IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an email to s dot adaszewski at gmail dot com. User accounts are meant only to report issues and/or generate pull requests. This is a purpose-specific Git hosting for ADARED projects. Thank you for your understanding!
Procházet zdrojové kódy

Start triacontagon.

master
Stanislaw Adaszewski před 4 roky
rodič
revize
d57af9c090
12 změnil soubory, kde provedl 1517 přidání a 0 odebrání
  1. +209
    -0
      src/triacontagon/data.py
  2. +123
    -0
      src/triacontagon/decode.py
  3. +42
    -0
      src/triacontagon/dropout.py
  4. +255
    -0
      src/triacontagon/fastconv.py
  5. +138
    -0
      src/triacontagon/fastdec.py
  6. +166
    -0
      src/triacontagon/fastloop.py
  7. +79
    -0
      src/triacontagon/fastmodel.py
  8. +79
    -0
      src/triacontagon/input.py
  9. +145
    -0
      src/triacontagon/normalize.py
  10. +47
    -0
      src/triacontagon/sampling.py
  11. +215
    -0
      src/triacontagon/trainprep.py
  12. +19
    -0
      src/triacontagon/weights.py

+ 209
- 0
src/triacontagon/data.py Zobrazit soubor

@@ -0,0 +1,209 @@
#
# Copyright (C) Stanislaw Adaszewski, 2020
# License: GPLv3
#
from collections import defaultdict
from dataclasses import dataclass, field
import torch
from typing import List, \
Dict, \
Tuple, \
Any, \
Type
from .decode import DEDICOMDecoder, \
BilinearDecoder
import numpy as np
def _equal(x: torch.Tensor, y: torch.Tensor):
if x.is_sparse ^ y.is_sparse:
raise ValueError('Cannot mix sparse and dense tensors')
if not x.is_sparse:
return (x == y)
return ((x - y).coalesce().values() == 0)
@dataclass
class NodeType(object):
name: str
count: int
@dataclass
class RelationTypeBase(object):
name: str
node_type_row: int
node_type_column: int
adjacency_matrix: torch.Tensor
adjacency_matrix_backward: torch.Tensor
@dataclass
class RelationType(RelationTypeBase):
pass
@dataclass
class RelationFamilyBase(object):
data: 'Data'
name: str
node_type_row: int
node_type_column: int
is_symmetric: bool
decoder_class: Type
@dataclass
class RelationFamily(RelationFamilyBase):
relation_types: List[RelationType] = None
def __post_init__(self) -> None:
if not self.is_symmetric and \
self.decoder_class != DEDICOMDecoder and \
self.decoder_class != BilinearDecoder:
raise TypeError('Family is assymetric but the specified decoder_class supports symmetric relations only')
self.relation_types = []
def add_relation_type(self,
name: str, adjacency_matrix: torch.Tensor,
adjacency_matrix_backward: torch.Tensor = None) -> None:
name = str(name)
node_type_row = self.node_type_row
node_type_column = self.node_type_column
if adjacency_matrix is None and adjacency_matrix_backward is None:
raise ValueError('adjacency_matrix and adjacency_matrix_backward cannot both be None')
if adjacency_matrix is not None and \
not isinstance(adjacency_matrix, torch.Tensor):
raise ValueError('adjacency_matrix must be a torch.Tensor')
if adjacency_matrix_backward is not None \
and not isinstance(adjacency_matrix_backward, torch.Tensor):
raise ValueError('adjacency_matrix_backward must be a torch.Tensor')
if adjacency_matrix is not None and \
adjacency_matrix.shape != (self.data.node_types[node_type_row].count,
self.data.node_types[node_type_column].count):
raise ValueError('adjacency_matrix shape must be (num_row_nodes, num_column_nodes)')
if adjacency_matrix_backward is not None and \
adjacency_matrix_backward.shape != (self.data.node_types[node_type_column].count,
self.data.node_types[node_type_row].count):
raise ValueError('adjacency_matrix_backward shape must be (num_column_nodes, num_row_nodes)')
if node_type_row == node_type_column and \
adjacency_matrix_backward is not None:
raise ValueError('Relation between nodes of the same type must be expressed using a single matrix')
if self.is_symmetric and adjacency_matrix_backward is not None:
raise ValueError('Cannot use a custom adjacency_matrix_backward in a symmetric relation family')
if self.is_symmetric and node_type_row == node_type_column and \
not torch.all(_equal(adjacency_matrix,
adjacency_matrix.transpose(0, 1))):
raise ValueError('Relation family is symmetric but adjacency_matrix is assymetric')
if not self.is_symmetric and node_type_row != node_type_column and \
adjacency_matrix_backward is None:
raise ValueError('Relation is asymmetric but adjacency_matrix_backward is None')
if self.is_symmetric and node_type_row != node_type_column:
adjacency_matrix_backward = adjacency_matrix.transpose(0, 1)
self.relation_types.append(RelationType(name,
node_type_row, node_type_column,
adjacency_matrix, adjacency_matrix_backward))
def node_name(self, index):
return self.data.node_types[index].name
def __repr__(self):
s = 'Relation family %s' % self.name
for r in self.relation_types:
s += '\n - %s%s' % (r.name, ' (two-way)' \
if (r.adjacency_matrix is not None \
and r.adjacency_matrix_backward is not None) \
or self.node_type_row == self.node_type_column \
else '%s <- %s' % (self.node_name(self.node_type_row),
self.node_name(self.node_type_column)))
return s
def repr_indented(self):
s = ' - %s' % self.name
for r in self.relation_types:
s += '\n - %s%s' % (r.name, ' (two-way)' \
if (r.adjacency_matrix is not None \
and r.adjacency_matrix_backward is not None) \
or self.node_type_row == self.node_type_column \
else '%s <- %s' % (self.node_name(self.node_type_row),
self.node_name(self.node_type_column)))
return s
class Data(object):
node_types: List[NodeType]
relation_families: List[RelationFamily]
def __init__(self) -> None:
self.node_types = []
self.relation_families = []
def add_node_type(self, name: str, count: int) -> None:
name = str(name)
count = int(count)
if not name:
raise ValueError('You must provide a non-empty node type name')
if count <= 0:
raise ValueError('You must provide a positive node count')
self.node_types.append(NodeType(name, count))
def add_relation_family(self, name: str, node_type_row: int,
node_type_column: int, is_symmetric: bool,
decoder_class: Type = DEDICOMDecoder):
name = str(name)
node_type_row = int(node_type_row)
node_type_column = int(node_type_column)
is_symmetric = bool(is_symmetric)
if node_type_row < 0 or node_type_row >= len(self.node_types):
raise ValueError('node_type_row outside of the valid range of node types')
if node_type_column < 0 or node_type_column >= len(self.node_types):
raise ValueError('node_type_column outside of the valid range of node types')
fam = RelationFamily(self, name, node_type_row, node_type_column,
is_symmetric, decoder_class)
self.relation_families.append(fam)
return fam
def __repr__(self):
n = len(self.node_types)
if n == 0:
return 'Empty Icosagon Data'
s = ''
s += 'Icosagon Data with:\n'
s += '- ' + str(n) + ' node type(s):\n'
for nt in self.node_types:
s += ' - ' + nt.name + '\n'
if len(self.relation_families) == 0:
s += '- No relation families\n'
return s.strip()
s += '- %d relation families:\n' % len(self.relation_families)
for fam in self.relation_families:
s += fam.repr_indented() + '\n'
return s.strip()

