julius/prototorch_models
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge branch 'dev' into main

Browse Source
This commit is contained in:
Alexander Engelsberger
2022-01-11 18:29:32 +01:00
parent aaa3c51e0a 1a17193b35
commit 1a0e697b27
27 changed files with 537 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
prototorch/models/__init__.py
View File

@@ -8,17 +8,34 @@ from .glvq import (
GLVQ21,
GMLVQ,
GRLVQ,
GTLVQ,
LGMLVQ,
LVQMLN,
ImageGLVQ,
ImageGMLVQ,
ImageGTLVQ,
SiameseGLVQ,
SiameseGMLVQ,
SiameseGTLVQ,
)
from .knn import KNN
from .lvq import LVQ1, LVQ21, MedianLVQ
from .probabilistic import CELVQ, PLVQ, RSLVQ, SLVQ
from .unsupervised import GrowingNeuralGas, HeskesSOM, KohonenSOM, NeuralGas
from .lvq import (
LVQ1,
LVQ21,
MedianLVQ,
)
from .probabilistic import (
CELVQ,
PLVQ,
RSLVQ,
SLVQ,
)
from .unsupervised import (
GrowingNeuralGas,
HeskesSOM,
KohonenSOM,
NeuralGas,
)
from .vis import *
__version__ = "0.4.0"

8
prototorch/models/abstract.py
View File

@@ -14,6 +14,7 @@ from ..nn.wrappers import LambdaLayer
class ProtoTorchBolt(pl.LightningModule):
"""All ProtoTorch models are ProtoTorch Bolts."""
def __init__(self, hparams, **kwargs):
super().__init__()
@@ -52,6 +53,7 @@ class ProtoTorchBolt(pl.LightningModule):
class PrototypeModel(ProtoTorchBolt):
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
@@ -81,6 +83,7 @@ class PrototypeModel(ProtoTorchBolt):
class UnsupervisedPrototypeModel(PrototypeModel):
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
@@ -103,6 +106,7 @@ class UnsupervisedPrototypeModel(PrototypeModel):
class SupervisedPrototypeModel(PrototypeModel):
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
@@ -135,7 +139,7 @@ class SupervisedPrototypeModel(PrototypeModel):
distances = self.compute_distances(x)
_, plabels = self.proto_layer()
winning = stratified_min_pooling(distances, plabels)
y_pred = torch.nn.functional.softmin(winning)
y_pred = torch.nn.functional.softmin(winning, dim=1)
return y_pred
def predict_from_distances(self, distances):
@@ -178,6 +182,7 @@ class ProtoTorchMixin(object):
class NonGradientMixin(ProtoTorchMixin):
"""Mixin for custom non-gradient optimization."""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.automatic_optimization = False
@@ -188,6 +193,7 @@ class NonGradientMixin(ProtoTorchMixin):
class ImagePrototypesMixin(ProtoTorchMixin):
"""Mixin for models with image prototypes."""
def on_train_batch_end(self, outputs, batch, batch_idx, dataloader_idx):
"""Constrain the components to the range [0, 1] by clamping after updates."""
self.proto_layer.components.data.clamp_(0.0, 1.0)

3
prototorch/models/callbacks.py
View File

@@ -11,6 +11,7 @@ from .extras import ConnectionTopology
class PruneLoserPrototypes(pl.Callback):
def __init__(self,
threshold=0.01,
idle_epochs=10,
@@ -67,6 +68,7 @@ class PruneLoserPrototypes(pl.Callback):
class PrototypeConvergence(pl.Callback):
def __init__(self, min_delta=0.01, idle_epochs=10, verbose=False):
self.min_delta = min_delta
self.idle_epochs = idle_epochs # epochs to wait
@@ -89,6 +91,7 @@ class GNGCallback(pl.Callback):
Based on "A Growing Neural Gas Network Learns Topologies" by Bernd Fritzke.
"""
def __init__(self, reduction=0.1, freq=10):
self.reduction = reduction
self.freq = freq

1
prototorch/models/cbc.py
View File

@@ -13,6 +13,7 @@ from .glvq import SiameseGLVQ
class CBC(SiameseGLVQ):
"""Classification-By-Components."""
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)

