Spaces:
Moustached
/
Runtime error

File size: 13,097 Bytes
f9a674e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from inspect import isfunction
import math
import torch
import torch.nn.functional as F
from torch import nn, einsum
from einops import rearrange, repeat

# from ldm.modules.diffusionmodules.util import checkpoint, FourierEmbedder
from torch.utils import checkpoint

try:
    import xformers
    import xformers.ops
    XFORMERS_IS_AVAILBLE = True
except:
    XFORMERS_IS_AVAILBLE = False


def exists(val):
    return val is not None


def uniq(arr):
    return{el: True for el in arr}.keys()


def default(val, d):
    if exists(val):
        return val
    return d() if isfunction(d) else d


def max_neg_value(t):
    return -torch.finfo(t.dtype).max


def init_(tensor):
    dim = tensor.shape[-1]
    std = 1 / math.sqrt(dim)
    tensor.uniform_(-std, std)
    return tensor


# feedforward
class GEGLU(nn.Module):
    def __init__(self, dim_in, dim_out):
        super().__init__()
        self.proj = nn.Linear(dim_in, dim_out * 2)

    def forward(self, x):
        x, gate = self.proj(x).chunk(2, dim=-1)
        return x * F.gelu(gate)


class FeedForward(nn.Module):
    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
        super().__init__()
        inner_dim = int(dim * mult)
        dim_out = default(dim_out, dim)
        project_in = nn.Sequential(
            nn.Linear(dim, inner_dim),
            nn.GELU()
        ) if not glu else GEGLU(dim, inner_dim)

        self.net = nn.Sequential(
            project_in,
            nn.Dropout(dropout),
            nn.Linear(inner_dim, dim_out)
        )

    def forward(self, x):
        return self.net(x)


def zero_module(module):
    """
    Zero out the parameters of a module and return it.
    """
    for p in module.parameters():
        p.detach().zero_()
    return module


def Normalize(in_channels):
    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)


class LinearAttention(nn.Module):
    def __init__(self, dim, heads=4, dim_head=32):
        super().__init__()
        self.heads = heads
        hidden_dim = dim_head * heads
        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
        self.to_out = nn.Conv2d(hidden_dim, dim, 1)

    def forward(self, x):
        b, c, h, w = x.shape
        qkv = self.to_qkv(x)
        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
        k = k.softmax(dim=-1)  
        context = torch.einsum('bhdn,bhen->bhde', k, v)
        out = torch.einsum('bhde,bhdn->bhen', context, q)
        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
        return self.to_out(out)




class CrossAttention(nn.Module):
    def __init__(self, query_dim, key_dim, value_dim, heads=8, dim_head=64, dropout=0):
        super().__init__()
        inner_dim = dim_head * heads
        self.scale = dim_head ** -0.5
        self.heads = heads
        self.dim_head = dim_head

        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
        self.to_k = nn.Linear(key_dim, inner_dim, bias=False)
        self.to_v = nn.Linear(value_dim, inner_dim, bias=False)


        self.to_out = nn.Sequential( nn.Linear(inner_dim, query_dim), nn.Dropout(dropout) )


    def fill_inf_from_mask(self, sim, mask):
        if mask is not None:
            B,M = mask.shape
            mask = mask.unsqueeze(1).repeat(1,self.heads,1).reshape(B*self.heads,1,-1)
            max_neg_value = -torch.finfo(sim.dtype).max
            sim.masked_fill_(~mask, max_neg_value)
        return sim 

    def forward_plain(self, x, key, value, mask=None):

        q = self.to_q(x)     # B*N*(H*C)
        k = self.to_k(key)   # B*M*(H*C)
        v = self.to_v(value) # B*M*(H*C)
   
        B, N, HC = q.shape 
        _, M, _ = key.shape
        H = self.heads
        C = HC // H 

        q = q.view(B,N,H,C).permute(0,2,1,3).reshape(B*H,N,C) # (B*H)*N*C
        k = k.view(B,M,H,C).permute(0,2,1,3).reshape(B*H,M,C) # (B*H)*M*C
        v = v.view(B,M,H,C).permute(0,2,1,3).reshape(B*H,M,C) # (B*H)*M*C

        sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale # (B*H)*N*M
        self.fill_inf_from_mask(sim, mask)
        attn = sim.softmax(dim=-1) # (B*H)*N*M

