File size: 6,018 Bytes
d36d50b |
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 |
from typing import (
Tuple,
List,
Union,
Dict,
Optional,
Callable,
)
from collections import namedtuple
from abc import ABC, abstractmethod
import torch as T
from torch import nn
from torch.nn import functional as F
from torch import Tensor
import pdb
from dataclasses import dataclass
class IRecurrentCell(ABC, nn.Module):
@abstractmethod
def get_init_state(self, input: Tensor):
pass
@abstractmethod
def loop(self, inputs, state_t0, mask=None):
pass
# def forward(self, input, state, mask=None):
# pass
@dataclass
class IRecurrentCellBuilder(ABC):
hidden_size: int
def make(self, input_size: int) -> IRecurrentCell:
pass
def make_scripted(self, *p, **ks) -> IRecurrentCell:
return T.jit.script(self.make(*p, **ks))
class RecurrentLayer(nn.Module):
def reorder_inputs(self, inputs: Union[List[T.Tensor], T.Tensor]):
#^ inputs : [t b i]
if self.direction == 'backward':
return inputs[::-1]
return inputs
def __init__(
self,
cell: IRecurrentCell,
direction='forward',
batch_first=False,
):
super().__init__()
if isinstance(batch_first, bool):
batch_first = (batch_first, batch_first)
self.batch_first = batch_first
self.direction = direction
self.cell_: IRecurrentCell = cell
@T.jit.ignore
def forward(self, input, state_t0, return_state=None):
if self.batch_first[0]:
#^ input : [b t i]
input = input.transpose(1, 0)
#^ input : [t b i]
inputs = input.unbind(0)
if state_t0 is None:
state_t0 = self.cell_.get_init_state(input)
inputs = self.reorder_inputs(inputs)
if return_state:
sequence, state = self.cell_.loop(inputs, state_t0)
else:
sequence, _ = self.cell_.loop(inputs, state_t0)
#^ sequence : t * [b h]
sequence = self.reorder_inputs(sequence)
sequence = T.stack(sequence)
#^ sequence : [t b h]
if self.batch_first[1]:
sequence = sequence.transpose(1, 0)
#^ sequence : [b t h]
if return_state:
return sequence, state
else:
return sequence, None
class BidirectionalRecurrentLayer(nn.Module):
def __init__(
self,
input_size: int,
cell_builder: IRecurrentCellBuilder,
batch_first=False,
return_states=False
):
super().__init__()
self.batch_first = batch_first
self.cell_builder = cell_builder
self.batch_first = batch_first
self.return_states = return_states
self.fwd = RecurrentLayer(
cell_builder.make_scripted(input_size),
direction='forward',
batch_first=batch_first
)
self.bwd = RecurrentLayer(
cell_builder.make_scripted(input_size),
direction='backward',
batch_first=batch_first
)
@T.jit.ignore
def forward(self, input, state_t0, is_last):
return_states = is_last and self.return_states
if return_states:
fwd, state_fwd = self.fwd(input, state_t0, return_states)
bwd, state_bwd = self.bwd(input, state_t0, return_states)
return T.cat([fwd, bwd], dim=-1), (T.cat([state_fwd[0], state_bwd[0]], dim=-1), T.cat([state_fwd[1], state_bwd[1]], dim=-1))
else:
fwd, _ = self.fwd(input, state_t0, return_states)
bwd, _ = self.bwd(input, state_t0, return_states)
return T.cat([fwd, bwd], dim=-1), None
class RecurrentLayerStack(nn.Module):
def __init__(
self,
cell_builder : Callable[..., IRecurrentCellBuilder],
input_size : int,
num_layers : int,
bidirectional : bool = False,
batch_first : bool = False,
scripted : bool = True,
return_states : bool = False,
*args, **kargs,
):
super().__init__()
cell_builder_: IRecurrentCellBuilder = cell_builder(*args, **kargs)
self._cell_builder = cell_builder_
if bidirectional:
Dh = cell_builder_.hidden_size * 2
def make(isize: int, last=False):
return BidirectionalRecurrentLayer(isize, cell_builder_,
batch_first=batch_first, return_states=return_states)
else:
Dh = cell_builder_.hidden_size
def make(isize: int, last=False):
cell = cell_builder_.make_scripted(isize)
return RecurrentLayer(cell, isize,
batch_first=batch_first)
if num_layers > 1:
rnns = [
make(input_size),
*[
make(Dh)
for _ in range(num_layers - 2)
],
make(Dh, last=True)
]
else:
rnns = [make(input_size, last=True)]
self.rnn = nn.Sequential(*rnns)
self.input_size = input_size
self.hidden_size = self._cell_builder.hidden_size
self.num_layers = num_layers
self.bidirectional = bidirectional
self.return_states = return_states
def __repr__(self):
return (
f'${self.__class__.__name__}'
+ '('
+ f'in={self.input_size}, '
+ f'hid={self.hidden_size}, '
+ f'layers={self.num_layers}, '
+ f'bi={self.bidirectional}'
+ '; '
+ str(self._cell_builder)
)
def forward(self, input, state_t0=None):
for layer_idx, rnn in enumerate(self.rnn):
is_last = (layer_idx == (len(self.rnn) - 1))
input, state = rnn(input, state_t0, is_last)
if self.return_states:
return input, state
return input
|