41
prototorch/models/extras.py
View File

@@ -15,7 +15,46 @@ def rank_scaled_gaussian(distances, lambd):
return torch.exp(-torch.exp(-ranks / lambd) * distances)
def orthogonalization(tensors):
"""Orthogonalization via polar decomposition """
u, _, v = torch.svd(tensors, compute_uv=True)
u_shape = tuple(list(u.shape))
v_shape = tuple(list(v.shape))
# reshape to (num x N x M)
u = torch.reshape(u, (-1, u_shape[-2], u_shape[-1]))
v = torch.reshape(v, (-1, v_shape[-2], v_shape[-1]))
out = u @ v.permute([0, 2, 1])
out = torch.reshape(out, u_shape[:-1] + (v_shape[-2], ))
return out
def ltangent_distance(x, y, omegas):
r"""Localized Tangent distance.
Compute Orthogonal Complement: math:`\bm P_k = \bm I - \Omega_k \Omega_k^T`
Compute Tangent Distance: math:`{\| \bm P \bm x - \bm P_k \bm y_k \|}_2`
:param `torch.tensor` omegas: Three dimensional matrix
:rtype: `torch.tensor`
"""
x, y = [arr.view(arr.size(0), -1) for arr in (x, y)]
p = torch.eye(omegas.shape[-2], device=omegas.device) - torch.bmm(
omegas, omegas.permute([0, 2, 1]))
projected_x = x @ p
projected_y = torch.diagonal(y @ p).T
expanded_y = torch.unsqueeze(projected_y, dim=1)
batchwise_difference = expanded_y - projected_x
differences_squared = batchwise_difference**2
distances = torch.sqrt(torch.sum(differences_squared, dim=2))
distances = distances.permute(1, 0)
return distances
class GaussianPrior(torch.nn.Module):
def __init__(self, variance):
super().__init__()
self.variance = variance
@@ -25,6 +64,7 @@ class GaussianPrior(torch.nn.Module):
class RankScaledGaussianPrior(torch.nn.Module):
def __init__(self, lambd):
super().__init__()
self.lambd = lambd
@@ -34,6 +74,7 @@ class RankScaledGaussianPrior(torch.nn.Module):
class ConnectionTopology(torch.nn.Module):
def __init__(self, agelimit, num_prototypes):
super().__init__()
self.agelimit = agelimit

