-
Notifications
You must be signed in to change notification settings - Fork 237
Expand file tree
/
Copy pathtensorflow_backend.py
More file actions
303 lines (249 loc) · 11.7 KB
/
tensorflow_backend.py
File metadata and controls
303 lines (249 loc) · 11.7 KB
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
import typing
from typing import TYPE_CHECKING, Callable, List, Optional, Tuple
import numpy as np
from docarray.computation import AbstractComputationalBackend
from docarray.computation.abstract_numpy_based_backend import AbstractNumpyBasedBackend
from docarray.typing import TensorFlowTensor
from docarray.utils._internal.misc import import_library
if TYPE_CHECKING:
import tensorflow as tf # type: ignore
import tensorflow._api.v2.experimental.numpy as tnp # type: ignore
else:
tf = import_library('tensorflow', raise_error=True)
tnp = tf._api.v2.experimental.numpy
def _unsqueeze_if_single_axis(*matrices: tf.Tensor) -> List[tf.Tensor]:
"""
Unsqueezes tensors that only have one axis, at dim 0.
This ensures that all outputs can be treated as matrices, not vectors.
:param matrices: Matrices to be unsqueezed
:return: List of the input matrices,
where single axis matrices are unsqueezed at dim 0.
"""
unsqueezed = []
for m in matrices:
if len(m.shape) == 1:
unsqueezed.append(tf.expand_dims(m, axis=0))
else:
unsqueezed.append(m)
return unsqueezed
def _unsqueeze_if_scalar(t: tf.Tensor) -> tf.Tensor:
"""
Unsqueezes tensor of a scalar, from shape () to shape (1,).
:param t: tensor to unsqueeze.
:return: unsqueezed tf.Tensor
"""
if len(t.shape) == 0: # avoid scalar output
t = tf.expand_dims(t, 0)
return t
def norm_left(t: tf.Tensor) -> TensorFlowTensor:
return TensorFlowTensor(tensor=t)
def norm_right(t: TensorFlowTensor) -> tf.Tensor:
return t.tensor
class TensorFlowCompBackend(AbstractNumpyBasedBackend[TensorFlowTensor]):
"""
Computational backend for TensorFlow.
"""
_module = tnp
_cast_output: Callable = norm_left
_get_tensor: Callable = norm_right
@classmethod
def to_numpy(cls, array: 'TensorFlowTensor') -> 'np.ndarray':
return cls._get_tensor(array).numpy()
@classmethod
def none_value(cls) -> typing.Any:
"""Provide a compatible value that represents None in numpy."""
return tf.constant(float('nan'))
@classmethod
def to_device(cls, tensor: 'TensorFlowTensor', device: str) -> 'TensorFlowTensor':
"""Move the tensor to the specified device."""
if cls.device(tensor) == device:
return tensor
else:
with tf.device(device):
return cls._cast_output(tf.identity(cls._get_tensor(tensor)))
@classmethod
def device(cls, tensor: 'TensorFlowTensor') -> Optional[str]:
"""Return device on which the tensor is allocated."""
return cls._get_tensor(tensor).device
@classmethod
def detach(cls, tensor: 'TensorFlowTensor') -> 'TensorFlowTensor':
"""
Returns the tensor detached from its current graph.
:param tensor: tensor to be detached
:return: a detached tensor with the same data.
"""
return cls._cast_output(tf.stop_gradient(cls._get_tensor(tensor)))
@classmethod
def dtype(cls, tensor: 'TensorFlowTensor') -> tf.dtypes:
"""Get the data type of the tensor."""