+ 123
- 0
src/triacontagon/decode.py Zobrazit soubor

@@ -0,0 +1,123 @@
#
# Copyright (C) Stanislaw Adaszewski, 2020
# License: GPLv3
#
import torch
from .weights import init_glorot
from .dropout import dropout
class DEDICOMDecoder(torch.nn.Module):
"""DEDICOM Tensor Factorization Decoder model layer for link prediction."""
def __init__(self, input_dim, num_relation_types, keep_prob=1.,
activation=torch.sigmoid, **kwargs):
super().__init__(**kwargs)
self.input_dim = input_dim
self.num_relation_types = num_relation_types
self.keep_prob = keep_prob
self.activation = activation
self.global_interaction = torch.nn.Parameter(init_glorot(input_dim, input_dim))
self.local_variation = torch.nn.ParameterList([
torch.nn.Parameter(torch.flatten(init_glorot(input_dim, 1))) \
for _ in range(num_relation_types)
])
def forward(self, inputs_row, inputs_col, relation_index):
inputs_row = dropout(inputs_row, self.keep_prob)
inputs_col = dropout(inputs_col, self.keep_prob)
relation = torch.diag(self.local_variation[relation_index])
product1 = torch.mm(inputs_row, relation)
product2 = torch.mm(product1, self.global_interaction)
product3 = torch.mm(product2, relation)
rec = torch.bmm(product3.view(product3.shape[0], 1, product3.shape[1]),
inputs_col.view(inputs_col.shape[0], inputs_col.shape[1], 1))
rec = torch.flatten(rec)
return self.activation(rec)
class DistMultDecoder(torch.nn.Module):
"""DEDICOM Tensor Factorization Decoder model layer for link prediction."""
def __init__(self, input_dim, num_relation_types, keep_prob=1.,
activation=torch.sigmoid, **kwargs):
super().__init__(**kwargs)
self.input_dim = input_dim
self.num_relation_types = num_relation_types
self.keep_prob = keep_prob
self.activation = activation
self.relation = torch.nn.ParameterList([
torch.nn.Parameter(torch.flatten(init_glorot(input_dim, 1))) \
for _ in range(num_relation_types)
])
def forward(self, inputs_row, inputs_col, relation_index):
inputs_row = dropout(inputs_row, self.keep_prob)
inputs_col = dropout(inputs_col, self.keep_prob)
relation = torch.diag(self.relation[relation_index])
intermediate_product = torch.mm(inputs_row, relation)
rec = torch.bmm(intermediate_product.view(intermediate_product.shape[0], 1, intermediate_product.shape[1]),
inputs_col.view(inputs_col.shape[0], inputs_col.shape[1], 1))
rec = torch.flatten(rec)
return self.activation(rec)
class BilinearDecoder(torch.nn.Module):
"""DEDICOM Tensor Factorization Decoder model layer for link prediction."""
def __init__(self, input_dim, num_relation_types, keep_prob=1.,
activation=torch.sigmoid, **kwargs):
super().__init__(**kwargs)
self.input_dim = input_dim
self.num_relation_types = num_relation_types
self.keep_prob = keep_prob
self.activation = activation
self.relation = torch.nn.ParameterList([
torch.nn.Parameter(init_glorot(input_dim, input_dim)) \
for _ in range(num_relation_types)
])
def forward(self, inputs_row, inputs_col, relation_index):
inputs_row = dropout(inputs_row, self.keep_prob)
inputs_col = dropout(inputs_col, self.keep_prob)
intermediate_product = torch.mm(inputs_row, self.relation[relation_index])
rec = torch.bmm(intermediate_product.view(intermediate_product.shape[0], 1, intermediate_product.shape[1]),
inputs_col.view(inputs_col.shape[0], inputs_col.shape[1], 1))
rec = torch.flatten(rec)
return self.activation(rec)
class InnerProductDecoder(torch.nn.Module):
"""DEDICOM Tensor Factorization Decoder model layer for link prediction."""
def __init__(self, input_dim, num_relation_types, keep_prob=1.,
activation=torch.sigmoid, **kwargs):
super().__init__(**kwargs)
self.input_dim = input_dim
self.num_relation_types = num_relation_types
self.keep_prob = keep_prob
self.activation = activation
def forward(self, inputs_row, inputs_col, _):
inputs_row = dropout(inputs_row, self.keep_prob)
inputs_col = dropout(inputs_col, self.keep_prob)
rec = torch.bmm(inputs_row.view(inputs_row.shape[0], 1, inputs_row.shape[1]),
inputs_col.view(inputs_col.shape[0], inputs_col.shape[1], 1))
rec = torch.flatten(rec)
return self.activation(rec)

+ 42
- 0
src/triacontagon/dropout.py Zobrazit soubor

@@ -0,0 +1,42 @@
#
# Copyright (C) Stanislaw Adaszewski, 2020
# License: GPLv3
#
import torch
from .normalize import _sparse_coo_tensor
def dropout_sparse(x, keep_prob):
x = x.coalesce()
i = x._indices()
v = x._values()
size = x.size()
n = keep_prob + torch.rand(len(v))
n = torch.floor(n).to(torch.bool)
i = i[:,n]
v = v[n]
x = _sparse_coo_tensor(i, v, size=size)
return x * (1./keep_prob)
def dropout_dense(x, keep_prob):
# print('dropout_dense()')
x = x.clone()
i = torch.nonzero(x)
n = keep_prob + torch.rand(len(i))
n = (1. - torch.floor(n)).to(torch.bool)
x[i[n, 0], i[n, 1]] = 0.
return x * (1./keep_prob)
def dropout(x, keep_prob):
if x.is_sparse:
return dropout_sparse(x, keep_prob)
else:
return dropout_dense(x, keep_prob)

