File size: 9,120 Bytes
d758c99 |
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 |
import itertools
import operator
import attr
import numpy as np
import torch
def argsort(items, key=lambda x: x, reverse=False):
# sorted_items: items sorted by key, descending
# orig_to_sort: tuple of integers, satisfies the following:
# tuple(items[i] for i in orig_to_sort) == sorted_items
# items[orig_to_sort[i]] == sorted_items[i]
# sort_to_orig: tuple of integers, satisfies the following:
# tuple(sorted_items[i] for i in sort_to_orig) == items
# sorted_items[sort_to_orig[i]] == items[i]
orig_to_sort, sorted_items = zip(*sorted(
enumerate(items), key=lambda x: key(x[1]), reverse=reverse))
sort_to_orig = tuple(
x[0] for x in sorted(
enumerate(orig_to_sort), key=operator.itemgetter(1)))
return sorted_items, sort_to_orig, orig_to_sort
def sort_lists_by_length(lists):
# Returns the following 3 items:
# - lists_sorted: lists sorted by length of each element, descending
# - orig_to_sort: tuple of integers, satisfies the following:
# tuple(lists[i] for i in orig_to_sort) == lists_sorted
# lists[orig_to_sort[sort_idx]] == lists_sorted[sort_idx]
# - sort_to_orig: list of integers, satisfies the following:
# [lists_sorted[i] for i in sort_to_orig] == lists
# lists_sorted[sort_to_orig[orig_idx]] == lists[orig_idx]
return argsort(lists, key=len, reverse=True)
def batch_bounds_for_packing(lengths):
'''Returns how many items in batch have length >= i at step i.
Examples:
[5] -> [1, 1, 1, 1, 1]
[5, 5] -> [2, 2, 2, 2, 2]
[5, 3] -> [2, 2, 2, 1, 1]
[5, 4, 1, 1] -> [4, 2, 2, 2, 1]
'''
last_length = 0
count = len(lengths)
result = []
for i, (length, group) in enumerate(itertools.groupby(reversed(lengths))):
# TODO: Check that things don't blow up when some lengths are 0
if i > 0 and length <= last_length:
raise ValueError('lengths must be decreasing and positive')
result.extend([count] * (length - last_length))
count -= sum(1 for _ in group)
last_length = length
return result
def _make_packed_sequence(data, batch_sizes):
return torch.nn.utils.rnn.PackedSequence(data,
torch.LongTensor(batch_sizes))
@attr.s(frozen=True)
class PackedSequencePlus:
ps = attr.ib()
lengths = attr.ib()
sort_to_orig = attr.ib(converter=np.array)
orig_to_sort = attr.ib(converter=np.array)
@lengths.validator
def descending(self, attribute, value):
for x, y in zip(value, value[1:]):
if not x >= y:
raise ValueError('Lengths are not descending: {}'.format(value))
def __attrs_post_init__(self):
self.__dict__['cum_batch_sizes'] = np.cumsum([0] + self.ps.batch_sizes[:-1].tolist()).astype(np.int_)
def apply(self, fn):
return attr.evolve(self, ps=torch.nn.utils.rnn.PackedSequence(
fn(self.ps.data), self.ps.batch_sizes))
def with_new_ps(self, ps):
return attr.evolve(self, ps=ps)
def pad(self, batch_first, others_to_unsort=(), padding_value=0.0):
padded, seq_lengths = torch.nn.utils.rnn.pad_packed_sequence(
self.ps, batch_first=batch_first, padding_value=padding_value)
results = padded[
self.sort_to_orig], [seq_lengths[i] for i in self.sort_to_orig]
return results + tuple(t[self.sort_to_orig] for t in others_to_unsort)
def cuda(self):
if self.ps.data.is_cuda:
return self
return self.apply(lambda d: d.cuda())
def raw_index(self, orig_batch_idx, seq_idx):
result = np.take(self.cum_batch_sizes, seq_idx) + np.take(
self.sort_to_orig, orig_batch_idx)
if self.ps.data is not None:
assert np.all(result < len(self.ps.data))
return result
def select(self, orig_batch_idx, seq_idx=None):
if seq_idx is None:
return self.ps.data[
self.raw_index(orig_batch_idx, range(self.lengths[self.sort_to_orig[orig_batch_idx]]))]
return self.ps.data[self.raw_index(orig_batch_idx, seq_idx)]
def select_subseq(self, orig_batch_indices):
lengths = [self.lengths[self.sort_to_orig[i]] for i in
orig_batch_indices]
return self.from_gather(
lengths=lengths,
map_index=self.raw_index,
gather_from_indices=lambda indices:
self.ps.data[torch.LongTensor(indices)])
def orig_index(self, raw_idx):
seq_idx = np.searchsorted(
self.cum_batch_sizes, raw_idx, side='right') - 1
batch_idx = raw_idx - self.cum_batch_sizes[seq_idx]
orig_batch_idx = self.sort_to_orig[batch_idx]
return orig_batch_idx, seq_idx
def orig_batch_indices(self):
result = []
for bs in self.ps.batch_sizes:
result.extend(self.orig_to_sort[:bs])
return np.array(result)
def orig_lengths(self):
for sort_idx in self.sort_to_orig:
yield self.lengths[sort_idx]
def expand(self, k):
