prototorch_models/prototorch/models/unsupervised.py
2021-06-14 20:08:08 +02:00

147 lines
4.8 KiB
Python

"""Unsupervised prototype learning algorithms."""
import numpy as np
import torch
from ..core.competitions import wtac
from ..core.distances import squared_euclidean_distance
from ..core.losses import NeuralGasEnergy
from ..nn.wrappers import LambdaLayer
from .abstract import NonGradientMixin, UnsupervisedPrototypeModel
from .callbacks import GNGCallback
from .extras import ConnectionTopology
class KohonenSOM(NonGradientMixin, UnsupervisedPrototypeModel):
"""Kohonen Self-Organizing-Map.
TODO Allow non-2D grids
"""
def __init__(self, hparams, **kwargs):
h, w = hparams.get("shape")
# Ignore `num_prototypes`
hparams["num_prototypes"] = h * w
distance_fn = kwargs.pop("distance_fn", squared_euclidean_distance)
super().__init__(hparams, distance_fn=distance_fn, **kwargs)
# Hyperparameters
self.save_hyperparameters(hparams)
# Default hparams
self.hparams.setdefault("alpha", 0.3)
self.hparams.setdefault("sigma", max(h, w) / 2.0)
# Additional parameters
x, y = torch.arange(h), torch.arange(w)
grid = torch.stack(torch.meshgrid(x, y), dim=-1)
self.register_buffer("_grid", grid)
self._sigma = self.hparams.sigma
self._lr = self.hparams.lr
def predict_from_distances(self, distances):
grid = self._grid.view(-1, 2)
wp = wtac(distances, grid)
return wp
def training_step(self, train_batch, batch_idx):
# x = train_batch
# TODO Check if the batch has labels
x = train_batch[0]
d = self.compute_distances(x)
wp = self.predict_from_distances(d)
grid = self._grid.view(-1, 2)
gd = squared_euclidean_distance(wp, grid)
nh = torch.exp(-gd / self._sigma**2)
protos = self.proto_layer.components
diff = x.unsqueeze(dim=1) - protos
delta = self._lr * self.hparams.alpha * nh.unsqueeze(-1) * diff
updated_protos = protos + delta.sum(dim=0)
self.proto_layer.load_state_dict({"_components": updated_protos},
strict=False)
def training_epoch_end(self, training_step_outputs):
self._sigma = self.hparams.sigma * np.exp(
-self.current_epoch / self.trainer.max_epochs)
def extra_repr(self):
return f"(grid): (shape: {tuple(self._grid.shape)})"
class HeskesSOM(UnsupervisedPrototypeModel):
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
def training_step(self, train_batch, batch_idx):
# TODO Implement me!
raise NotImplementedError()
class NeuralGas(UnsupervisedPrototypeModel):
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
# Hyperparameters
self.save_hyperparameters(hparams)
# Default hparams
self.hparams.setdefault("agelimit", 10)
self.hparams.setdefault("lm", 1)
self.energy_layer = NeuralGasEnergy(lm=self.hparams.lm)
self.topology_layer = ConnectionTopology(
agelimit=self.hparams.agelimit,
num_prototypes=self.hparams.num_prototypes,
)
def training_step(self, train_batch, batch_idx):
# x = train_batch
# TODO Check if the batch has labels
x = train_batch[0]
d = self.compute_distances(x)
loss, _ = self.energy_layer(d)
self.topology_layer(d)
self.log("loss", loss)
return loss
# def training_epoch_end(self, training_step_outputs):
# print(f"{self.trainer.lr_schedulers}")
# print(f"{self.trainer.lr_schedulers[0]['scheduler'].optimizer}")
class GrowingNeuralGas(NeuralGas):
def __init__(self, hparams, **kwargs):
super().__init__(hparams, **kwargs)
# Defaults
self.hparams.setdefault("step_reduction", 0.5)
self.hparams.setdefault("insert_reduction", 0.1)
self.hparams.setdefault("insert_freq", 10)
errors = torch.zeros(self.hparams.num_prototypes, device=self.device)
self.register_buffer("errors", errors)
def training_step(self, train_batch, _batch_idx):
# x = train_batch
# TODO Check if the batch has labels
x = train_batch[0]
d = self.compute_distances(x)
loss, order = self.energy_layer(d)
winner = order[:, 0]
mask = torch.zeros_like(d)
mask[torch.arange(len(mask)), winner] = 1.0
dp = d * mask
self.errors += torch.sum(dp * dp, dim=0)
self.errors *= self.hparams.step_reduction
self.topology_layer(d)
self.log("loss", loss)
return loss
def configure_callbacks(self):
return [
GNGCallback(reduction=self.hparams.insert_reduction,
freq=self.hparams.insert_freq)
]