File size: 11,223 Bytes
522606a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright 2022 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import flax.linen as nn
import jax.numpy as jnp


class FlaxAttentionBlock(nn.Module):
    r"""
    A Flax multi-head attention module as described in: https://arxiv.org/abs/1706.03762

    Parameters:
        query_dim (:obj:`int`):
            Input hidden states dimension
        heads (:obj:`int`, *optional*, defaults to 8):
            Number of heads
        dim_head (:obj:`int`, *optional*, defaults to 64):
            Hidden states dimension inside each head
        dropout (:obj:`float`, *optional*, defaults to 0.0):
            Dropout rate
        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
            Parameters `dtype`

    """
    query_dim: int
    heads: int = 8
    dim_head: int = 64
    dropout: float = 0.0
    dtype: jnp.dtype = jnp.float32

    def setup(self):
        inner_dim = self.dim_head * self.heads
        self.scale = self.dim_head**-0.5

        # Weights were exported with old names {to_q, to_k, to_v, to_out}
        self.query = nn.Dense(inner_dim, use_bias=False, dtype=self.dtype, name="to_q")
        self.key = nn.Dense(inner_dim, use_bias=False, dtype=self.dtype, name="to_k")
        self.value = nn.Dense(inner_dim, use_bias=False, dtype=self.dtype, name="to_v")

        self.proj_attn = nn.Dense(self.query_dim, dtype=self.dtype, name="to_out_0")

    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 = jnp.transpose(tensor, (0, 2, 1, 3))
        tensor = tensor.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 = jnp.transpose(tensor, (0, 2, 1, 3))
        tensor = tensor.reshape(batch_size // head_size, seq_len, dim * head_size)
        return tensor

    def __call__(self, hidden_states, context=None, deterministic=True):
        context = hidden_states if context is None else context

        query_proj = self.query(hidden_states)
        key_proj = self.key(context)
        value_proj = self.value(context)

        query_states = self.reshape_heads_to_batch_dim(query_proj)
        key_states = self.reshape_heads_to_batch_dim(key_proj)
        value_states = self.reshape_heads_to_batch_dim(value_proj)

        # compute attentions
        attention_scores = jnp.einsum("b i d, b j d->b i j", query_states, key_states)
        attention_scores = attention_scores * self.scale
        attention_probs = nn.softmax(attention_scores, axis=2)

        # attend to values
        hidden_states = jnp.einsum("b i j, b j d -> b i d", attention_probs, value_states)
        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
        hidden_states = self.proj_attn(hidden_states)
        return hidden_states


class FlaxBasicTransformerBlock(nn.Module):
    r"""
    A Flax transformer block layer with `GLU` (Gated Linear Unit) activation function as described in:
    https://arxiv.org/abs/1706.03762


    Parameters:
        dim (:obj:`int`):
            Inner hidden states dimension
        n_heads (:obj:`int`):
            Number of heads
        d_head (:obj:`int`):
            Hidden states dimension inside each head
        dropout (:obj:`float`, *optional*, defaults to 0.0):
            Dropout rate
        only_cross_attention (`bool`, defaults to `False`):
            Whether to only apply cross attention.
        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
            Parameters `dtype`
    """
    dim: int
    n_heads: int
    d_head: int
    dropout: float = 0.0
    only_cross_attention: bool = False
    dtype: jnp.dtype = jnp.float32

    def setup(self):
        # self attention (or cross_attention if only_cross_attention is True)
        self.attn1 = FlaxAttentionBlock(self.dim, self.n_heads, self.d_head, self.dropout, dtype=self.dtype)
        # cross attention
        self.attn2 = FlaxAttentionBlock(self.dim, self.n_heads, self.d_head, self.dropout, dtype=self.dtype)
        self.ff = FlaxGluFeedForward(dim=self.dim, dropout=self.dropout, dtype=self.dtype)
        self.norm1 = nn.LayerNorm(epsilon=1e-5, dtype=self.dtype)
        self.norm2 = nn.LayerNorm(epsilon=1e-5, dtype=self.dtype)
        self.norm3 = nn.LayerNorm(epsilon=1e-5, dtype=self.dtype)

    def __call__(self, hidden_states, context, deterministic=True):
        # self attention
        residual = hidden_states
        if self.only_cross_attention:
            hidden_states = self.attn1(self.norm1(hidden_states), context, deterministic=deterministic)
        else:
            hidden_states = self.attn1(self.norm1(hidden_states), deterministic=deterministic)
        hidden_states = hidden_states + residual

        # cross attention
        residual = hidden_states
        hidden_states = self.attn2(self.norm2(hidden_states), context, deterministic=deterministic)
        hidden_states = hidden_states + residual