76
prototorch/models/glvq.py
View File

@@ -4,16 +4,26 @@ import torch
from torch.nn.parameter import Parameter
from ..core.competitions import wtac
from ..core.distances import lomega_distance, omega_distance, squared_euclidean_distance
from ..core.distances import (
lomega_distance,
omega_distance,
squared_euclidean_distance,
)
from ..core.initializers import EyeTransformInitializer
from ..core.losses import GLVQLoss, lvq1_loss, lvq21_loss
from ..core.losses import (
GLVQLoss,
lvq1_loss,
lvq21_loss,
)
from ..core.transforms import LinearTransform
from ..nn.wrappers import LambdaLayer, LossLayer
from .abstract import ImagePrototypesMixin, SupervisedPrototypeModel
from .extras import ltangent_distance, orthogonalization
class GLVQ(SupervisedPrototypeModel):
"""Generalized Learning Vector Quantization."""
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
@@ -98,6 +108,7 @@ class SiameseGLVQ(GLVQ):
transformation pipeline are only learned from the inputs.
"""
def __init__(self,
hparams,
backbone=torch.nn.Identity(),
@@ -164,6 +175,7 @@ class LVQMLN(SiameseGLVQ):
rather in the embedding space.
"""
def compute_distances(self, x):
latent_protos, _ = self.proto_layer()
latent_x = self.backbone(x)
@@ -179,6 +191,7 @@ class GRLVQ(SiameseGLVQ):
TODO Make a RelevanceLayer. `bb_lr` is ignored otherwise.
"""
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
@@ -204,6 +217,7 @@ class SiameseGMLVQ(SiameseGLVQ):
Implemented as a Siamese network with a linear transformation backbone.
"""
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
@@ -234,6 +248,7 @@ class GMLVQ(GLVQ):
function. This makes it easier to implement a localized variant.
"""
def __init__(self, hparams, **kwargs):
distance_fn = kwargs.pop("distance_fn", omega_distance)
super().__init__(hparams, distance_fn=distance_fn, **kwargs)
@@ -268,6 +283,7 @@ class GMLVQ(GLVQ):
class LGMLVQ(GMLVQ):
"""Localized and Generalized Matrix Learning Vector Quantization."""
def __init__(self, hparams, **kwargs):
distance_fn = kwargs.pop("distance_fn", lomega_distance)
super().__init__(hparams, distance_fn=distance_fn, **kwargs)
@@ -282,8 +298,48 @@ class LGMLVQ(GMLVQ):
self.register_parameter("_omega", Parameter(omega))
class GTLVQ(LGMLVQ):
"""Localized and Generalized Tangent Learning Vector Quantization."""
def __init__(self, hparams, **kwargs):
distance_fn = kwargs.pop("distance_fn", ltangent_distance)
super().__init__(hparams, distance_fn=distance_fn, **kwargs)
omega_initializer = kwargs.get("omega_initializer")
if omega_initializer is not None:
subspace = omega_initializer.generate(self.hparams.input_dim,
self.hparams.latent_dim)
omega = torch.repeat_interleave(subspace.unsqueeze(0),
self.num_prototypes,
dim=0)
else:
omega = torch.rand(
self.num_prototypes,
self.hparams.input_dim,
self.hparams.latent_dim,
device=self.device,
)
# Re-register `_omega` to override the one from the super class.
self.register_parameter("_omega", Parameter(omega))
def on_train_batch_end(self, outputs, batch, batch_idx, dataloader_idx):
with torch.no_grad():
self._omega.copy_(orthogonalization(self._omega))
class SiameseGTLVQ(SiameseGLVQ, GTLVQ):
"""Generalized Tangent Learning Vector Quantization.
Implemented as a Siamese network with a linear transformation backbone.
"""
class GLVQ1(GLVQ):
"""Generalized Learning Vector Quantization 1."""
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
self.loss = LossLayer(lvq1_loss)
@@ -292,6 +348,7 @@ class GLVQ1(GLVQ):
class GLVQ21(GLVQ):
"""Generalized Learning Vector Quantization 2.1."""
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
self.loss = LossLayer(lvq21_loss)
@@ -314,3 +371,18 @@ class ImageGMLVQ(ImagePrototypesMixin, GMLVQ):
after updates.
"""
class ImageGTLVQ(ImagePrototypesMixin, GTLVQ):
"""GTLVQ for training on image data.
GTLVQ model that constrains the prototypes to the range [0, 1] by clamping
after updates.
"""
def on_train_batch_end(self, outputs, batch, batch_idx, dataloader_idx):
"""Constrain the components to the range [0, 1] by clamping after updates."""
self.proto_layer.components.data.clamp_(0.0, 1.0)
with torch.no_grad():
self._omega.copy_(orthogonalization(self._omega))

6
prototorch/models/knn.py
View File

@@ -4,13 +4,17 @@ import warnings
from ..core.competitions import KNNC
from ..core.components import LabeledComponents
from ..core.initializers import LiteralCompInitializer, LiteralLabelsInitializer
from ..core.initializers import (
LiteralCompInitializer,
LiteralLabelsInitializer,
)
from ..utils.utils import parse_data_arg
from .abstract import SupervisedPrototypeModel
class KNN(SupervisedPrototypeModel):
"""K-Nearest-Neighbors classification algorithm."""
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)

3
prototorch/models/lvq.py
View File

@@ -9,6 +9,7 @@ from .glvq import GLVQ
class LVQ1(NonGradientMixin, GLVQ):
"""Learning Vector Quantization 1."""
def training_step(self, train_batch, batch_idx, optimizer_idx=None):
protos, plables = self.proto_layer()
x, y = train_batch
@@ -38,6 +39,7 @@ class LVQ1(NonGradientMixin, GLVQ):
class LVQ21(NonGradientMixin, GLVQ):
"""Learning Vector Quantization 2.1."""
def training_step(self, train_batch, batch_idx, optimizer_idx=None):
protos, plabels = self.proto_layer()
@@ -70,6 +72,7 @@ class MedianLVQ(NonGradientMixin, GLVQ):
# TODO Avoid computing distances over and over
"""
def __init__(self, hparams, verbose=True, **kwargs):
self.verbose = verbose
super().__init__(hparams, **kwargs)

