attention.py 13 KB
Newer Older
1
import math
Kashif Rasul's avatar
Kashif Rasul committed
2
from typing import Optional
3
4

import torch
Patrick von Platen's avatar
Patrick von Platen committed
5
import torch.nn.functional as F
6
7
8
from torch import nn


9
class AttentionBlock(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
10
11
12
13
    """
    An attention block that allows spatial positions to attend to each other. Originally ported from here, but adapted
    to the N-d case.
    https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
Kashif Rasul's avatar
Kashif Rasul committed
14
15
16
17
18
19
20
21
22
    Uses three q, k, v linear layers to compute attention.

    Parameters:
        channels (:obj:`int`): The number of channels in the input and output.
        num_head_channels (:obj:`int`, *optional*):
            The number of channels in each head. If None, then `num_heads` = 1.
        num_groups (:obj:`int`, *optional*, defaults to 32): The number of groups to use for group norm.
        rescale_output_factor (:obj:`float`, *optional*, defaults to 1.0): The factor to rescale the output by.
        eps (:obj:`float`, *optional*, defaults to 1e-5): The epsilon value to use for group norm.
Patrick von Platen's avatar
Patrick von Platen committed
23
24
25
26
    """

    def __init__(
        self,
Kashif Rasul's avatar
Kashif Rasul committed
27
28
29
30
31
        channels: int,
        num_head_channels: Optional[int] = None,
        num_groups: int = 32,
        rescale_output_factor: float = 1.0,
        eps: float = 1e-5,
Patrick von Platen's avatar
Patrick von Platen committed
32
33
34
35
    ):
        super().__init__()
        self.channels = channels

Patrick von Platen's avatar
Patrick von Platen committed
36
        self.num_heads = channels // num_head_channels if num_head_channels is not None else 1
Patrick von Platen's avatar
Patrick von Platen committed
37
38
39
40
41
42
43
44
45
        self.num_head_size = num_head_channels
        self.group_norm = nn.GroupNorm(num_channels=channels, num_groups=num_groups, eps=eps, affine=True)

        # define q,k,v as linear layers
        self.query = nn.Linear(channels, channels)
        self.key = nn.Linear(channels, channels)
        self.value = nn.Linear(channels, channels)

        self.rescale_output_factor = rescale_output_factor
Patrick von Platen's avatar
Patrick von Platen committed
46
        self.proj_attn = nn.Linear(channels, channels, 1)
Patrick von Platen's avatar
Patrick von Platen committed
47
48

    def transpose_for_scores(self, projection: torch.Tensor) -> torch.Tensor:
49
        new_projection_shape = projection.size()[:-1] + (self.num_heads, -1)
Patrick von Platen's avatar
Patrick von Platen committed
50
51
52
53
54
55
56
57
58
59
        # move heads to 2nd position (B, T, H * D) -> (B, T, H, D) -> (B, H, T, D)
        new_projection = projection.view(new_projection_shape).permute(0, 2, 1, 3)
        return new_projection

    def forward(self, hidden_states):
        residual = hidden_states
        batch, channel, height, width = hidden_states.shape

        # norm
        hidden_states = self.group_norm(hidden_states)
60

Patrick von Platen's avatar
Patrick von Platen committed
61
62
63
64
65
66
67
68
69
70
71
72
73
        hidden_states = hidden_states.view(batch, channel, height * width).transpose(1, 2)

        # proj to q, k, v
        query_proj = self.query(hidden_states)
        key_proj = self.key(hidden_states)
        value_proj = self.value(hidden_states)

        # transpose
        query_states = self.transpose_for_scores(query_proj)
        key_states = self.transpose_for_scores(key_proj)
        value_states = self.transpose_for_scores(value_proj)

        # get scores
74
        scale = 1 / math.sqrt(math.sqrt(self.channels / self.num_heads))
75

76
77
        attention_scores = torch.matmul(query_states * scale, key_states.transpose(-1, -2) * scale)
        attention_probs = torch.softmax(attention_scores.float(), dim=-1).type(attention_scores.dtype)
Patrick von Platen's avatar
Patrick von Platen committed
78
79

        # compute attention output
80
        hidden_states = torch.matmul(attention_probs, value_states)
Patrick von Platen's avatar
Patrick von Platen committed
81

82
83
84
        hidden_states = hidden_states.permute(0, 2, 1, 3).contiguous()
        new_hidden_states_shape = hidden_states.size()[:-2] + (self.channels,)
        hidden_states = hidden_states.view(new_hidden_states_shape)
Patrick von Platen's avatar
Patrick von Platen committed
85
86

        # compute next hidden_states
87
        hidden_states = self.proj_attn(hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
88
89
90
91
92
93
        hidden_states = hidden_states.transpose(-1, -2).reshape(batch, channel, height, width)

        # res connect and rescale
        hidden_states = (hidden_states + residual) / self.rescale_output_factor
        return hidden_states

Patrick von Platen's avatar
Patrick von Platen committed
94

Patrick von Platen's avatar
Patrick von Platen committed
95
96
97
class SpatialTransformer(nn.Module):
    """
    Transformer block for image-like data. First, project the input (aka embedding) and reshape to b, t, d. Then apply
Kashif Rasul's avatar
Kashif Rasul committed
98
99
100
101
102
103
104
105
106
    standard transformer action. Finally, reshape to image.