        out = torch.einsum('b i j, b j d -> b i d', attn, v) # (B*H)*N*C
        out = out.view(B,H,N,C).permute(0,2,1,3).reshape(B,N,(H*C)) # B*N*(H*C)

        return self.to_out(out)

    def forward(self, x, key, value, mask=None):
        if not XFORMERS_IS_AVAILBLE:
            return self.forward_plain(x, key, value, mask)

        q = self.to_q(x)     # B*N*(H*C)
        k = self.to_k(key)   # B*M*(H*C)
        v = self.to_v(value) # B*M*(H*C)

        b, _, _ = q.shape
        q, k, v = map(
            lambda t: t.unsqueeze(3)
            .reshape(b, t.shape[1], self.heads, self.dim_head)
            .permute(0, 2, 1, 3)
            .reshape(b * self.heads, t.shape[1], self.dim_head)
            .contiguous(),
            (q, k, v),
        )

        # actually compute the attention, what we cannot get enough of
        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=None)

        if exists(mask):
            raise NotImplementedError
        out = (
            out.unsqueeze(0)
            .reshape(b, self.heads, out.shape[1], self.dim_head)
            .permute(0, 2, 1, 3)
            .reshape(b, out.shape[1], self.heads * self.dim_head)
        )
        return self.to_out(out)





class SelfAttention(nn.Module):
    def __init__(self, query_dim, heads=8, dim_head=64, dropout=0.):
        super().__init__()
        inner_dim = dim_head * heads
        self.scale = dim_head ** -0.5
        self.heads = heads
        self.dim_head = dim_head

        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
        self.to_k = nn.Linear(query_dim, inner_dim, bias=False)
        self.to_v = nn.Linear(query_dim, inner_dim, bias=False)

        self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout) )

    def forward_plain(self, x):
        q = self.to_q(x) # B*N*(H*C)
        k = self.to_k(x) # B*N*(H*C)
        v = self.to_v(x) # B*N*(H*C)

        B, N, HC = q.shape 
        H = self.heads
        C = HC // H 

        q = q.view(B,N,H,C).permute(0,2,1,3).reshape(B*H,N,C) # (B*H)*N*C
        k = k.view(B,N,H,C).permute(0,2,1,3).reshape(B*H,N,C) # (B*H)*N*C
        v = v.view(B,N,H,C).permute(0,2,1,3).reshape(B*H,N,C) # (B*H)*N*C

        sim = torch.einsum('b i c, b j c -> b i j', q, k) * self.scale  # (B*H)*N*N
        attn = sim.softmax(dim=-1) # (B*H)*N*N

        out = torch.einsum('b i j, b j c -> b i c', attn, v) # (B*H)*N*C
        out = out.view(B,H,N,C).permute(0,2,1,3).reshape(B,N,(H*C)) # B*N*(H*C)

        return self.to_out(out)

    def forward(self, x, context=None, mask=None):
        if not XFORMERS_IS_AVAILBLE:
            return self.forward_plain(x)

        q = self.to_q(x)
        context = default(context, x)
        k = self.to_k(context)
        v = self.to_v(context)

        b, _, _ = q.shape
        q, k, v = map(
            lambda t: t.unsqueeze(3)
            .reshape(b, t.shape[1], self.heads, self.dim_head)
            .permute(0, 2, 1, 3)
            .reshape(b * self.heads, t.shape[1], self.dim_head)
            .contiguous(),
            (q, k, v),
        )

        # actually compute the attention, what we cannot get enough of
        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=None)

        if exists(mask):
            raise NotImplementedError
        out = (
            out.unsqueeze(0)
            .reshape(b, self.heads, out.shape[1], self.dim_head)
            .permute(0, 2, 1, 3)
            .reshape(b, out.shape[1], self.heads * self.dim_head)
        )
        return self.to_out(out)


class GatedCrossAttentionDense(nn.Module):
    def __init__(self, query_dim, key_dim, value_dim, n_heads, d_head):
        super().__init__()
        
        self.attn = CrossAttention(query_dim=query_dim, key_dim=key_dim, value_dim=value_dim, heads=n_heads, dim_head=d_head) 
        self.ff = FeedForward(query_dim, glu=True)

        self.norm1 = nn.LayerNorm(query_dim)
        self.norm2 = nn.LayerNorm(query_dim)

        self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.)) )
        self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.)) )

