + shift

class Attention(nn.Module):
    def __init__(
            self,
            dim: int,
            num_heads: int = 8,
            qkv_bias: bool = False,
            attn_drop: float = 0.,
            proj_drop: float = 0.,
    ) -> None:
        super().__init__()
        assert dim % num_heads == 0, 'dim should be divisible by num_heads'

        self.dim = dim
        self.num_heads = num_heads
        self.head_dim = dim // num_heads
        self.scale = self.head_dim ** -0.5
        self.qkv_x = nn.Linear(dim, dim*3, bias=qkv_bias)
        self.kv_y = nn.Linear(dim, dim*2, bias=qkv_bias)

        self.q_norm = Norm(self.head_dim)
        self.k_norm = Norm(self.head_dim)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x: torch.Tensor, y, pos) -> torch.Tensor:
        B, N, C = x.shape
        qkv_x = self.qkv_x(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        q, kx, vx = qkv_x[0], qkv_x[1], qkv_x[2]
        q = self.q_norm(q.contiguous())
        kx = self.k_norm(kx.contiguous())
        q, kx = apply_rotary_emb(q, kx, freqs_cis=pos)
        kv_y = self.kv_y(y).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        ky, vy = kv_y[0], kv_y[1]
        ky = self.k_norm(ky.contiguous())

        k = torch.cat([kx, ky], dim=2)
        v = torch.cat([vx, vy], dim=2)

        q = q.view(B, self.num_heads, -1, C // self.num_heads)  # B, H, N, Hc
        k = k.view(B, self.num_heads, -1, C // self.num_heads).contiguous()  # B, H, N, Hc
        v = v.view(B, self.num_heads, -1, C // self.num_heads).contiguous()

        x = attention(q, k, v)
        x = x.transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x

class FlattenDiTBlock(nn.Module):
    def __init__(self, hidden_size, groups,  mlp_ratio=4, ):
        super().__init__()
        self.norm1 = Norm(hidden_size, eps=1e-6)
        self.attn = Attention(hidden_size, num_heads=groups, qkv_bias=False)
        self.norm2 = Norm(hidden_size, eps=1e-6)
        mlp_hidden_dim = int(hidden_size * mlp_ratio)
        self.mlp = FeedForward(hidden_size, mlp_hidden_dim)
        self.adaLN_modulation = nn.Sequential(
            nn.Linear(hidden_size, 6 * hidden_size, bias=True)
        )

    def forward(self, x, y, c, pos):
        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=-1)
        x = x + gate_msa * self.attn(modulate(self.norm1(x), shift_msa, scale_msa), y, pos)
        x = x + gate_mlp * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
        return x

class NerfEmbedder(nn.Module):
    def __init__(self, in_channels, hidden_size_input, max_freqs):
        super().__init__()
        self.max_freqs = max_freqs
        self.hidden_size_input = hidden_size_input
        self.embedder = nn.Sequential(
            nn.Linear(in_channels+max_freqs**2, hidden_size_input, bias=True),
        )

    @lru_cache
    def fetch_pos(self, patch_size_h, patch_size_w, device, dtype):
        pos = precompute_freqs_cis_2d(self.max_freqs ** 2 * 2, patch_size_h, patch_size_w, scale=(16/patch_size_h, 16/patch_size_w))
        pos = pos[None, :, :].to(device=device, dtype=dtype)
        return pos


    def forward(self, inputs, patch_size_h, patch_size_w):
        B, _, C = inputs.shape
        device = inputs.device
        dtype = inputs.dtype
        dct = self.fetch_pos(patch_size_h, patch_size_w, device, dtype)
        dct = dct.repeat(B, 1, 1)
        inputs = torch.cat([inputs, dct], dim=-1)
        inputs = self.embedder(inputs)
        return inputs

class NerfBlock(nn.Module):
    def __init__(self, hidden_size_s, hidden_size_x, mlp_ratio=4):
        super().__init__()
        self.param_generator1 = nn.Sequential(
            nn.Linear(hidden_size_s, 2*hidden_size_x**2*mlp_ratio, bias=True),
        )
        self.norm = Norm(hidden_size_x, eps=1e-6)
        self.mlp_ratio = mlp_ratio
    def forward(self, x, s):
        batch_size, num_x, hidden_size_x = x.shape
        mlp_params1 = self.param_generator1(s)
        fc1_param1, fc2_param1 = mlp_params1.chunk(2, dim=-1)
        fc1_param1 = fc1_param1.view(batch_size, hidden_size_x, hidden_size_x*self.mlp_ratio)
        fc2_param1 = fc2_param1.view(batch_size, hidden_size_x*self.mlp_ratio, hidden_size_x)

