FilesExpand file tree

main

/

embeddings_flax.py

Copy path

Blame

More file actions

Blame

More file actions

Latest commit

 

History

History

History

95 lines (81 loc) · 3.36 KB

main

/

embeddings_flax.py

Copy path

Top

File metadata and controls

• Code
• Blame

95 lines (81 loc) · 3.36 KB

Raw

Copy raw file

Download raw file

Open symbols panel

Edit and raw actions

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

# Copyright 2023 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 math

import flax.linen as nn

import jax.numpy as jnp

def get_sinusoidal_embeddings(

timesteps: jnp.ndarray,

embedding_dim: int,

freq_shift: float = 1,

min_timescale: float = 1,

max_timescale: float = 1.0e4,

flip_sin_to_cos: bool = False,

scale: float = 1.0,

) -> jnp.ndarray:

"""Returns the positional encoding (same as Tensor2Tensor).

Args:

timesteps: a 1-D Tensor of N indices, one per batch element.

These may be fractional.

embedding_dim: The number of output channels.

min_timescale: The smallest time unit (should probably be 0.0).

max_timescale: The largest time unit.

Returns:

a Tensor of timing signals [N, num_channels]

"""

assert timesteps.ndim == 1, "Timesteps should be a 1d-array"

assert embedding_dim % 2 == 0, f"Embedding dimension {embedding_dim} should be even"

num_timescales = float(embedding_dim // 2)

log_timescale_increment = math.log(max_timescale / min_timescale) / (num_timescales - freq_shift)

inv_timescales = min_timescale * jnp.exp(jnp.arange(num_timescales, dtype=jnp.float32) * -log_timescale_increment)

emb = jnp.expand_dims(timesteps, 1) * jnp.expand_dims(inv_timescales, 0)

# scale embeddings

scaled_time = scale * emb

if flip_sin_to_cos:

signal = jnp.concatenate([jnp.cos(scaled_time), jnp.sin(scaled_time)], axis=1)

else:

signal = jnp.concatenate([jnp.sin(scaled_time), jnp.cos(scaled_time)], axis=1)

signal = jnp.reshape(signal, [jnp.shape(timesteps)[0], embedding_dim])

return signal

class FlaxTimestepEmbedding(nn.Module):

r"""

Time step Embedding Module. Learns embeddings for input time steps.

Args:

time_embed_dim (`int`, *optional*, defaults to `32`):

Time step embedding dimension

dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):

Parameters `dtype`

"""

time_embed_dim: int = 32

dtype: jnp.dtype = jnp.float32

@nn.compact

def __call__(self, temb):

temb = nn.Dense(self.time_embed_dim, dtype=self.dtype, name="linear_1")(temb)

temb = nn.silu(temb)

temb = nn.Dense(self.time_embed_dim, dtype=self.dtype, name="linear_2")(temb)

return temb

class FlaxTimesteps(nn.Module):

r"""

Wrapper Module for sinusoidal Time step Embeddings as described in https://arxiv.org/abs/2006.11239

Args:

dim (`int`, *optional*, defaults to `32`):

Time step embedding dimension

"""

dim: int = 32

flip_sin_to_cos: bool = False

freq_shift: float = 1

@nn.compact

def __call__(self, timesteps):

return get_sinusoidal_embeddings(

timesteps, embedding_dim=self.dim, flip_sin_to_cos=self.flip_sin_to_cos, freq_shift=self.freq_shift

)