        # this can be useful: we can externally change magnitude of tanh(alpha)
        # for example, when it is set to 0, then the entire model is same as original one 
        self.scale = 1  

    def forward(self, x, objs):

        x = x + self.scale*torch.tanh(self.alpha_attn) * self.attn( self.norm1(x), objs, objs)  
        x = x + self.scale*torch.tanh(self.alpha_dense) * self.ff( self.norm2(x) ) 
        
        return x 


class GatedSelfAttentionDense(nn.Module):
    def __init__(self, query_dim, context_dim,  n_heads, d_head):
        super().__init__()
        
        # we need a linear projection since we need cat visual feature and obj feature
        self.linear = nn.Linear(context_dim, query_dim)

        self.attn = SelfAttention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
        self.ff = FeedForward(query_dim, glu=True)

        self.norm1 = nn.LayerNorm(query_dim)
        self.norm2 = nn.LayerNorm(query_dim)

        self.register_parameter('alpha_attn', nn.Parameter(torch.tensor(0.)) )
        self.register_parameter('alpha_dense', nn.Parameter(torch.tensor(0.)) )

        # this can be useful: we can externally change magnitude of tanh(alpha)
        # for example, when it is set to 0, then the entire model is same as original one 
        self.scale = 1  


    def forward(self, x, objs):

        N_visual = x.shape[1]
        objs = self.linear(objs)
        
        x = x + self.scale*torch.tanh(self.alpha_attn) * self.attn(  self.norm1(torch.cat([x,objs],dim=1))  )[:,0:N_visual,:]
        x = x + self.scale*torch.tanh(self.alpha_dense) * self.ff( self.norm2(x) )  
        
        return x 


class BasicTransformerBlock(nn.Module):
    def __init__(self, query_dim, key_dim, value_dim, n_heads, d_head, fuser_type, use_checkpoint=True):
        super().__init__()
        self.attn1 = SelfAttention(query_dim=query_dim, heads=n_heads, dim_head=d_head)  
        self.ff = FeedForward(query_dim, glu=True)
        self.attn2 = CrossAttention(query_dim=query_dim, key_dim=key_dim, value_dim=value_dim, heads=n_heads, dim_head=d_head)  
        self.norm1 = nn.LayerNorm(query_dim)
        self.norm2 = nn.LayerNorm(query_dim)
        self.norm3 = nn.LayerNorm(query_dim)
        self.use_checkpoint = use_checkpoint

        if fuser_type == "gatedSA":
            # note key_dim here actually is context_dim
            self.fuser = GatedSelfAttentionDense(query_dim, key_dim, n_heads, d_head) 
        elif fuser_type == "gatedCA":
            self.fuser = GatedCrossAttentionDense(query_dim, key_dim, value_dim, n_heads, d_head) 
        else:
            assert False 


    def forward(self, x, context, objs):
#        return checkpoint(self._forward, (x, context, objs), self.parameters(), self.use_checkpoint)
        if self.use_checkpoint and x.requires_grad:
            return checkpoint.checkpoint(self._forward, x, context, objs)
        else:
            return self._forward(x, context, objs)

    def _forward(self, x, context, objs): 
        x = self.attn1( self.norm1(x) ) + x 
        x = self.fuser(x, objs) # identity mapping in the beginning 
        x = self.attn2(self.norm2(x), context, context) + x
        x = self.ff(self.norm3(x)) + x
        return x


class SpatialTransformer(nn.Module):
    def __init__(self, in_channels, key_dim, value_dim, n_heads, d_head, depth=1, fuser_type=None, use_checkpoint=True):
        super().__init__()
        self.in_channels = in_channels
        query_dim = n_heads * d_head
        self.norm = Normalize(in_channels)


        self.proj_in = nn.Conv2d(in_channels,
                                 query_dim,
                                 kernel_size=1,
                                 stride=1,
                                 padding=0)

        self.transformer_blocks = nn.ModuleList(
            [BasicTransformerBlock(query_dim, key_dim, value_dim, n_heads, d_head, fuser_type, use_checkpoint=use_checkpoint)
                for d in range(depth)]
        )

        self.proj_out = zero_module(nn.Conv2d(query_dim,
                                              in_channels,
                                              kernel_size=1,
                                              stride=1,
                                              padding=0))

    def forward(self, x, context, objs):
        b, c, h, w = x.shape
        x_in = x
        x = self.norm(x)
        x = self.proj_in(x)
        x = rearrange(x, 'b c h w -> b (h w) c')
        for block in self.transformer_blocks:
            x = block(x, context, objs)
        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w)
        x = self.proj_out(x)
        return x + x_in