        # normalize fc1
        normalized_fc1_param1 = torch.nn.functional.normalize(fc1_param1, dim=-2)
        # mlp 1
        res_x = x
        x = self.norm(x)
        x = torch.bmm(x, normalized_fc1_param1)
        x = torch.nn.functional.silu(x)
        x = torch.bmm(x, fc2_param1)
        x = x + res_x
        return x

class NerfFinalLayer(nn.Module):
    def __init__(self, hidden_size, out_channels):
        super().__init__()
        self.linear = nn.Linear(hidden_size, out_channels, bias=True)
    def forward(self, x):
        x = self.linear(x)
        return x

class TextRefineAttention(nn.Module):
    def __init__(
            self,
            dim: int,
            num_heads: int = 8,
            qkv_bias: bool = False,
            attn_drop: float = 0.,
            proj_drop: float = 0.,
    ) -> None:
        super().__init__()
        assert dim % num_heads == 0, 'dim should be divisible by num_heads'
        self.dim = dim
        self.num_heads = num_heads
        self.head_dim = dim // num_heads
        self.scale = self.head_dim ** -0.5
        self.qkv = nn.Linear(dim, dim*3, bias=qkv_bias)
        self.q_norm = Norm(self.head_dim)
        self.k_norm = Norm(self.head_dim)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        B, N, C = x.shape
        qkv_x = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        q, k, v = qkv_x[0], qkv_x[1], qkv_x[2]
        q = self.q_norm(q)
        k = self.k_norm(k)
        q = q.view(B, self.num_heads, -1, C // self.num_heads)  # B, H, N, Hc
        k = k.view(B, self.num_heads, -1, C // self.num_heads).contiguous()  # B, H, N, Hc
        v = v.view(B, self.num_heads, -1, C // self.num_heads).contiguous()
        x = attention(q, k, v)
        x = x.transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x

class TextRefineBlock(nn.Module):
    def __init__(self, hidden_size, groups,  mlp_ratio=4, ):
        super().__init__()
        self.norm1 = Norm(hidden_size, eps=1e-6)
        self.attn = TextRefineAttention(hidden_size, num_heads=groups, qkv_bias=False)
        self.norm2 = Norm(hidden_size, eps=1e-6)
        mlp_hidden_dim = int(hidden_size * mlp_ratio)
        self.mlp = FeedForward(hidden_size, mlp_hidden_dim)

        self.adaLN_modulation = nn.Sequential(
            nn.Linear(hidden_size, 6 * hidden_size, bias=True)
        )

    def forward(self, x, c):
        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=-1)
        x = x + gate_msa * self.attn(modulate(self.norm1(x), shift_msa, scale_msa))
        x = x + gate_mlp * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
        return x


class PixNerDiT(nn.Module):
    def __init__(
            self,
            in_channels=4,
            num_groups=12,
            hidden_size=1152,
            decoder_hidden_size=64,
            num_encoder_blocks=18,
            num_decoder_blocks=4,
            num_text_blocks=4,
            patch_size=2,
            txt_embed_dim=1024,
            txt_max_length=100,
            weight_path=None,
            load_ema=False,
    ):
        super().__init__()
        self.in_channels = in_channels
        self.out_channels = in_channels
        self.hidden_size = hidden_size
        self.num_groups = num_groups
        self.decoder_hidden_size = decoder_hidden_size
        self.num_encoder_blocks = num_encoder_blocks
        self.num_decoder_blocks = num_decoder_blocks
        self.num_blocks = self.num_encoder_blocks + self.num_decoder_blocks
        self.num_text_blocks = num_text_blocks
        self.decoder_patch_scaling_h = 1.0
        self.decoder_patch_scaling_w = 1.0
        self.patch_size = patch_size
        self.txt_embed_dim = txt_embed_dim
        self.txt_max_length = txt_max_length
        self.s_embedder = Embed(in_channels*patch_size**2, hidden_size, bias=True)
        self.x_embedder = NerfEmbedder(in_channels, decoder_hidden_size, max_freqs=8)
        self.t_embedder = TimestepEmbedder(hidden_size)
        self.y_embedder = Embed(txt_embed_dim, hidden_size, bias=True, norm_layer=Norm)
        self.y_pos_embedding = torch.nn.Parameter(
            torch.randn(1, txt_max_length, hidden_size),
            requires_grad=True
        )
        self.final_layer = NerfFinalLayer(decoder_hidden_size, in_channels)
        encoder_blocks = nn.ModuleList([
            FlattenDiTBlock(self.hidden_size, self.num_groups) for _ in range(self.num_encoder_blocks)
        ])
        decoder_blocks = nn.ModuleList([
            NerfBlock(self.hidden_size, self.decoder_hidden_size, mlp_ratio=2) for _ in range(self.num_decoder_blocks)
        ])
        self.blocks = nn.ModuleList(encoder_blocks + decoder_blocks)
        self.text_refine_blocks = nn.ModuleList([
            TextRefineBlock(self.hidden_size, self.num_groups) for _ in range(self.num_text_blocks)
        ])
        self.initialize_weights()
        self.precompute_pos = dict()
        self.weight_path = weight_path
        self.load_ema = load_ema