    Parameters:
        in_channels (:obj:`int`): The number of channels in the input and output.
        n_heads (:obj:`int`): The number of heads to use for multi-head attention.
        d_head (:obj:`int`): The number of channels in each head.
        depth (:obj:`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
        dropout (:obj:`float`, *optional*, defaults to 0.1): The dropout probability to use.
        context_dim (:obj:`int`, *optional*): The number of context dimensions to use.
Patrick von Platen's avatar
Patrick von Platen committed
107
108
    """

Kashif Rasul's avatar
Kashif Rasul committed
109
110
111
112
113
114
115
116
117
    def __init__(
        self,
        in_channels: int,
        n_heads: int,
        d_head: int,
        depth: int = 1,
        dropout: float = 0.0,
        context_dim: Optional[int] = None,
    ):
Patrick von Platen's avatar
Patrick von Platen committed
118
        super().__init__()
Patrick von Platen's avatar
Patrick von Platen committed
119
120
        self.n_heads = n_heads
        self.d_head = d_head
Patrick von Platen's avatar
Patrick von Platen committed
121
122
123
124
125
126
127
128
129
130
131
132
133
        self.in_channels = in_channels
        inner_dim = n_heads * d_head
        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)

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

        self.transformer_blocks = nn.ModuleList(
            [
                BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim)
                for d in range(depth)
            ]
        )

Patrick von Platen's avatar
Patrick von Platen committed
134
        self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
Patrick von Platen's avatar
Patrick von Platen committed
135

136
137
138
139
    def _set_attention_slice(self, slice_size):
        for block in self.transformer_blocks:
            block._set_attention_slice(slice_size)

Patrick von Platen's avatar
Patrick von Platen committed
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
    def forward(self, x, context=None):
        # note: if no context is given, cross-attention defaults to self-attention
        b, c, h, w = x.shape
        x_in = x
        x = self.norm(x)
        x = self.proj_in(x)
        x = x.permute(0, 2, 3, 1).reshape(b, h * w, c)
        for block in self.transformer_blocks:
            x = block(x, context=context)
        x = x.reshape(b, h, w, c).permute(0, 3, 1, 2)
        x = self.proj_out(x)
        return x + x_in


class BasicTransformerBlock(nn.Module):
Kashif Rasul's avatar
Kashif Rasul committed
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
    r"""
    A basic Transformer block.

    Parameters:
        dim (:obj:`int`): The number of channels in the input and output.
        n_heads (:obj:`int`): The number of heads to use for multi-head attention.
        d_head (:obj:`int`): The number of channels in each head.
        dropout (:obj:`float`, *optional*, defaults to 0.0): The dropout probability to use.
        context_dim (:obj:`int`, *optional*): The size of the context vector for cross attention.
        gated_ff (:obj:`bool`, *optional*, defaults to :obj:`False`): Whether to use a gated feed-forward network.
        checkpoint (:obj:`bool`, *optional*, defaults to :obj:`False`): Whether to use checkpointing.
    """

    def __init__(
        self,
        dim: int,
        n_heads: int,
        d_head: int,
        dropout=0.0,
        context_dim: Optional[int] = None,
        gated_ff: bool = True,
        checkpoint: bool = True,
    ):
Patrick von Platen's avatar
Patrick von Platen committed
178
179
180
181
182
183
184
185
186
187
188
189
190
        super().__init__()
        self.attn1 = CrossAttention(
            query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout
        )  # is a self-attention
        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
        self.attn2 = CrossAttention(
            query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, dropout=dropout
        )  # is self-attn if context is none
        self.norm1 = nn.LayerNorm(dim)
        self.norm2 = nn.LayerNorm(dim)
        self.norm3 = nn.LayerNorm(dim)
        self.checkpoint = checkpoint

191
192
193
194
    def _set_attention_slice(self, slice_size):
        self.attn1._slice_size = slice_size
        self.attn2._slice_size = slice_size

Patrick von Platen's avatar
Patrick von Platen committed
195
    def forward(self, x, context=None):
196
        x = x.contiguous() if x.device.type == "mps" else x
Patrick von Platen's avatar
Patrick von Platen committed
197
198
199
200
201
202
203
        x = self.attn1(self.norm1(x)) + x
        x = self.attn2(self.norm2(x), context=context) + x
        x = self.ff(self.norm3(x)) + x
        return x


class CrossAttention(nn.Module):
Kashif Rasul's avatar
Kashif Rasul committed
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
    r"""
    A cross attention layer.