+ 255
- 0
src/triacontagon/fastconv.py Zobrazit soubor

@@ -0,0 +1,255 @@
from typing import List, \
Union, \
Callable
from .data import Data, \
RelationFamily
from .trainprep import PreparedData, \
PreparedRelationFamily
import torch
from .weights import init_glorot
from .normalize import _sparse_coo_tensor
import types
def _sparse_diag_cat(matrices: List[torch.Tensor]):
if len(matrices) == 0:
raise ValueError('The list of matrices must be non-empty')
if not all(m.is_sparse for m in matrices):
raise ValueError('All matrices must be sparse')
if not all(len(m.shape) == 2 for m in matrices):
raise ValueError('All matrices must be 2D')
indices = []
values = []
row_offset = 0
col_offset = 0
for m in matrices:
ind = m._indices().clone()
ind[0] += row_offset
ind[1] += col_offset
indices.append(ind)
values.append(m._values())
row_offset += m.shape[0]
col_offset += m.shape[1]
indices = torch.cat(indices, dim=1)
values = torch.cat(values)
return _sparse_coo_tensor(indices, values, size=(row_offset, col_offset))
def _cat(matrices: List[torch.Tensor]):
if len(matrices) == 0:
raise ValueError('Empty list passed to _cat()')
n = sum(a.is_sparse for a in matrices)
if n != 0 and n != len(matrices):
raise ValueError('All matrices must have the same layout (dense or sparse)')
if not all(a.shape[1:] == matrices[0].shape[1:] for a in matrices):
raise ValueError('All matrices must have the same dimensions apart from dimension 0')
if not matrices[0].is_sparse:
return torch.cat(matrices)
total_rows = sum(a.shape[0] for a in matrices)
indices = []
values = []
row_offset = 0
for a in matrices:
ind = a._indices().clone()
val = a._values()
ind[0] += row_offset
ind = ind.transpose(0, 1)
indices.append(ind)
values.append(val)
row_offset += a.shape[0]
indices = torch.cat(indices).transpose(0, 1)
values = torch.cat(values)
res = _sparse_coo_tensor(indices, values, size=(row_offset, matrices[0].shape[1]))
return res
class FastGraphConv(torch.nn.Module):
def __init__(self,
in_channels: int,
out_channels: int,
adjacency_matrices: List[torch.Tensor],
keep_prob: float = 1.,
activation: Callable[[torch.Tensor], torch.Tensor] = lambda x: x,
**kwargs) -> None:
super().__init__(**kwargs)
in_channels = int(in_channels)
out_channels = int(out_channels)
if not isinstance(adjacency_matrices, list):
raise TypeError('adjacency_matrices must be a list')
if len(adjacency_matrices) == 0:
raise ValueError('adjacency_matrices must not be empty')
if not all(isinstance(m, torch.Tensor) for m in adjacency_matrices):
raise TypeError('adjacency_matrices elements must be of class torch.Tensor')
if not all(m.is_sparse for m in adjacency_matrices):
raise ValueError('adjacency_matrices elements must be sparse')
keep_prob = float(keep_prob)
if not isinstance(activation, types.FunctionType):
raise TypeError('activation must be a function')
self.in_channels = in_channels
self.out_channels = out_channels
self.adjacency_matrices = adjacency_matrices
self.keep_prob = keep_prob
self.activation = activation
self.num_row_nodes = len(adjacency_matrices[0])
self.num_relation_types = len(adjacency_matrices)
self.adjacency_matrices = _sparse_diag_cat(adjacency_matrices)
self.weights = torch.cat([
init_glorot(in_channels, out_channels) \
for _ in range(self.num_relation_types)
], dim=1)
def forward(self, x) -> torch.Tensor:
if self.keep_prob < 1.:
x = dropout(x, self.keep_prob)
res = torch.sparse.mm(x, self.weights) \
if x.is_sparse \
else torch.mm(x, self.weights)
res = torch.split(res, res.shape[1] // self.num_relation_types, dim=1)
res = torch.cat(res)
res = torch.sparse.mm(self.adjacency_matrices, res) \
if self.adjacency_matrices.is_sparse \
else torch.mm(self.adjacency_matrices, res)
res = res.view(self.num_relation_types, self.num_row_nodes, self.out_channels)
if self.activation is not None:
res = self.activation(res)
return res
class FastConvLayer(torch.nn.Module):
def __init__(self,
input_dim: List[int],
output_dim: List[int],
data: Union[Data, PreparedData],
keep_prob: float = 1.,
rel_activation: Callable[[torch.Tensor], torch.Tensor] = lambda x: x,
layer_activation: Callable[[torch.Tensor], torch.Tensor] = torch.nn.functional.relu,
**kwargs):
super().__init__(**kwargs)
self._check_params(input_dim, output_dim, data, keep_prob,
rel_activation, layer_activation)
self.input_dim = input_dim
self.output_dim = output_dim
self.data = data
self.keep_prob = keep_prob
self.rel_activation = rel_activation
self.layer_activation = layer_activation
self.is_sparse = False
self.next_layer_repr = None
self.build()
def build(self):
self.next_layer_repr = torch.nn.ModuleList([
torch.nn.ModuleList() \
for _ in range(len(self.data.node_types))
])
for fam in self.data.relation_families:
self.build_family(fam)
def build_family(self, fam) -> None:
if fam.node_type_row == fam.node_type_column:
self.build_fam_one_node_type(fam)
else:
self.build_fam_two_node_types(fam)
def build_fam_one_node_type(self, fam) -> None:
adjacency_matrices = [
r.adjacency_matrix \
for r in fam.relation_types
]
conv = FastGraphConv(self.input_dim[fam.node_type_column],
self.output_dim[fam.node_type_row],
adjacency_matrices,
self.keep_prob,
self.rel_activation)
conv.input_node_type = fam.node_type_column
self.next_layer_repr[fam.node_type_row].append(conv)
def build_fam_two_node_types(self, fam) -> None:
adjacency_matrices = [
r.adjacency_matrix \
for r in fam.relation_types \
if r.adjacency_matrix is not None
]
adjacency_matrices_backward = [
r.adjacency_matrix_backward \
for r in fam.relation_types \
if r.adjacency_matrix_backward is not None
]
conv = FastGraphConv(self.input_dim[fam.node_type_column],
self.output_dim[fam.node_type_row],
adjacency_matrices,
self.keep_prob,
self.rel_activation)
conv_backward = FastGraphConv(self.input_dim[fam.node_type_row],
self.output_dim[fam.node_type_column],
adjacency_matrices_backward,
self.keep_prob,
self.rel_activation)
conv.input_node_type = fam.node_type_column
conv_backward.input_node_type = fam.node_type_row
self.next_layer_repr[fam.node_type_row].append(conv)
self.next_layer_repr[fam.node_type_column].append(conv_backward)
def forward(self, prev_layer_repr):
next_layer_repr = [ [] \
for _ in range(len(self.data.node_types)) ]
for output_node_type in range(len(self.data.node_types)):
for conv in self.next_layer_repr[output_node_type]:
rep = conv(prev_layer_repr[conv.input_node_type])
rep = torch.sum(rep, dim=0)
rep = torch.nn.functional.normalize(rep, p=2, dim=1)
next_layer_repr[output_node_type].append(rep)
if len(next_layer_repr[output_node_type]) == 0:
next_layer_repr[output_node_type] = \
torch.zeros(self.data.node_types[output_node_type].count, self.output_dim[output_node_type])
else:
next_layer_repr[output_node_type] = \
sum(next_layer_repr[output_node_type])
next_layer_repr[output_node_type] = \
self.layer_activation(next_layer_repr[output_node_type])
return next_layer_repr
@staticmethod
def _check_params(input_dim, output_dim, data, keep_prob,
rel_activation, layer_activation):
if not isinstance(input_dim, list):
raise ValueError('input_dim must be a list')
if not output_dim:
raise ValueError('output_dim must be specified')
if not isinstance(output_dim, list):
output_dim = [output_dim] * len(data.node_types)
if not isinstance(data, Data) and not isinstance(data, PreparedData):
raise ValueError('data must be of type Data or PreparedData')