        # feed forward
        residual = hidden_states
        hidden_states = self.ff(self.norm3(hidden_states), deterministic=deterministic)
        hidden_states = hidden_states + residual

        return hidden_states


class FlaxTransformer2DModel(nn.Module):
    r"""
    A Spatial Transformer layer with Gated Linear Unit (GLU) activation function as described in:
    https://arxiv.org/pdf/1506.02025.pdf


    Parameters:
        in_channels (:obj:`int`):
            Input number of channels
        n_heads (:obj:`int`):
            Number of heads
        d_head (:obj:`int`):
            Hidden states dimension inside each head
        depth (:obj:`int`, *optional*, defaults to 1):
            Number of transformers block
        dropout (:obj:`float`, *optional*, defaults to 0.0):
            Dropout rate
        use_linear_projection (`bool`, defaults to `False`): tbd
        only_cross_attention (`bool`, defaults to `False`): tbd
        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
            Parameters `dtype`
    """
    in_channels: int
    n_heads: int
    d_head: int
    depth: int = 1
    dropout: float = 0.0
    use_linear_projection: bool = False
    only_cross_attention: bool = False
    dtype: jnp.dtype = jnp.float32

    def setup(self):
        self.norm = nn.GroupNorm(num_groups=32, epsilon=1e-5)

        inner_dim = self.n_heads * self.d_head
        if self.use_linear_projection:
            self.proj_in = nn.Dense(inner_dim, dtype=self.dtype)
        else:
            self.proj_in = nn.Conv(
                inner_dim,
                kernel_size=(1, 1),
                strides=(1, 1),
                padding="VALID",
                dtype=self.dtype,
            )

        self.transformer_blocks = [
            FlaxBasicTransformerBlock(
                inner_dim,
                self.n_heads,
                self.d_head,
                dropout=self.dropout,
                only_cross_attention=self.only_cross_attention,
                dtype=self.dtype,
            )
            for _ in range(self.depth)
        ]

        if self.use_linear_projection:
            self.proj_out = nn.Dense(inner_dim, dtype=self.dtype)
        else:
            self.proj_out = nn.Conv(
                inner_dim,
                kernel_size=(1, 1),
                strides=(1, 1),
                padding="VALID",
                dtype=self.dtype,
            )

    def __call__(self, hidden_states, context, deterministic=True):
        batch, height, width, channels = hidden_states.shape
        residual = hidden_states
        hidden_states = self.norm(hidden_states)
        if self.use_linear_projection:
            hidden_states = hidden_states.reshape(batch, height * width, channels)
            hidden_states = self.proj_in(hidden_states)
        else:
            hidden_states = self.proj_in(hidden_states)
            hidden_states = hidden_states.reshape(batch, height * width, channels)

        for transformer_block in self.transformer_blocks:
            hidden_states = transformer_block(hidden_states, context, deterministic=deterministic)

        if self.use_linear_projection:
            hidden_states = self.proj_out(hidden_states)
            hidden_states = hidden_states.reshape(batch, height, width, channels)
        else:
            hidden_states = hidden_states.reshape(batch, height, width, channels)
            hidden_states = self.proj_out(hidden_states)

        hidden_states = hidden_states + residual
        return hidden_states


class FlaxGluFeedForward(nn.Module):
    r"""
    Flax module that encapsulates two Linear layers separated by a gated linear unit activation from:
    https://arxiv.org/abs/2002.05202

    Parameters:
        dim (:obj:`int`):
            Inner hidden states dimension
        dropout (:obj:`float`, *optional*, defaults to 0.0):
            Dropout rate
        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
            Parameters `dtype`
    """
    dim: int
    dropout: float = 0.0
    dtype: jnp.dtype = jnp.float32

    def setup(self):
        # The second linear layer needs to be called
        # net_2 for now to match the index of the Sequential layer
        self.net_0 = FlaxGEGLU(self.dim, self.dropout, self.dtype)
        self.net_2 = nn.Dense(self.dim, dtype=self.dtype)

    def __call__(self, hidden_states, deterministic=True):
        hidden_states = self.net_0(hidden_states)
        hidden_states = self.net_2(hidden_states)
        return hidden_states


class FlaxGEGLU(nn.Module):
    r"""
    Flax implementation of a Linear layer followed by the variant of the gated linear unit activation function from
    https://arxiv.org/abs/2002.05202.

    Parameters:
        dim (:obj:`int`):
            Input hidden states dimension
        dropout (:obj:`float`, *optional*, defaults to 0.0):
            Dropout rate
        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
            Parameters `dtype`
    """
    dim: int
    dropout: float = 0.0
    dtype: jnp.dtype = jnp.float32

    def setup(self):
        inner_dim = self.dim * 4
        self.proj = nn.Dense(inner_dim * 2, dtype=self.dtype)

    def __call__(self, hidden_states, deterministic=True):
        hidden_states = self.proj(hidden_states)
        hidden_linear, hidden_gelu = jnp.split(hidden_states, 2, axis=2)
        return hidden_linear * nn.gelu(hidden_gelu)