-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathinput_embedders.py
223 lines (196 loc) · 10.9 KB
/
input_embedders.py
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
import torch
import torch.nn as nn
import numpy as np
from datasets.gauss_datasets import get_mus_label_class
from models import get_sinusoidal_positional_embeddings_2
import h5py as h5
import os
from definitions import ROOT_FOLDER
class GaussianEmbedder(nn.Module):
"""A simple Gaussian embedder for the input space of the transformer model.
Note: this inherits from nn.Module, but it doesn't have any trainable parameters.
"""
def __init__(self, config):
super(GaussianEmbedder, self).__init__()
self.config = config
self.e_fac = 1 / np.sqrt(1 + config.data.eps ** 2)
self.Nmax = config.data.Nmax
self.N = config.seq.N
self.D = config.data.D
self.L = config.data.L
self.S = config.train.batch_size
self.mus_label, self.mus_class, self.labels_class = self.get_mus_label_class(config.data.K, config.data.L, config.data.D)
def get_mus_label_class(self, K, L, D):
n_possible_labels = K
mus_label = torch.normal(0, 1, size=(n_possible_labels, D)) / np.sqrt(D)
mus_class = torch.normal(0, 1, size=(K, D)) / np.sqrt(D)
if K < L or K % L != 0:
raise ValueError("K >= L and K%L == 0 is required")
labels_class = torch.tile(torch.arange(L), (1, int(K / L))).squeeze().int()
return mus_label, mus_class, labels_class
def forward(self, batch):
examples = batch['example']
labels = batch['label']
inputs = torch.zeros((self.config.train.batch_size, 2 * self.N + 1, 2 * self.Nmax + 1 + self.config.data.D))
# fill even indices with class examples
inputs[:, :-1:2, 2 * self.Nmax + 1:] = \
(self.e_fac * (self.mus_class[examples[:, :-1]] + self.config.data.eps * torch.Tensor(
np.random.normal(size=(self.S, self.N, self.D))).double() / np.sqrt(self.D)))
# fill odd indices with label examples
inputs[:, 1:-1:2, 2 * self.Nmax + 1:] = self.mus_label[labels[:, :-1]]
# fill last index with target examples
inputs[:, -1, 2 * self.Nmax + 1:] = \
(self.e_fac * (self.mus_class[examples[:, -1]] + self.config.data.eps * np.random.normal(size=(self.S, self.D)) / np.sqrt(self.D)))
# add positional encoding (random shifts)
shifts = np.random.choice((2 * self.Nmax + 1) - (2 * self.N + 1) + 1, size=(self.S))
for s in range(self.S):
inputs[s, :, shifts[s]:shifts[s] + 2 * self.N + 1] = torch.Tensor(np.identity(2 * self.N + 1))
return inputs
class GaussianEmbedderForOrdering(nn.Module):
"""A simple Gaussian embedder for the input space of the transformer model.
Note: this inherits from nn.Module, but it doesn't have any trainable parameters.
"""
def __init__(self, config, device):
super(GaussianEmbedderForOrdering, self).__init__()
self.config = config
self.e_fac = 1 / np.sqrt(1 + config.data.eps ** 2)
self.Nmax = config.data.Nmax
self.N = config.seq.N
self.D = config.data.D
self.L = config.data.L
self.S = config.train.batch_size
self.pos_embedding_type = config.model.pos_emb_type
self.device = device
if config.model.pos_emb_type == 'sinusoidal':
self.positional_embedding = get_sinusoidal_positional_embeddings_2(config.model.max_T * 3, self.D).to(device)
self.mus_label, self.mus_class, self.labels_class = self.get_mus_label_class(config.data.K, config.data.L, config.data.D)
self.mus_label = self.mus_label.to(device)
self.mus_class = self.mus_class.to(device)
self.labels_class = self.labels_class.to(device)
def get_mus_label_class(self, K, L, D):
n_possible_labels = K
mus_label = torch.normal(0, 1, size=(n_possible_labels, D)) / np.sqrt(D)
mus_class = torch.normal(0, 1, size=(K, D)) / np.sqrt(D)
if K < L or K % L != 0:
raise ValueError("K >= L and K%L == 0 is required")
labels_class = torch.tile(torch.arange(L), (1, int(K / L))).squeeze().int()
return mus_label, mus_class, labels_class
def forward(self, batch):
examples = batch['example']
labels = batch['label']
if self.config.seq.include_flipped:
n_flips = 2
else:
n_flips = 1
seq_len = (self.N-1) * 3 * n_flips + 2 # N-1 combos, 3 items per combo, 2 orderings, 2 targets
n_example_pairs = (self.N-1) * n_flips + 1 # this includes the target
inputs = torch.zeros((self.config.train.batch_size, seq_len, 2 * self.Nmax + self.config.data.D))
# fill every first 2 indices with class examples
inputs[:, ::3, 2 * self.Nmax:] = \
(self.e_fac * (self.mus_class[examples[:, ::2]] + self.config.data.eps * torch.Tensor(
np.random.normal(size=(self.S, n_example_pairs, self.D))).double().to(self.device) / np.sqrt(self.D)))
inputs[:, 1::3, 2 * self.Nmax:] = \