+ 138
- 0
src/triacontagon/fastdec.py Zobrazit soubor

@@ -0,0 +1,138 @@
import torch
from typing import List
from .trainprep import PreparedData
from dataclasses import dataclass
import random
from collections import defaultdict
@dataclass
class TrainingBatch(object):
relation_family_index: int
relation_type_index: int
node_type_row: int
node_type_column: int
edges: torch.Tensor
class FastBatcher(object):
def __init__(self,
prep_d: PreparedData,
batch_size: int) -> None:
if not isinstance(prep_d, PreparedData):
raise TypeError('prep_d must be an instance of PreparedData')
self.prep_d = prep_d
self.batch_size = int(batch_size)
self.edges = None
self.build()
def build(self):
self.edges = []
for fam_idx, fam in enumerate(self.prep_d.relation_families):
edges = []
targets = []
edges_back = []
targets_back = []
for rel_idx, rel in enumerate(fam.relation_types):
edges.append(rel.edges_pos.train)
edges.append(rel.edges_neg.train)
targets.append(torch.ones(len(rel.edges_pos.train)))
targets.append(torch.zeros(len(rel.edges_neg.train)))
edges_back.append(rel.edges_back_pos.train)
edges_back.append(rel.edges_back_neg.train)
targets_back.apend(torch.zeros(len(rel.edges_back_pos.train)))
targets_back.apend(torch.zeros(len(rel.edges_back_neg.train)))
edges = torch.cat(edges)
targets = torch.cat(targets)
edges_back = torch.cat(edges_back)
targets_back = torch.cat(targets_back)
order = torch.randperm(len(edges))
edges = edges[order]
targets = targets[order]
order_back = torch.randperm(len(edges_back))
edges_back = edges_back[order_back]
targets_back = targets_back[order_back]
self.edges.append({'fam_idx': fam_idx, 'rel_idx': rel_idx, 'back': False,
'edges': edges, 'targets': targets, 'ofs': 0})
self.edges.append({'fam_idx': fam_idx, 'rel_idx': rel_idx, 'back': True,
'edges': edges_back, 'targets': targets_back, 'ofs': 0})
def __iter__(self):
while True:
edges = [ e for e in self.edges \
if e['ofs'] < len(e['edges']) ]
# TODO: need to finish this
def __iter_old__(self):
edge_types = ['edges_pos', 'edges_neg', 'edges_back_pos', 'edges_back_neg']
offsets = {}
orders = {}
done = {}
for fam_idx, fam in enumerate(self.prep_d.relation_families):
for rel_idx, rel in enumerate(fam.relation_types):
for et in edge_types:
done[fam_idx, rel_idx, et] = False
while True:
fam_idx = torch.randint(0, len(self.prep_d.relation_families), (1,)).item()
fam = self.prep_d.relation_families[fam_idx]
rel_idx = torch.randint(0, len(fam.relation_types), (1,)).item()
rel = fam.relation_types[rel_idx]
et = random.choice(edge_types)
edges = getattr(rel, et).train
key = (fam_idx, rel_idx, et)
if key not in orders:
orders[key] = torch.randperm(len(edges))
offsets[key] = 0
ord = orders[key]
ofs = offsets[key]
nt_row = rel.node_type_row
nt_col = rel.node_type_column
if 'back' in et:
nt_row, nt_col = nt_col, nt_row
if ofs < len(edges):
offsets[key] += self.batch_size
ord = ord[ofs:ofs+self.batch_size]
edges = edges[ord]
yield TrainingBatch(fam_idx, rel_idx, nt_row, nt_column, edges)
else:
done[key] = True
for fam in self.prep_d.relation_families:
edges = []
for rel in fam.relation_types:
edges.append(rel.edges_pos.train)
edges.append(rel.edges_back_pos.train)
edges.append(rel.edges_neg.train)
edges.append(rel.edges_back_neg.train)
edges = torch.cat(e)
class FastDecLayer(torch.nn.Module):
def __init__(self, **kwargs):
super().__init__(**kwargs)
def forward(self,
last_layer_repr: List[torch.Tensor],
training_batch: TrainingBatch):