    Parameters:
        query_dim (:obj:`int`): The number of channels in the query.
        context_dim (:obj:`int`, *optional*):
            The number of channels in the context. If not given, defaults to `query_dim`.
        heads (:obj:`int`,  *optional*, defaults to 8): The number of heads to use for multi-head attention.
        dim_head (:obj:`int`,  *optional*, defaults to 64): The number of channels in each head.
        dropout (:obj:`float`, *optional*, defaults to 0.0): The dropout probability to use.
    """

    def __init__(
        self, query_dim: int, context_dim: Optional[int] = None, heads: int = 8, dim_head: int = 64, dropout: int = 0.0
    ):
Patrick von Platen's avatar
Patrick von Platen committed
219
220
        super().__init__()
        inner_dim = dim_head * heads
221
        context_dim = context_dim if context_dim is not None else query_dim
Patrick von Platen's avatar
Patrick von Platen committed
222
223
224

        self.scale = dim_head**-0.5
        self.heads = heads
225
226
227
228
        # for slice_size > 0 the attention score computation
        # is split across the batch axis to save memory
        # You can set slice_size with `set_attention_slice`
        self._slice_size = None
Patrick von Platen's avatar
Patrick von Platen committed
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

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

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

    def reshape_heads_to_batch_dim(self, tensor):
        batch_size, seq_len, dim = tensor.shape
        head_size = self.heads
        tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size)
        return tensor

    def reshape_batch_dim_to_heads(self, tensor):
        batch_size, seq_len, dim = tensor.shape
        head_size = self.heads
        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
        return tensor

    def forward(self, x, context=None, mask=None):
        batch_size, sequence_length, dim = x.shape

        q = self.to_q(x)
254
        context = context if context is not None else x
Patrick von Platen's avatar
Patrick von Platen committed
255
256
257
258
259
260
261
        k = self.to_k(context)
        v = self.to_v(context)

        q = self.reshape_heads_to_batch_dim(q)
        k = self.reshape_heads_to_batch_dim(k)
        v = self.reshape_heads_to_batch_dim(v)

262
        # TODO(PVP) - mask is currently never used. Remember to re-implement when used
Patrick von Platen's avatar
Patrick von Platen committed
263
264

        # attention, what we cannot get enough of
265
266
267
        hidden_states = self._attention(q, k, v, sequence_length, dim)

        return self.to_out(hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
268

269
270
271
272
273
274
275
276
277
    def _attention(self, query, key, value, sequence_length, dim):
        batch_size_attention = query.shape[0]
        hidden_states = torch.zeros(
            (batch_size_attention, sequence_length, dim // self.heads), device=query.device, dtype=query.dtype
        )
        slice_size = self._slice_size if self._slice_size is not None else hidden_states.shape[0]
        for i in range(hidden_states.shape[0] // slice_size):
            start_idx = i * slice_size
            end_idx = (i + 1) * slice_size
278
            attn_slice = torch.matmul(query[start_idx:end_idx], key[start_idx:end_idx].transpose(1, 2)) * self.scale
279
            attn_slice = attn_slice.softmax(dim=-1)
280
            attn_slice = torch.matmul(attn_slice, value[start_idx:end_idx])
281
282
283
284
285
286

            hidden_states[start_idx:end_idx] = attn_slice

        # reshape hidden_states
        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
        return hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
287
288
289


class FeedForward(nn.Module):
Kashif Rasul's avatar
Kashif Rasul committed
290
291
292
293
294
295
296
297
298
299
300
301
302
303
    r"""
    A feed-forward layer.

    Parameters:
        dim (:obj:`int`): The number of channels in the input.
        dim_out (:obj:`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
        mult (:obj:`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
        glu (:obj:`bool`, *optional*, defaults to :obj:`False`): Whether to use GLU activation.
        dropout (:obj:`float`, *optional*, defaults to 0.0): The dropout probability to use.
    """

    def __init__(
        self, dim: int, dim_out: Optional[int] = None, mult: int = 4, glu: bool = False, dropout: float = 0.0
    ):
Patrick von Platen's avatar
Patrick von Platen committed
304
305
        super().__init__()
        inner_dim = int(dim * mult)
306
307
        dim_out = dim_out if dim_out is not None else dim
        project_in = GEGLU(dim, inner_dim)
Patrick von Platen's avatar
Patrick von Platen committed
308
309
310
311
312

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

    def forward(self, x):
        return self.net(x)
Patrick von Platen's avatar
Patrick von Platen committed
313

Patrick von Platen's avatar
Patrick von Platen committed
314

Patrick von Platen's avatar
Patrick von Platen committed
315
316
# feedforward
class GEGLU(nn.Module):
Kashif Rasul's avatar
Kashif Rasul committed
317
318
319
320
321
322
323
324
325
    r"""
    A variant of the gated linear unit activation function from https://arxiv.org/abs/2002.05202.

    Parameters:
        dim_in (:obj:`int`): The number of channels in the input.
        dim_out (:obj:`int`): The number of channels in the output.
    """

    def __init__(self, dim_in: int, dim_out: int):
Patrick von Platen's avatar
Patrick von Platen committed
326
327
328
329
330
331
        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)