# Conceptually, this function does the following:
# Input: d1 x ...
# Output: d1 * k x ... where
# out[0] = out[1] = ... out[k],
# out[k + 0] = out[k + 1] = ... out[k + k],
# and so on.
v = self.ps.data
ps_data = v.unsqueeze(1).repeat(1, k, *(
[1] * (v.dim() - 1))).view(-1, *v.shape[1:])
batch_sizes = (np.array(self.ps.batch_sizes) * k).tolist()
lengths = np.repeat(self.lengths, k).tolist()
sort_to_orig = [
exp_i for i in self.sort_to_orig for exp_i in range(i * k, i * k + k)
]
orig_to_sort = [
exp_i for i in self.orig_to_sort for exp_i in range(i * k, i * k + k)
]
return PackedSequencePlus(
_make_packed_sequence(ps_data, batch_sizes),
lengths, sort_to_orig, orig_to_sort)
@classmethod
def from_lists(cls, lists, item_shape, device, item_to_tensor):
# result = tensor_type(sum(len(lst) for lst in lists), *item_shape)
result_list = []
sorted_lists, sort_to_orig, orig_to_sort = sort_lists_by_length(lists)
lengths = [len(lst) for lst in sorted_lists]
batch_bounds = batch_bounds_for_packing(lengths)
idx = 0
for i, bound in enumerate(batch_bounds):
for batch_idx, lst in enumerate(sorted_lists[:bound]):
# item_to_tensor(lst[i], batch_idx, result[idx])
embed = item_to_tensor(lst[i], batch_idx)
result_list.append(embed)
idx += 1
result = torch.stack(result_list, 0)
return cls(
_make_packed_sequence(result, batch_bounds),
lengths, sort_to_orig, orig_to_sort)
@classmethod
def from_gather(cls, lengths, map_index, gather_from_indices):
sorted_lengths, sort_to_orig, orig_to_sort = argsort(lengths, reverse=True)
batch_bounds = batch_bounds_for_packing(sorted_lengths)
indices = []
for seq_idx, bound in enumerate(batch_bounds):
for batch_idx in orig_to_sort[:bound]:
# batch_idx: index into batch, when sequences in batch are in unsorted order
# seq_idx: index of item in sequence
assert seq_idx < lengths[batch_idx]
indices.append(map_index(batch_idx, seq_idx))
result = gather_from_indices(indices)
return cls(
_make_packed_sequence(result, batch_bounds),
sorted_lengths, sort_to_orig, orig_to_sort)
@classmethod
def cat_seqs(cls, items):
# Check that all items have the same batch size
batch_size = len(items[0].lengths)
assert all(len(item.lengths) == batch_size for item in items[1:])
# Get length of each sequence after concatenation
unsorted_concat_lengths = np.zeros(batch_size, dtype=np.int)
for item in items:
unsorted_concat_lengths += list(item.orig_lengths())
# For each sequence in the result, figure out which item each seq_idx belongs to
concat_data = torch.cat([item.ps.data for item in items], dim=0)
concat_data_base_indices = np.cumsum([0] + [item.ps.data.shape[0] for item in items])
item_map_per_batch_item = []
for batch_idx in range(batch_size):
item_map_per_batch_item.append([
(item_idx, item, i)
for item_idx, item in enumerate(items)
for i in range(item.lengths[item.sort_to_orig[batch_idx]])])
def map_index(batch_idx, seq_idx):
item_idx, item, seq_idx_within_item = item_map_per_batch_item[batch_idx][seq_idx]
return concat_data_base_indices[item_idx] + item.raw_index(batch_idx, seq_idx_within_item)
return cls.from_gather(
lengths=unsorted_concat_lengths,
map_index=map_index,
gather_from_indices=lambda indices: concat_data[torch.LongTensor(indices)])
|