+ 166
- 0
src/triacontagon/fastloop.py Zobrazit soubor

@@ -0,0 +1,166 @@
from .fastmodel import FastModel
from .trainprep import PreparedData
import torch
from typing import Callable
from types import FunctionType
import time
import random
class FastBatcher(object):
def __init__(self, prep_d: PreparedData, batch_size: int,
shuffle: bool, generator: torch.Generator,
part_type: str) -> None:
if not isinstance(prep_d, PreparedData):
raise TypeError('prep_d must be an instance of PreparedData')
if not isinstance(generator, torch.Generator):
raise TypeError('generator must be an instance of torch.Generator')
if part_type not in ['train', 'val', 'test']:
raise ValueError('part_type must be set to train, val or test')
self.prep_d = prep_d
self.batch_size = int(batch_size)
self.shuffle = bool(shuffle)
self.generator = generator
self.part_type = part_type
self.edges = None
self.targets = None
self.build()
def build(self):
self.edges = []
self.targets = []
for fam in self.prep_d.relation_families:
edges = []
targets = []
for i, rel in enumerate(fam.relation_types):
edges_pos = getattr(rel.edges_pos, self.part_type)
edges_neg = getattr(rel.edges_neg, self.part_type)
edges_back_pos = getattr(rel.edges_back_pos, self.part_type)
edges_back_neg = getattr(rel.edges_back_neg, self.part_type)
e = torch.cat([ edges_pos,
torch.cat([edges_back_pos[:, 1], edges_back_pos[:, 0]], dim=1) ])
e = torch.cat([torch.ones(len(e), 1, dtype=torch.long) * i , e ], dim=1)
t = torch.ones(len(e))
edges.append(e)
targets.append(t)
e = torch.cat([ edges_neg,
torch.cat([edges_back_neg[:, 1], edges_back_neg[:, 0]], dim=1) ])
e = torch.cat([ torch.ones(len(e), 1, dtype=torch.long) * i, e ], dim=1)
t = torch.zeros(len(e))
edges.append(e)
targets.append(t)
edges = torch.cat(edges)
targets = torch.cat(targets)
self.edges.append(edges)
self.targets.append(targets)
# print(self.edges)
# print(self.targets)
if self.shuffle:
self.shuffle_families()
def shuffle_families(self):
for i in range(len(self.edges)):
edges = self.edges[i]
targets = self.targets[i]
order = torch.randperm(len(edges), generator=self.generator)
self.edges[i] = edges[order]
self.targets[i] = targets[order]
def __iter__(self):
offsets = [ 0 for _ in self.edges ]
while True:
choice = [ i for i in range(len(offsets)) \
if offsets[i] < len(self.edges[i]) ]
if len(choice) == 0:
break
fam_idx = torch.randint(len(choice), (1,), generator=self.generator).item()
ofs = offsets[fam_idx]
edges = self.edges[fam_idx][ofs:ofs + self.batch_size]
targets = self.targets[fam_idx][ofs:ofs + self.batch_size]
offsets[fam_idx] += self.batch_size
yield (fam_idx, edges, targets)
class FastLoop(object):
def __init__(
self,
model: FastModel,
lr: float = 0.001,
loss: Callable[[torch.Tensor, torch.Tensor], torch.Tensor] = \
torch.nn.functional.binary_cross_entropy_with_logits,
batch_size: int = 100,
shuffle: bool = True,
generator: torch.Generator = None) -> None:
self._check_params(model, loss, generator)
self.model = model
self.lr = float(lr)
self.loss = loss
self.batch_size = int(batch_size)
self.shuffle = bool(shuffle)
self.generator = generator or torch.default_generator
self.opt = None
self.build()
def _check_params(self, model, loss, generator):
if not isinstance(model, FastModel):
raise TypeError('model must be an instance of FastModel')
if not isinstance(loss, FunctionType):
raise TypeError('loss must be a function')
if generator is not None and not isinstance(generator, torch.Generator):
raise TypeError('generator must be an instance of torch.Generator')
def build(self) -> None:
opt = torch.optim.Adam(self.model.parameters(), lr=self.lr)
self.opt = opt
def run_epoch(self):
prep_d = self.model.prep_d
batcher = FastBatcher(self.model.prep_d, batch_size=self.batch_size,
shuffle = self.shuffle, generator=self.generator)
# pred = self.model(None)
# n = len(list(iter(batch)))
loss_sum = 0
for fam_idx, edges, targets in batcher:
self.opt.zero_grad()
pred = self.model(None)
# process pred, get input and targets
input = pred[fam_idx][edges[:, 0], edges[:, 1]]
loss = self.loss(input, targets)
loss.backward()
self.opt.step()
loss_sum += loss.detach().cpu().item()
return loss_sum
def train(self, max_epochs):
best_loss = None
best_epoch = None
for i in range(max_epochs):
loss = self.run_epoch()
if best_loss is None or loss < best_loss:
best_loss = loss
best_epoch = i
return loss, best_loss, best_epoch

