forked from facebookresearch/EGG
-
Notifications
You must be signed in to change notification settings - Fork 0
/
features.py
216 lines (180 loc) · 7.41 KB
/
features.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
# Copyright (c) Facebook, Inc. and its affiliates.
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import itertools
import os
import pathlib
from functools import reduce
import numpy as np
import torch
from torch.utils import data
from egg.zoo.objects_game.util import compute_binomial
class VectorsLoader:
def __init__(
self,
perceptual_dimensions=[4, 4, 4, 4, 4],
n_distractors=1,
batch_size=32,
train_samples=128000,
validation_samples=4096,
test_samples=1024,
shuffle_train_data=False,
dump_data_folder=None,
load_data_path=None,
seed=None,
):
self.perceptual_dimensions = perceptual_dimensions
self._n_features = len(self.perceptual_dimensions)
self.n_distractors = n_distractors
self.batch_size = batch_size
self.train_samples = train_samples
self.validation_samples = validation_samples
self.test_samples = test_samples
self.shuffle_train_data = shuffle_train_data
self.load_data_path = load_data_path
self.dump_data_folder = (
pathlib.Path(dump_data_folder) if dump_data_folder is not None else None
)
seed = seed if seed else np.random.randint(0, 2 ** 31)
self.random_state = np.random.RandomState(seed)
@property
def n_features(self):
return self._n_features
@n_features.setter
def n_features(self, n_features):
self._n_features = n_features
def upd_cl_options(self, opts):
opts.perceptual_dimensions = self.perceptual_dimensions
opts.train_samples = self.train_samples
opts.validation_samples = self.validation_samples
opts.test_samples = self.test_samples
opts.n_distractors = self.n_distractors
def load_data(self, data_file):
data = np.load(data_file)
train, train_labels = data["train"], data["train_labels"]
valid, valid_labels = data["valid"], data["valid_labels"]
test, test_labels = data["test"], data["test_labels"]
# train valid and test are of shape b_size X n_distractors+1 X n_features
self.train_samples = train.shape[0]
self.validation_samples = valid.shape[0]
self.test_samples = test.shape[0]
self.n_distractors = train.shape[1] - 1
self.perceptual_dimensions = [-1] * train.shape[-1]
self._n_features = len(self.perceptual_dimensions)
return (train, train_labels), (valid, valid_labels), (test, test_labels)
def _fill_split(self, all_vectors, n_samples, tuple_dict):
split_list = []
len_all_vectors = len(all_vectors)
tuple_dim = self.n_distractors + 1
done = 0
while done < n_samples:
candidates_tuple = self.random_state.choice(
len_all_vectors, replace=False, size=tuple_dim
)
key = ""
for vector_idx in candidates_tuple:
key += f"{str(vector_idx)}-"
key = key[:-1]
if key not in tuple_dict:
tuple_dict[key] = True
possible_batch = all_vectors[candidates_tuple]
split_list.append(possible_batch)
done += 1
else:
continue
target_idxs = self.random_state.choice(self.n_distractors + 1, n_samples)
return (np.array(split_list), target_idxs), tuple_dict
def generate_tuples(self, data):
data = np.array(data)
train_data, tuple_dict = self._fill_split(data, self.train_samples, {})
valid_data, tuple_dict = self._fill_split(
data, self.validation_samples, tuple_dict
)
test_data, _ = self._fill_split(data, self.test_samples, tuple_dict)
return train_data, valid_data, test_data
def collate(self, batch):
tuples, target_idxs = [elem[0] for elem in batch], [elem[1] for elem in batch]
receiver_input = np.reshape(
tuples, (self.batch_size, self.n_distractors + 1, -1)
)
labels = np.array(target_idxs)
targets = receiver_input[np.arange(self.batch_size), labels]
return (
torch.from_numpy(targets).float(),
torch.from_numpy(labels).long(),
torch.from_numpy(receiver_input).float(),
)
def get_iterators(self):
if self.load_data_path:
train, valid, test = self.load_data(self.load_data_path)
else: # if load_data_path wasn't given then I need to generate the tuple
world_dim = reduce(lambda x, y: x * y, self.perceptual_dimensions)
possible_tuples = compute_binomial(world_dim, self.n_distractors + 1)
list_of_dim = [range(1, elem + 1) for elem in self.perceptual_dimensions]
all_vectors = list(itertools.product(*list_of_dim))
assert (
self.train_samples > 0
and self.validation_samples > 0
and self.test_samples > 0
), "Train size, validation size and test size must all be greater than 0"
assert (
possible_tuples
> self.train_samples + self.validation_samples + self.test_samples
), f"Not enough data for requested split sizes. Reduced split samples or increase perceptual_dimensions"
train, valid, test = self.generate_tuples(data=all_vectors)
assert (
self.train_samples >= self.batch_size
and self.validation_samples >= self.batch_size
and self.test_samples >= self.batch_size
), "Batch size cannot be smaller than any split size"
train_dataset = TupleDataset(*train)
valid_dataset = TupleDataset(*valid)
test_dataset = TupleDataset(*test)
train_it = data.DataLoader(
train_dataset,
batch_size=self.batch_size,
collate_fn=self.collate,
drop_last=True,
shuffle=self.shuffle_train_data,
)
validation_it = data.DataLoader(
valid_dataset,
batch_size=self.batch_size,
collate_fn=self.collate,
drop_last=True,
)
test_it = data.DataLoader(
test_dataset,
batch_size=self.batch_size,
collate_fn=self.collate,
drop_last=True,
)
if self.dump_data_folder:
self.dump_data_folder.mkdir(exist_ok=True)
path = (
self.dump_data_folder
/ f"{self.perceptual_dimensions}_{self.n_distractors}_distractors"
)
np.savez_compressed(
path,
train=train[0],
train_labels=train[1],
valid=valid[0],
valid_labels=valid[1],
test=test[0],
test_labels=test[1],
n_distractors=self.n_distractors,
)
return train_it, validation_it, test_it
class TupleDataset(data.Dataset):
def __init__(self, tuples, target_idxs):
self.list_of_tuples = tuples
self.target_idxs = target_idxs
def __len__(self):
return len(self.list_of_tuples)
def __getitem__(self, idx):
if idx < 0 or idx >= len(self.list_of_tuples):
raise RuntimeError(
"Accessing dataset through wrong index: < 0 or >= max_len"
)
return self.list_of_tuples[idx], self.target_idxs[idx]