d_type = cls._get_tensor(tensor).dtype
return d_type.name
@classmethod
def minmax_normalize(
cls,
tensor: 'TensorFlowTensor',
t_range: Tuple = (0.0, 1.0),
x_range: Optional[Tuple] = None,
eps: float = 1e-7,
) -> 'TensorFlowTensor':
a, b = t_range
t = tf.cast(cls._get_tensor(tensor), tf.float32)
min_d = x_range[0] if x_range else tnp.min(t, axis=-1, keepdims=True)
max_d = x_range[1] if x_range else tnp.max(t, axis=-1, keepdims=True)
i = (b - a) * (t - min_d) / (max_d - min_d + tf.constant(eps) + a)
normalized = tnp.clip(i, *((a, b) if a < b else (b, a)))
return cls._cast_output(tf.cast(normalized, tensor.tensor.dtype))
@classmethod
def equal(cls, tensor1: 'TensorFlowTensor', tensor2: 'TensorFlowTensor') -> bool:
"""
Check if two tensors are equal.
:param tensor1: the first tensor
:param tensor2: the second tensor
:return: True if two tensors are equal, False otherwise.
If one or more of the inputs is not a TensorFlowTensor, return False.
"""
t1, t2 = getattr(tensor1, 'tensor', None), getattr(tensor2, 'tensor', None)
if tf.is_tensor(t1) and tf.is_tensor(t2):
# mypy doesn't know that tf.is_tensor implies that t1, t2 are not None
return t1.shape == t2.shape and tf.math.reduce_all(tf.equal(t1, t1)) # type: ignore
return False
class Retrieval(AbstractComputationalBackend.Retrieval[TensorFlowTensor]):
"""
Abstract class for retrieval and ranking functionalities
"""
@staticmethod
def top_k(
values: 'TensorFlowTensor',
k: int,
descending: bool = False,
device: Optional[str] = None,
) -> Tuple['TensorFlowTensor', 'TensorFlowTensor']:
"""
Retrieves the top k smallest values in `values`,
and returns them alongside their indices in the input `values`.
Can also be used to retrieve the top k largest values,
by setting the `descending` flag.
:param values: TensorFlowTensor of values to rank.
Should be of shape (n_queries, n_values_per_query).
Inputs of shape (n_values_per_query,) will be expanded
to (1, n_values_per_query).
:param k: number of values to retrieve
:param descending: retrieve largest values instead of smallest values
:param device: the computational device to use.
:return: Tuple of TensorFlowTensors containing the retrieved values, and
their indices. Both are of shape (n_queries, k)
"""
comp_be = TensorFlowCompBackend
if device is not None:
values = comp_be.to_device(values, device)
tf_values: tf.Tensor = comp_be._get_tensor(values)
if len(tf_values.shape) <= 1:
tf_values = tf.expand_dims(tf_values, axis=0)
len_tf_values = (
tf_values.shape[-1] if len(tf_values.shape) > 1 else len(tf_values)
)
k = min(k, len_tf_values)
if not descending:
tf_values = -tf_values
result = tf.math.top_k(input=tf_values, k=k, sorted=True)
res_values = result.values
res_indices = result.indices
if not descending:
res_values = -result.values
return comp_be._cast_output(res_values), comp_be._cast_output(res_indices)
class Metrics(AbstractComputationalBackend.Metrics[TensorFlowTensor]):
"""
Abstract base class for metrics (distances and similarities).
"""
@staticmethod
def cosine_sim(
x_mat: 'TensorFlowTensor',
y_mat: 'TensorFlowTensor',
eps: float = 1e-7,
device: Optional[str] = None,
) -> 'TensorFlowTensor':
"""Pairwise cosine similarities between all vectors in x_mat and y_mat.
:param x_mat: tensor of shape (n_vectors, n_dim), where n_vectors is the
number of vectors and n_dim is the number of dimensions of each example.
:param y_mat: tensor of shape (n_vectors, n_dim), where n_vectors is the
number of vectors and n_dim is the number of dimensions of each example.
:param eps: a small jitter to avoid divde by zero
:param device: the device to use for computations.
If not provided, the devices of x_mat and y_mat are used.
:return: Tensor of shape (n_vectors, n_vectors) containing all pairwise
cosine distances.