+ 79
- 0
src/triacontagon/fastmodel.py Zobrazit soubor

@@ -0,0 +1,79 @@
from .fastconv import FastConvLayer
from .bulkdec import BulkDecodeLayer
from .input import OneHotInputLayer
from .trainprep import PreparedData
import torch
import types
from typing import List, \
Union, \
Callable
class FastModel(torch.nn.Module):
def __init__(self, prep_d: PreparedData,
layer_dimensions: List[int] = [32, 64],
keep_prob: float = 1.,
rel_activation: Callable[[torch.Tensor], torch.Tensor] = lambda x: x,
layer_activation: Callable[[torch.Tensor], torch.Tensor] = torch.nn.functional.relu,
dec_activation: Callable[[torch.Tensor], torch.Tensor] = lambda x: x,
**kwargs) -> None:
super().__init__(**kwargs)
self._check_params(prep_d, layer_dimensions, rel_activation,
layer_activation, dec_activation)
self.prep_d = prep_d
self.layer_dimensions = layer_dimensions
self.keep_prob = float(keep_prob)
self.rel_activation = rel_activation
self.layer_activation = layer_activation
self.dec_activation = dec_activation
self.seq = None
self.build()
def build(self):
in_layer = OneHotInputLayer(self.prep_d)
last_output_dim = in_layer.output_dim
seq = [ in_layer ]
for dim in self.layer_dimensions:
conv_layer = FastConvLayer(input_dim = last_output_dim,
output_dim = [dim] * len(self.prep_d.node_types),
data = self.prep_d,
keep_prob = self.keep_prob,
rel_activation = self.rel_activation,
layer_activation = self.layer_activation)
last_output_dim = conv_layer.output_dim
seq.append(conv_layer)
dec_layer = BulkDecodeLayer(input_dim = last_output_dim,
data = self.prep_d,
keep_prob = self.keep_prob,
activation = self.dec_activation)
seq.append(dec_layer)
seq = torch.nn.Sequential(*seq)
self.seq = seq
def forward(self, _):
return self.seq(None)
def _check_params(self, prep_d, layer_dimensions, rel_activation,
layer_activation, dec_activation):
if not isinstance(prep_d, PreparedData):
raise TypeError('prep_d must be an instanced of PreparedData')
if not isinstance(layer_dimensions, list):
raise TypeError('layer_dimensions must be a list')
if not isinstance(rel_activation, types.FunctionType):
raise TypeError('rel_activation must be a function')
if not isinstance(layer_activation, types.FunctionType):
raise TypeError('layer_activation must be a function')
if not isinstance(dec_activation, types.FunctionType):
raise TypeError('dec_activation must be a function')

+ 79
- 0
src/triacontagon/input.py Zobrazit soubor

@@ -0,0 +1,79 @@
#
# Copyright (C) Stanislaw Adaszewski, 2020
# License: GPLv3
#
import torch
from typing import Union, \
List
from .data import Data
class InputLayer(torch.nn.Module):
def __init__(self, data: Data, output_dim: Union[int, List[int]] = None,
**kwargs) -> None:
output_dim = output_dim or \
list(map(lambda a: a.count, data.node_types))
if not isinstance(output_dim, list):
output_dim = [output_dim,] * len(data.node_types)
super().__init__(**kwargs)
self.output_dim = output_dim
self.data = data
self.is_sparse=False
self.node_reps = None
self.build()
def build(self) -> None:
self.node_reps = []
for i, nt in enumerate(self.data.node_types):
reps = torch.rand(nt.count, self.output_dim[i])
reps = torch.nn.Parameter(reps)
self.register_parameter('node_reps[%d]' % i, reps)
self.node_reps.append(reps)
def forward(self, x) -> List[torch.nn.Parameter]:
return self.node_reps
def __repr__(self) -> str:
s = ''
s += 'Icosagon input layer with output_dim: %s\n' % self.output_dim
s += ' # of node types: %d\n' % len(self.data.node_types)
for nt in self.data.node_types:
s += ' - %s (%d)\n' % (nt.name, nt.count)
return s.strip()
class OneHotInputLayer(torch.nn.Module):
def __init__(self, data: Data, **kwargs) -> None:
output_dim = [ a.count for a in data.node_types ]
super().__init__(**kwargs)
self.output_dim = output_dim
self.data = data
self.is_sparse=True
self.node_reps = None
self.build()
def build(self) -> None:
self.node_reps = torch.nn.ParameterList()
for i, nt in enumerate(self.data.node_types):
reps = torch.eye(nt.count).to_sparse()
reps = torch.nn.Parameter(reps, requires_grad=False)
# self.register_parameter('node_reps[%d]' % i, reps)
self.node_reps.append(reps)
def forward(self, x) -> List[torch.nn.Parameter]:
return self.node_reps
def __repr__(self) -> str:
s = ''
s += 'Icosagon one-hot input layer\n'
s += ' # of node types: %d\n' % len(self.data.node_types)
for nt in self.data.node_types:
s += ' - %s (%d)\n' % (nt.name, nt.count)
return s.strip()