5
prototorch/models/probabilistic.py
View File

@@ -11,6 +11,7 @@ from .glvq import GLVQ, SiameseGMLVQ
class CELVQ(GLVQ):
"""Cross-Entropy Learning Vector Quantization."""
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
@@ -29,6 +30,7 @@ class CELVQ(GLVQ):
class ProbabilisticLVQ(GLVQ):
def __init__(self, hparams, rejection_confidence=0.0, **kwargs):
super().__init__(hparams, **kwargs)
@@ -62,6 +64,7 @@ class ProbabilisticLVQ(GLVQ):
class SLVQ(ProbabilisticLVQ):
"""Soft Learning Vector Quantization."""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.loss = LossLayer(nllr_loss)
@@ -70,6 +73,7 @@ class SLVQ(ProbabilisticLVQ):
class RSLVQ(ProbabilisticLVQ):
"""Robust Soft Learning Vector Quantization."""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.loss = LossLayer(rslvq_loss)
@@ -81,6 +85,7 @@ class PLVQ(ProbabilisticLVQ, SiameseGMLVQ):
TODO: Use Backbone LVQ instead
"""
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.conditional_distribution = RankScaledGaussianPrior(

4
prototorch/models/unsupervised.py
View File

@@ -18,6 +18,7 @@ class KohonenSOM(NonGradientMixin, UnsupervisedPrototypeModel):
TODO Allow non-2D grids
"""
def __init__(self, hparams, **kwargs):
h, w = hparams.get("shape")
# Ignore `num_prototypes`
@@ -69,6 +70,7 @@ class KohonenSOM(NonGradientMixin, UnsupervisedPrototypeModel):
class HeskesSOM(UnsupervisedPrototypeModel):
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
@@ -78,6 +80,7 @@ class HeskesSOM(UnsupervisedPrototypeModel):
class NeuralGas(UnsupervisedPrototypeModel):
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
@@ -110,6 +113,7 @@ class NeuralGas(UnsupervisedPrototypeModel):
class GrowingNeuralGas(NeuralGas):
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)

7
prototorch/models/vis.py
View File

@@ -11,6 +11,7 @@ from ..utils.utils import mesh2d
class Vis2DAbstract(pl.Callback):
def __init__(self,
data,
title="Prototype Visualization",
@@ -118,6 +119,7 @@ class Vis2DAbstract(pl.Callback):
class VisGLVQ2D(Vis2DAbstract):
def on_epoch_end(self, trainer, pl_module):
if not self.precheck(trainer):
return True
@@ -141,6 +143,7 @@ class VisGLVQ2D(Vis2DAbstract):
class VisSiameseGLVQ2D(Vis2DAbstract):
def __init__(self, *args, map_protos=True, **kwargs):
super().__init__(*args, **kwargs)
self.map_protos = map_protos
@@ -179,6 +182,7 @@ class VisSiameseGLVQ2D(Vis2DAbstract):
class VisGMLVQ2D(Vis2DAbstract):
def __init__(self, *args, ev_proj=True, **kwargs):
super().__init__(*args, **kwargs)
self.ev_proj = ev_proj
@@ -212,6 +216,7 @@ class VisGMLVQ2D(Vis2DAbstract):
class VisCBC2D(Vis2DAbstract):
def on_epoch_end(self, trainer, pl_module):
if not self.precheck(trainer):
return True
@@ -235,6 +240,7 @@ class VisCBC2D(Vis2DAbstract):
class VisNG2D(Vis2DAbstract):
def on_epoch_end(self, trainer, pl_module):
if not self.precheck(trainer):
return True
@@ -262,6 +268,7 @@ class VisNG2D(Vis2DAbstract):
class VisImgComp(Vis2DAbstract):
def __init__(self,
*args,
random_data=0,