The index [i_x, i_y] contains the cosine distance between
x_mat[i_x] and y_mat[i_y].
"""
comp_be = TensorFlowCompBackend
x_mat_tf: tf.Tensor = comp_be._get_tensor(x_mat)
y_mat_tf: tf.Tensor = comp_be._get_tensor(y_mat)
with tf.device(device):
x_mat_tf = tf.identity(x_mat_tf)
y_mat_tf = tf.identity(y_mat_tf)
x_mat_tf, y_mat_tf = _unsqueeze_if_single_axis(x_mat_tf, y_mat_tf)
a_n = tf.linalg.normalize(x_mat_tf, axis=1)[1]
b_n = tf.linalg.normalize(y_mat_tf, axis=1)[1]
a_norm = x_mat_tf / tf.clip_by_value(
a_n, clip_value_min=eps, clip_value_max=tf.float32.max
)
b_norm = y_mat_tf / tf.clip_by_value(
b_n, clip_value_min=eps, clip_value_max=tf.float32.max
)
sims = tf.squeeze(tf.linalg.matmul(a_norm, tf.transpose(b_norm)))
sims = _unsqueeze_if_scalar(sims)
return comp_be._cast_output(sims)
@staticmethod
def euclidean_dist(
x_mat: 'TensorFlowTensor',
y_mat: 'TensorFlowTensor',
device: Optional[str] = None,
) -> 'TensorFlowTensor':
"""Pairwise Euclidian distances between all vectors in x_mat and y_mat.
:param x_mat: tensor of shape (n_vectors, n_dim), where n_vectors is the
number of vectors and n_dim is the number of dimensions of each example.
:param y_mat: tensor of shape (n_vectors, n_dim), where n_vectors is the
number of vectors and n_dim is the number of dimensions of each example.
:param device: the device to use for pytorch computations.
If not provided, the devices of x_mat and y_mat are used.
:return: Tensor of shape (n_vectors, n_vectors) containing all pairwise
euclidian distances.
The index [i_x, i_y] contains the euclidian distance between
x_mat[i_x] and y_mat[i_y].
"""
comp_be = TensorFlowCompBackend
x_mat_tf: tf.Tensor = comp_be._get_tensor(x_mat)
y_mat_tf: tf.Tensor = comp_be._get_tensor(y_mat)
with tf.device(device):
x_mat_tf = tf.identity(x_mat_tf)
y_mat_tf = tf.identity(y_mat_tf)
x_mat_tf, y_mat_tf = _unsqueeze_if_single_axis(x_mat_tf, y_mat_tf)
dists = tf.squeeze(tf.norm(tf.subtract(x_mat_tf, y_mat_tf), axis=-1))
dists = _unsqueeze_if_scalar(dists)
return comp_be._cast_output(dists)
@staticmethod
def sqeuclidean_dist(
x_mat: 'TensorFlowTensor',
y_mat: 'TensorFlowTensor',
device: Optional[str] = None,
) -> 'TensorFlowTensor':
"""Pairwise Squared Euclidian distances between all vectors
in x_mat and y_mat.
:param x_mat: tensor of shape (n_vectors, n_dim), where n_vectors is the
number of vectors and n_dim is the number of dimensions of each
example.
:param y_mat: tensor of shape (n_vectors, n_dim), where n_vectors is the
number of vectors and n_dim is the number of dimensions of each
example.
:param device: the device to use for pytorch computations.
If not provided, the devices of x_mat and y_mat are used.
:return: Tensor of shape (n_vectors, n_vectors) containing all pairwise
euclidian distances.
The index [i_x, i_y] contains the euclidian distance between
x_mat[i_x] and y_mat[i_y].
"""
dists = TensorFlowCompBackend.Metrics.euclidean_dist(x_mat, y_mat)
squared: tf.Tensor = tf.math.square(
TensorFlowCompBackend._get_tensor(dists)
)
return TensorFlowCompBackend._cast_output(squared)