lenoctambule/autoencoder
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: NoiseLayer class + keep loss across trainings

Browse Source
This commit is contained in:
Lenoctambule
2026-04-12 20:31:31 +02:00
parent a4334568ec
commit 82acc6b4f3
2 changed files with 25 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
src/easyvae/autoencoder.py
View File

@@ -1,6 +1,6 @@
import numpy as np import numpy as np
from tqdm import tqdm from tqdm import tqdm
from .layers import DeepNNLayer, SampleLayer from .layers import DeepNNLayer, SampleLayer, NoiseLayer
from .activations import ActivationFunc, Identity from .activations import ActivationFunc, Identity
from .plotters import Plotter, CAPlotter, VAEPlotter from .plotters import Plotter, CAPlotter, VAEPlotter
from .utils import interruptable from .utils import interruptable
@@ -17,13 +17,15 @@ class AAutoencoder(ABC):
encoder_layers: list[int], encoder_layers: list[int],
decoder_layers: list[int], decoder_layers: list[int],
lr: float, lr: float,
activation_func: ActivationFunc): activation_func: ActivationFunc,
noise=0):
if encoder_layers[-1] != decoder_layers[0]: if encoder_layers[-1] != decoder_layers[0]:
raise Exception( raise Exception(
f"Encoder output and decoder input don't match {encoder_layers[-1]} != {encoder_layers[0]}" # noqa f"Encoder output and decoder input don't match {encoder_layers[-1]} != {encoder_layers[0]}" # noqa
) )
self.encoder = DeepNNLayer(encoder_layers, lr, activation_func) self.encoder = DeepNNLayer(encoder_layers, lr, activation_func)
self.decoder = DeepNNLayer(decoder_layers, lr, activation_func) self.decoder = DeepNNLayer(decoder_layers, lr, activation_func)
self.noise = NoiseLayer(noise)
self.space_dim = decoder_layers[0] self.space_dim = decoder_layers[0]
self.lr = lr self.lr = lr
self.losses = [0] self.losses = [0]
@@ -78,8 +80,8 @@ class ClassicalAutoencoder(AAutoencoder):
return loss / len(data_set) return loss / len(data_set)
def train(self, v: np.ndarray): def train(self, v: np.ndarray):
out = self.decoder.forward( out, _ = self.forward(
self.encoder.forward(v) self.noise.forward(v)
) )
error = out - v error = out - v
self.encoder.back( self.encoder.back(
@@ -96,7 +98,8 @@ class ClassicalAutoencoder(AAutoencoder):
patience: int, patience: int,
display_loss: bool = False) -> list[float]: display_loss: bool = False) -> list[float]:
plotter = self.plotter_cls(self) if display_loss else Plotter(self) plotter = self.plotter_cls(self) if display_loss else Plotter(self)
self.losses = [self.loss(data_set)] if len(self.losses) == 0:
self.losses = [self.loss(data_set)]
epoch = 0 epoch = 0
no_improv = 0 no_improv = 0
prev_error = self.losses[0] prev_error = self.losses[0]
@@ -167,7 +170,9 @@ class VariationalAutoencoder(AAutoencoder):
return recon_loss, kl_loss return recon_loss, kl_loss
def train(self, v: np.ndarray) -> tuple[float, float]: def train(self, v: np.ndarray) -> tuple[float, float]:
out, _ = self.forward(v) out, _ = self.forward(
self.noise.forward(v)
)
error = out - v error = out - v
self.encoder.back( self.encoder.back(
self.sampler.back( self.sampler.back(
@@ -186,9 +191,10 @@ class VariationalAutoencoder(AAutoencoder):
patience: int, patience: int,
display_loss: bool = False) -> list[float]: display_loss: bool = False) -> list[float]:
plotter = self.plotter_cls(self) if display_loss else Plotter(self) plotter = self.plotter_cls(self) if display_loss else Plotter(self)
recon_0, kl_0 = self.loss(data_set) if len(self.recon_losses) == 0:
self.recon_losses = [recon_0] recon_0, kl_0 = self.loss(data_set)
self.KL_losses = [kl_0] self.recon_losses = [recon_0]
self.KL_losses = [kl_0]
epoch = 0 epoch = 0
no_improv = 0 no_improv = 0
prev_loss = self.recon_losses[0] + self.KL_losses[0] prev_loss = self.recon_losses[0] + self.KL_losses[0]

11
src/easyvae/layers.py
View File

@@ -28,7 +28,7 @@ class NNLayer:
self.output_linear self.output_linear
) )
return self.output return self.output
def back(self, error: np.ndarray) -> np.ndarray: def back(self, error: np.ndarray) -> np.ndarray:
self.error = error * self.activation_func.d(self.output_linear) self.error = error * self.activation_func.d(self.output_linear)
return self.W @ self.error return self.W @ self.error
@@ -115,3 +115,12 @@ class DeepNNLayer:
for layer in self.layers: for layer in self.layers:
layer.backprop() layer.backprop()
class NoiseLayer:
def __init__(self, amount=0.1):
self.amount = amount
def forward(self, v: np.ndarray):
if self.amount == 0:
return v
return v + np.random.normal(0, self.amount, v.shape)