+ 145
- 0
src/triacontagon/normalize.py Zobrazit soubor

@@ -0,0 +1,145 @@
#
# Copyright (C) Stanislaw Adaszewski, 2020
# License: GPLv3
#
import numpy as np
import scipy.sparse as sp
import torch
def _check_tensor(adj_mat):
if not isinstance(adj_mat, torch.Tensor):
raise ValueError('adj_mat must be a torch.Tensor')
def _check_sparse(adj_mat):
if not adj_mat.is_sparse:
raise ValueError('adj_mat must be sparse')
def _check_dense(adj_mat):
if adj_mat.is_sparse:
raise ValueError('adj_mat must be dense')
def _check_square(adj_mat):
if len(adj_mat.shape) != 2 or \
adj_mat.shape[0] != adj_mat.shape[1]:
raise ValueError('adj_mat must be a square matrix')
def _check_2d(adj_mat):
if len(adj_mat.shape) != 2:
raise ValueError('adj_mat must be a square matrix')
def _sparse_coo_tensor(indices, values, size):
ctor = { torch.float32: torch.sparse.FloatTensor,
torch.float32: torch.sparse.DoubleTensor,
torch.uint8: torch.sparse.ByteTensor,
torch.long: torch.sparse.LongTensor,
torch.int: torch.sparse.IntTensor,
torch.short: torch.sparse.ShortTensor,
torch.bool: torch.sparse.ByteTensor }[values.dtype]
return ctor(indices, values, size)
def add_eye_sparse(adj_mat: torch.Tensor) -> torch.Tensor:
_check_tensor(adj_mat)
_check_sparse(adj_mat)
_check_square(adj_mat)
adj_mat = adj_mat.coalesce()
indices = adj_mat.indices()
values = adj_mat.values()
eye_indices = torch.arange(adj_mat.shape[0], dtype=indices.dtype,
device=adj_mat.device).view(1, -1)
eye_indices = torch.cat((eye_indices, eye_indices), 0)
eye_values = torch.ones(adj_mat.shape[0], dtype=values.dtype,
device=adj_mat.device)
indices = torch.cat((indices, eye_indices), 1)
values = torch.cat((values, eye_values), 0)
adj_mat = _sparse_coo_tensor(indices, values, adj_mat.shape)
return adj_mat
def norm_adj_mat_one_node_type_sparse(adj_mat: torch.Tensor) -> torch.Tensor:
_check_tensor(adj_mat)
_check_sparse(adj_mat)
_check_square(adj_mat)
adj_mat = add_eye_sparse(adj_mat)
adj_mat = norm_adj_mat_two_node_types_sparse(adj_mat)
return adj_mat
def norm_adj_mat_one_node_type_dense(adj_mat: torch.Tensor) -> torch.Tensor:
_check_tensor(adj_mat)
_check_dense(adj_mat)
_check_square(adj_mat)
adj_mat = adj_mat + torch.eye(adj_mat.shape[0], dtype=adj_mat.dtype,
device=adj_mat.device)
adj_mat = norm_adj_mat_two_node_types_dense(adj_mat)
return adj_mat
def norm_adj_mat_one_node_type(adj_mat: torch.Tensor) -> torch.Tensor:
_check_tensor(adj_mat)
_check_square(adj_mat)
if adj_mat.is_sparse:
return norm_adj_mat_one_node_type_sparse(adj_mat)
else:
return norm_adj_mat_one_node_type_dense(adj_mat)
def norm_adj_mat_two_node_types_sparse(adj_mat: torch.Tensor) -> torch.Tensor:
_check_tensor(adj_mat)
_check_sparse(adj_mat)
_check_2d(adj_mat)
adj_mat = adj_mat.coalesce()
indices = adj_mat.indices()
values = adj_mat.values()
degrees_row = torch.zeros(adj_mat.shape[0], device=adj_mat.device)
degrees_row = degrees_row.index_add(0, indices[0], values.to(degrees_row.dtype))
degrees_col = torch.zeros(adj_mat.shape[1], device=adj_mat.device)
degrees_col = degrees_col.index_add(0, indices[1], values.to(degrees_col.dtype))
values = values.to(degrees_row.dtype) / torch.sqrt(degrees_row[indices[0]] * degrees_col[indices[1]])
adj_mat = _sparse_coo_tensor(indices, values, adj_mat.shape)
return adj_mat
def norm_adj_mat_two_node_types_dense(adj_mat: torch.Tensor) -> torch.Tensor:
_check_tensor(adj_mat)
_check_dense(adj_mat)
_check_2d(adj_mat)
degrees_row = adj_mat.sum(1).view(-1, 1).to(torch.float32)
degrees_col = adj_mat.sum(0).view(1, -1).to(torch.float32)
degrees_row = torch.sqrt(degrees_row)
degrees_col = torch.sqrt(degrees_col)
adj_mat = adj_mat.to(degrees_row.dtype) / degrees_row
adj_mat = adj_mat / degrees_col
return adj_mat
def norm_adj_mat_two_node_types(adj_mat: torch.Tensor) -> torch.Tensor:
_check_tensor(adj_mat)
_check_2d(adj_mat)
if adj_mat.is_sparse:
return norm_adj_mat_two_node_types_sparse(adj_mat)
else:
return norm_adj_mat_two_node_types_dense(adj_mat)

+ 47
- 0
src/triacontagon/sampling.py Zobrazit soubor

@@ -0,0 +1,47 @@
#
# Copyright (C) Stanislaw Adaszewski, 2020
# License: GPLv3
#
import numpy as np
import torch
import torch.utils.data
from typing import List, \
Union
def fixed_unigram_candidate_sampler(
true_classes: Union[np.array, torch.Tensor],
unigrams: List[Union[int, float]],
distortion: float = 1.):
if isinstance(true_classes, torch.Tensor):
true_classes = true_classes.detach().cpu().numpy()
if isinstance(unigrams, torch.Tensor):
unigrams = unigrams.detach().cpu().numpy()
if len(true_classes.shape) != 2:
raise ValueError('true_classes must be a 2D matrix with shape (num_samples, num_true)')
num_samples = true_classes.shape[0]
unigrams = np.array(unigrams)
if distortion != 1.:
unigrams = unigrams.astype(np.float64) ** distortion
# print('unigrams:', unigrams)
indices = np.arange(num_samples)
result = np.zeros(num_samples, dtype=np.int64)
while len(indices) > 0:
# print('len(indices):', len(indices))
sampler = torch.utils.data.WeightedRandomSampler(unigrams, len(indices))
candidates = np.array(list(sampler))
candidates = np.reshape(candidates, (len(indices), 1))
# print('candidates:', candidates)
# print('true_classes:', true_classes[indices, :])
result[indices] = candidates.T
mask = (candidates == true_classes[indices, :])
mask = mask.sum(1).astype(np.bool)
# print('mask:', mask)
indices = indices[mask]
return torch.tensor(result)