    def fetch_pos(self, height, width, device):
        if (height, width) in self.precompute_pos:
            return self.precompute_pos[(height, width)].to(device)
        else:
            pos = precompute_freqs_cis_2d(self.hidden_size // self.num_groups, height, width).to(device)
            self.precompute_pos[(height, width)] = pos
            return pos

    def initialize_weights(self):
        # Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
        w = self.s_embedder.proj.weight.data
        nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
        nn.init.constant_(self.s_embedder.proj.bias, 0)

        # Initialize timestep embedding MLP:
        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)

        # Zero-out output layers:
        nn.init.constant_(self.final_layer.linear.weight, 0)
        nn.init.constant_(self.final_layer.linear.bias, 0)

    def forward(self, x, t, y):
        B, _, H, W = x.shape
        encoder_h, encoder_w = int(H/self.decoder_patch_scaling_h), int(W/self.decoder_patch_scaling_w)
        decoder_patch_size_h = int(self.patch_size * self.decoder_patch_scaling_h)
        decoder_patch_size_w = int(self.patch_size * self.decoder_patch_scaling_w)
        x_for_encoder = torch.nn.functional.interpolate(x, (encoder_h, encoder_w))

        x_for_encoder = torch.nn.functional.unfold(x_for_encoder, kernel_size=self.patch_size, stride=self.patch_size).transpose(1, 2)
        x_for_decoder = torch.nn.functional.unfold(x, kernel_size=(decoder_patch_size_h, decoder_patch_size_w), stride=(decoder_patch_size_h, decoder_patch_size_w)).transpose(1, 2)
        xpos = self.fetch_pos(encoder_h // self.patch_size, encoder_w // self.patch_size, x.device)
        ypos = self.y_pos_embedding
        t = self.t_embedder(t.view(-1)).view(B, -1, self.hidden_size)
        y = self.y_embedder(y).view(B, -1, self.hidden_size) + ypos.to(y.dtype)

        condition = nn.functional.silu(t)
        for i, block in enumerate(self.text_refine_blocks):
            y = block(y, condition)


        s = self.s_embedder(x_for_encoder)
        for i in range(self.num_encoder_blocks):
            s = self.blocks[i](s, y, condition, xpos)

        s = torch.nn.functional.silu(t + s)
        batch_size, length, _ = s.shape
        x = x_for_decoder.reshape(batch_size * length, self.in_channels, decoder_patch_size_h * decoder_patch_size_w)
        x = x.transpose(1, 2)
        s = s.view(batch_size * length, self.hidden_size)
        x = self.x_embedder(x, decoder_patch_size_h, decoder_patch_size_w)

        for i in range(self.num_decoder_blocks):
            def checkpoint_forward(x, s, block=self.blocks[i + self.num_encoder_blocks]):
                return block(x, s)
            x = checkpoint_forward(x, s)
        x = self.final_layer(x)
        x = x.transpose(1, 2)
        x = x.reshape(batch_size, length, -1)
        x = torch.nn.functional.fold(x.transpose(1, 2).contiguous(),
                                     (H, W),
                                     kernel_size=(decoder_patch_size_h, decoder_patch_size_w),
                                     stride=(decoder_patch_size_h, decoder_patch_size_w))
        return x