(self.e_fac * (self.mus_class[examples[:, 1::2]] + self.config.data.eps * torch.Tensor(
np.random.normal(size=(self.S, n_example_pairs, self.D))).double().to(self.device) / np.sqrt(self.D)))
# fill every 3rd index with label examples
inputs[:, 2:-2:3, 2 * self.Nmax:] = self.mus_label[labels[:, :-1]]
# add positional encoding (random shifts)
if self.config.model.add_pos_encodings:
shifts = np.random.choice((2 * self.Nmax + 2) - (2 * self.N + 2) + 1, size=(self.S))
for s in range(self.S):
if self.config.model.pos_emb_randomization == 'per_batch':
shift_choice = 0
write_to_example = s
elif self.config.model.pos_emb_randomization == 'per_sequence':
shift_choice = s
write_to_example = s
elif self.config.model.pos_emb_randomization == 'only_first':
shift_choice = 0
write_to_example = 0
elif self.config.model.pos_emb_randomization == 'only_last':
shift_choice = 0
write_to_example = self.S - 1
elif self.config.model.pos_emb_randomization == 'no_shift':
shifts = np.zeros(self.S, dtype=int)
shift_choice = 0
write_to_example = s
else:
raise ValueError('Invalid positional embedding randomization: {}'.format(
self.config.model.pos_emb_randomization))
if self.pos_embedding_type == 'onehot':
inputs[write_to_example, :,
shifts[shift_choice]:shifts[shift_choice] + seq_len] = torch.Tensor(
np.identity(seq_len)).to(self.device)
else:
inputs[write_to_example, :, :2 * self.Nmax] = self.positional_embedding[0,
shifts[shift_choice]: shifts[
shift_choice] + seq_len]
return inputs.to(self.device)
class OmniglotEmbedder(nn.Module):
"""A simple Omniglot embedder for the input space of the transformer model.
Note: this inherits from nn.Module, but it doesn't have any trainable parameters.
"""
def __init__(self, config, device):
super(OmniglotEmbedder, self).__init__()
self.config = config
# embedidngs are stored in a file. this class simply reads them
embeddings_file = os.path.join(ROOT_FOLDER, 'datasets/omniglot_resnet18_randomized_order_s0.h5')
with h5.File(embeddings_file, 'r') as f:
embeddings = torch.Tensor(np.array(f['resnet18/224/feat']))
# we only take the first exemplar from each class, for now
embeddings = embeddings[:, 0, :]
self.pos_embedding_type = config.model.pos_emb_type
num_classes, emb_dim = embeddings.shape
assert emb_dim == config.data.D
self.embeddings = embeddings.to(device)
self.D = config.data.D # dimension of the embeddings (512 in the omniglot case)
self.Nmax = config.data.Nmax
self.N = config.seq.N
self.L = config.data.L
self.label_embeddings = (torch.normal(0, 1, size=(self.L, self.D)) / np.sqrt(self.D)).to(device)
self.S = config.train.batch_size
self.device = device
def forward(self, batch):
examples = batch['example']
labels = batch['label']
seq_len = examples.shape[1] + labels.shape[1] - 1
input_dim = self.D + 2 * self.Nmax if self.config.model.add_pos_encodings else self.D
example_idx = 2 * self.Nmax if self.config.model.add_pos_encodings else 0
inputs = torch.zeros((self.config.train.batch_size, seq_len, input_dim)).to(self.device)
# fill every first 2 indices with class examples
inputs[:, ::3, example_idx:] = self.embeddings[examples[:, ::2]]
inputs[:, 1::3, example_idx:] = self.embeddings[examples[:, 1::2]]
# fill every 3rd index with label examples (except the last one)
inputs[:, 2::3, example_idx:] = self.label_embeddings[labels[:, :-1]]
# add positional encoding (random shifts)
if self.config.model.add_pos_encodings:
shifts = np.random.choice((2 * self.Nmax + 2) - (2 * self.N + 2) + 1, size=(self.S))
for s in range(self.S):
if self.config.model.pos_emb_randomization == 'per_batch':
shift_choice = 0
write_to_example = s
elif self.config.model.pos_emb_randomization == 'per_sequence':
shift_choice = s
write_to_example = s
elif self.config.model.pos_emb_randomization == 'only_first':
shift_choice = 0
write_to_example = 0
elif self.config.model.pos_emb_randomization == 'only_last':
shift_choice = 0
write_to_example = self.S - 1
elif self.config.model.pos_emb_randomization == 'no_shift':
shifts = np.zeros(self.S, dtype=int)
shift_choice = 0
write_to_example = s
else:
raise ValueError('Invalid positional embedding randomization: {}'.format(
self.config.model.pos_emb_randomization))
if self.pos_embedding_type == 'onehot':
inputs[write_to_example, :,
shifts[shift_choice]:shifts[shift_choice] + seq_len] = torch.Tensor(
np.identity(seq_len)).to(self.device)
else:
inputs[write_to_example, :, :2 * self.Nmax] = self.positional_embedding[0,
shifts[shift_choice]: shifts[
shift_choice] + seq_len]
return inputs.to(self.device)