+ 215
- 0
src/triacontagon/trainprep.py Zobrazit soubor

@@ -0,0 +1,215 @@
#
# Copyright (C) Stanislaw Adaszewski, 2020
# License: GPLv3
#
from .sampling import fixed_unigram_candidate_sampler
import torch
from dataclasses import dataclass, \
field
from typing import Any, \
List, \
Tuple, \
Dict
from .data import NodeType, \
RelationType, \
RelationTypeBase, \
RelationFamily, \
RelationFamilyBase, \
Data
from collections import defaultdict
from .normalize import norm_adj_mat_one_node_type, \
norm_adj_mat_two_node_types
import numpy as np
@dataclass
class TrainValTest(object):
train: Any
val: Any
test: Any
@dataclass
class PreparedRelationType(RelationTypeBase):
edges_pos: TrainValTest
edges_neg: TrainValTest
edges_back_pos: TrainValTest
edges_back_neg: TrainValTest
@dataclass
class PreparedRelationFamily(RelationFamilyBase):
relation_types: List[PreparedRelationType]
@dataclass
class PreparedData(object):
node_types: List[NodeType]
relation_families: List[PreparedRelationFamily]
def _empty_edge_list_tvt() -> TrainValTest:
return TrainValTest(*[ torch.zeros((0, 2), dtype=torch.long) for _ in range(3) ])
def train_val_test_split_edges(edges: torch.Tensor,
ratios: TrainValTest) -> TrainValTest:
if not isinstance(edges, torch.Tensor):
raise ValueError('edges must be a torch.Tensor')
if len(edges.shape) != 2 or edges.shape[1] != 2:
raise ValueError('edges shape must be (num_edges, 2)')
if not isinstance(ratios, TrainValTest):
raise ValueError('ratios must be a TrainValTest')
if ratios.train + ratios.val + ratios.test != 1.0:
raise ValueError('Train, validation and test ratios must add up to 1')
order = torch.randperm(len(edges))
edges = edges[order, :]
n = round(len(edges) * ratios.train)
edges_train = edges[:n]
n_1 = round(len(edges) * (ratios.train + ratios.val))
edges_val = edges[n:n_1]
edges_test = edges[n_1:]
return TrainValTest(edges_train, edges_val, edges_test)
def get_edges_and_degrees(adj_mat: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
if adj_mat.is_sparse:
adj_mat = adj_mat.coalesce()
degrees = torch.zeros(adj_mat.shape[1], dtype=torch.int64,
device=adj_mat.device)
degrees = degrees.index_add(0, adj_mat.indices()[1],
torch.ones(adj_mat.indices().shape[1], dtype=torch.int64,
device=adj_mat.device))
edges_pos = adj_mat.indices().transpose(0, 1)
else:
degrees = adj_mat.sum(0)
edges_pos = torch.nonzero(adj_mat)
return edges_pos, degrees
def prepare_adj_mat(adj_mat: torch.Tensor,
ratios: TrainValTest) -> Tuple[TrainValTest, TrainValTest]:
if not isinstance(adj_mat, torch.Tensor):
raise ValueError('adj_mat must be a torch.Tensor')
edges_pos, degrees = get_edges_and_degrees(adj_mat)
neg_neighbors = fixed_unigram_candidate_sampler(
edges_pos[:, 1].view(-1, 1), degrees, 0.75).to(adj_mat.device)
print(edges_pos.dtype)
print(neg_neighbors.dtype)
edges_neg = torch.cat((edges_pos[:, 0].view(-1, 1), neg_neighbors.view(-1, 1)), 1)
edges_pos = train_val_test_split_edges(edges_pos, ratios)
edges_neg = train_val_test_split_edges(edges_neg, ratios)
adj_mat_train = torch.sparse_coo_tensor(indices = edges_pos.train.transpose(0, 1),
values=torch.ones(len(edges_pos.train)), size=adj_mat.shape, dtype=adj_mat.dtype,
device=adj_mat.device)
return adj_mat_train, edges_pos, edges_neg
def prep_rel_one_node_type(r: RelationType,
ratios: TrainValTest) -> PreparedRelationType:
adj_mat = r.adjacency_matrix
adj_mat_train, edges_pos, edges_neg = prepare_adj_mat(adj_mat, ratios)
adj_mat_back_train, edges_back_pos, edges_back_neg = \
None, _empty_edge_list_tvt(), _empty_edge_list_tvt()
print('adj_mat_train:', adj_mat_train)
adj_mat_train = norm_adj_mat_one_node_type(adj_mat_train)
return PreparedRelationType(r.name, r.node_type_row, r.node_type_column,
adj_mat_train, adj_mat_back_train, edges_pos, edges_neg,
edges_back_pos, edges_back_neg)
def prep_rel_two_node_types_sym(r: RelationType,
ratios: TrainValTest) -> PreparedRelationType:
adj_mat = r.adjacency_matrix
adj_mat_train, edges_pos, edges_neg = prepare_adj_mat(adj_mat, ratios)
edges_back_pos, edges_back_neg = \
_empty_edge_list_tvt(), _empty_edge_list_tvt()
return PreparedRelationType(r.name, r.node_type_row,
r.node_type_column,
norm_adj_mat_two_node_types(adj_mat_train),
norm_adj_mat_two_node_types(adj_mat_train.transpose(0, 1)),
edges_pos, edges_neg, edges_back_pos, edges_back_neg)
def prep_rel_two_node_types_asym(r: RelationType,
ratios: TrainValTest) -> PreparedRelationType:
if r.adjacency_matrix is not None:
adj_mat_train, edges_pos, edges_neg =\
prepare_adj_mat(r.adjacency_matrix, ratios)
else:
adj_mat_train, edges_pos, edges_neg = \
None, _empty_edge_list_tvt(), _empty_edge_list_tvt()
if r.adjacency_matrix_backward is not None:
adj_mat_back_train, edges_back_pos, edges_back_neg = \
prepare_adj_mat(r.adjacency_matrix_backward, ratios)
else:
adj_mat_back_train, edges_back_pos, edges_back_neg = \
None, _empty_edge_list_tvt(), _empty_edge_list_tvt()
return PreparedRelationType(r.name, r.node_type_row,
r.node_type_column,
norm_adj_mat_two_node_types(adj_mat_train),
norm_adj_mat_two_node_types(adj_mat_back_train),
edges_pos, edges_neg, edges_back_pos, edges_back_neg)
def prepare_relation_type(r: RelationType,
ratios: TrainValTest, is_symmetric: bool) -> PreparedRelationType:
if not isinstance(r, RelationType):
raise ValueError('r must be a RelationType')
if not isinstance(ratios, TrainValTest):
raise ValueError('ratios must be a TrainValTest')
if r.node_type_row == r.node_type_column:
return prep_rel_one_node_type(r, ratios)
elif is_symmetric:
return prep_rel_two_node_types_sym(r, ratios)
else:
return prep_rel_two_node_types_asym(r, ratios)
def prepare_relation_family(fam: RelationFamily,
ratios: TrainValTest) -> PreparedRelationFamily:
relation_types = []
for r in fam.relation_types:
relation_types.append(prepare_relation_type(r, ratios, fam.is_symmetric))
return PreparedRelationFamily(fam.data, fam.name,
fam.node_type_row, fam.node_type_column,
fam.is_symmetric, fam.decoder_class,
relation_types)
def prepare_training(data: Data, ratios: TrainValTest) -> PreparedData:
if not isinstance(data, Data):
raise ValueError('data must be of class Data')
relation_families = [ prepare_relation_family(fam, ratios) \
for fam in data.relation_families ]
return PreparedData(data.node_types, relation_families)

+ 19
- 0
src/triacontagon/weights.py Zobrazit soubor

@@ -0,0 +1,19 @@
#
# Copyright (C) Stanislaw Adaszewski, 2020
# License: GPLv3
#
import torch
import numpy as np
def init_glorot(in_channels, out_channels, dtype=torch.float32):
"""Create a weight variable with Glorot & Bengio (AISTATS 2010)
initialization.
"""
init_range = np.sqrt(6.0 / (in_channels + out_channels))
initial = -init_range + 2 * init_range * \
torch.rand(( in_channels, out_channels ), dtype=dtype)
initial = initial.requires_grad_(True)
return initial

Načítá se…
Zrušit
Uložit