-
Notifications
You must be signed in to change notification settings - Fork 10
/
masking.py
185 lines (158 loc) · 7.53 KB
/
masking.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
"""Masking module: random input masking, table masking, etc."""
import numpy as np
import torch
import torch.nn as nn
import common
# TODO: refactor made.py to use this class.
class Masking(object):
"""Column masking logic."""
@classmethod
def Params(cls):
p = {}
p['draw_dropout_per_col'] = False
p['per_row_dropout'] = False
return p
def __init__(self, params):
for k, v in params.items():
setattr(self, k, v)
# Optimization.
self._constant_ones_cache = {}
def input_mask(self, x, is_training):
"""Calculates a random input mask for training.
Args:
x: input tokens, shaped [batch_size, num_cols].
is_training: bool.
Returns:
batch_mask, bools, shaped [batch_size, num_cols, 1], where 1 means
use original representation and 0 means use MASK representaions.
During inference this should be all 1s.
"""
assert x.ndim == 2, x.shape
if not is_training:
# During inference, short-circuit to immediately return 1s.
return torch.ones_like(x).unsqueeze(2)
if self.table_dropout:
return self._table_dropout(x, is_training)
return self._vanilla_dropout(x, is_training)
def _get_cached_constant_ones(self, shape, device):
"""Returns a cached all-one tensor with desired shape and device."""
key = (shape, device)
if key not in self._constant_ones_cache:
self._constant_ones_cache[key] = torch.ones(*shape, device=device)
return self._constant_ones_cache[key]
def _table_dropout(self, x, is_training):
bs, inp_seq_len = x.shape
kOnes = self._get_cached_constant_ones((bs, 1), x.device)
if self.per_row_dropout:
# NOTE: torch.rand* funcs on GPU are ~4% slower than
# generating them on CPU via np.random.
num_dropped_tables = np.random.randint(1, self.num_joined_tables,
(bs, 1)).astype(np.float32,
copy=False)
table_dropped = np.random.rand(bs, self.num_joined_tables) <= (
num_dropped_tables / self.num_joined_tables)
if self.table_primary_index is not None:
table_dropped[:, self.table_primary_index] = False
normal_drop_rands = np.random.rand(bs, inp_seq_len)
table_dropped = table_dropped.astype(np.float32, copy=False)
else:
# 1 means drop that table.
num_dropped_tables = np.random.randint(1, self.num_joined_tables)
table_dropped = np.random.rand(
self.num_joined_tables
) <= num_dropped_tables / self.num_joined_tables
if self.table_primary_index is not None:
table_dropped[self.table_primary_index] = False
table_dropped = table_dropped.astype(np.float32, copy=False)
batch_masks = []
for i in range(inp_seq_len):
# Table dropout. Logic:
# First, draw the tables to be dropped.
# If a table T is dropped:
# Drop its content columns & indicator only.
# Don't drop its fanout.
# Otherwise:
# Uniformly wraw # content columns to drop.
# Don't drop its indicator.
# Drop its fanout.
table_index = self.table_indexes[i]
if self.per_row_dropout:
# table_dropped[table_index]: shaped [BS, 1]
# elem 0 : True
# elem 1 : True
# elem 2 : False, etc.
is_content = float(
self.table_column_types[i] == common.TYPE_NORMAL_ATTR)
is_fanout = float(
self.table_column_types[i] == common.TYPE_FANOUT)
use_unk = table_dropped[:, table_index]
if is_fanout:
# Column i is a fanout column. Drop iff table not dropped.
batch_mask = torch.tensor(use_unk).float().unsqueeze(1).to(
x.device)
else:
# Handle batch elements where this table is not dropped.
normal_drop_prob = np.random.randint(
0, self.table_num_columns[table_index] + 1,
(bs,)) * 1. / self.table_num_columns[table_index]
normal_drop = normal_drop_rands[:, i] <= normal_drop_prob
# Make sure we drop content only.
normal_drop = normal_drop * is_content
not_dropped_pos = (use_unk == 0.0)
use_unk[not_dropped_pos] = normal_drop[not_dropped_pos]
# Shaped [bs, 1].
batch_mask = torch.as_tensor(1.0 - use_unk).unsqueeze(1).to(
x.device)
else:
# Make decisions for entire batch.
if table_dropped[table_index]:
# Drop all its normal attributes + indicator. Don't drop
# fanout.
batch_mask = torch.clamp(
torch.dropout(
kOnes,
p=1.0 -
(self.table_column_types[i] == common.TYPE_FANOUT),
train=is_training), 0, 1)
else:
# Drop each normal attribute with drawn propability.
# Don't drop indicator.
# Drop fanout.
drop_p = 0.0
if self.table_column_types[i] == common.TYPE_NORMAL_ATTR:
# Possible to drop all columns of this
# table (it participates in join but no
# attributes are filtered).
drop_p = np.random.randint(
0, self.table_num_columns[table_index] +
1) / self.table_num_columns[table_index]
elif self.table_column_types[i] == common.TYPE_FANOUT:
drop_p = 1.0
batch_mask = torch.clamp(
torch.dropout(kOnes, p=drop_p, train=is_training), 0, 1)
batch_masks.append(batch_mask)
# [bs, num cols, 1].
return torch.cat(batch_masks, 1).unsqueeze(-1)
def _vanilla_dropout(self, x, is_training):
bs, inp_seq_len = x.shape
if self.draw_dropout_per_col:
kOnes = self._get_cached_constant_ones((bs, 1, 1), x.device)
vecs = []
for _ in range(inp_seq_len):
vecs.append(
torch.dropout(kOnes,
p=np.random.randint(0, inp_seq_len) /
inp_seq_len,
train=is_training))
dropout_vec = torch.cat(vecs, dim=1)
else:
kOnes = self._get_cached_constant_ones((bs, inp_seq_len, 1),
x.device)
dropout_vec = torch.dropout(kOnes,
p=np.random.randint(0, inp_seq_len) /
inp_seq_len,
train=is_training)
# During training, non-dropped 1's are scaled by 1/(1-p), so we
# clamp back to 1. Shaped [bs, num cols, 1].
batch_mask = torch.clamp(dropout_vec, 0, 1)
return batch_mask