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

refactor: separate gradient back and weight updates

Browse Source
This commit is contained in:
Lenoctambule
2026-04-12 19:40:04 +02:00
parent d23f3a903a
commit a4334568ec
2 changed files with 33 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
src/easyvae/autoencoder.py
View File

@@ -82,9 +82,11 @@ class ClassicalAutoencoder(AAutoencoder):
self.encoder.forward(v) self.encoder.forward(v)
) )
error = out - v error = out - v
self.encoder.backprop( self.encoder.back(
self.decoder.backprop(error) self.decoder.back(error)
) )
self.encoder.backprop()
self.decoder.backprop()
return np.sum(np.abs(error)) / len(v) return np.sum(np.abs(error)) / len(v)
@interruptable @interruptable
@@ -109,7 +111,7 @@ class ClassicalAutoencoder(AAutoencoder):
error += self.train(x) error += self.train(x)
error /= len(data_set) error /= len(data_set)
derror = prev_error - error derror = prev_error - error
if derror <= 0 or abs(derror) < 1e-4: if abs(derror) < 1e-4:
no_improv += 1 no_improv += 1
else: else:
no_improv = 0 no_improv = 0
@@ -167,11 +169,14 @@ class VariationalAutoencoder(AAutoencoder):
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(v)
error = out - v error = out - v
self.encoder.backprop( self.encoder.back(
self.sampler.backprop( self.sampler.back(
self.decoder.backprop(error) self.decoder.back(error)
) )
) )
self.encoder.backprop()
self.sampler.backprop()
self.decoder.backprop()
return np.mean(error ** 2), self.sampler.DKL() return np.mean(error ** 2), self.sampler.DKL()
@interruptable @interruptable

33
src/easyvae/layers.py
View File

@@ -15,6 +15,7 @@ class NNLayer:
self.input = None self.input = None
self.output = None self.output = None
self.output_linear = None self.output_linear = None
self.error = None
self.activation_func = activation_func self.activation_func = activation_func
def __str__(self): def __str__(self):
@@ -27,15 +28,16 @@ class NNLayer:
self.output_linear self.output_linear
) )
return self.output return self.output
def back(self, error: np.ndarray) -> np.ndarray:
self.error = error * self.activation_func.d(self.output_linear)
return self.W @ self.error
def backprop(self, error: np.ndarray) -> np.ndarray: def backprop(self) -> np.ndarray:
error *= self.activation_func.d(self.output_linear) dW = np.outer(self.input, self.error) * self.lr
ret = self.W @ error dB = self.error * self.lr
dW = np.outer(self.input, error) * self.lr
dB = error * self.lr
self.W -= dW self.W -= dW
self.B -= dB self.B -= dB
return ret
class SampleLayer: class SampleLayer:
@@ -66,13 +68,17 @@ class SampleLayer:
self.eps = np.random.normal(0, 1, self.mean.shape) self.eps = np.random.normal(0, 1, self.mean.shape)
return 0.5 * self.eps * self.std + self.mean return 0.5 * self.eps * self.std + self.mean
def backprop(self, error: np.ndarray) -> np.ndarray: def back(self, error: np.ndarray) -> np.ndarray:
dmean = error + self.mean dmean = error + self.mean
dstd = error * self.eps + 0.5 * (np.exp(self.logvar) - 1) dstd = error * self.eps + 0.5 * (np.exp(self.logvar) - 1)
mean_error = self.mean_nn.backprop(dmean) mean_error = self.mean_nn.back(dmean)
logvar_error = self.std_nn.backprop(dstd * self.std) logvar_error = self.std_nn.back(dstd * self.std)
return mean_error + logvar_error return mean_error + logvar_error
def backprop(self):
self.mean_nn.backprop()
self.std_nn.backprop()
class DeepNNLayer: class DeepNNLayer:
def __init__(self, def __init__(self,
@@ -100,7 +106,12 @@ class DeepNNLayer:
v = layer.forward(v) v = layer.forward(v)
return v return v
def backprop(self, error: np.ndarray) -> np.ndarray: def back(self, error: np.ndarray):
for layer in self.layers[::-1]: for layer in self.layers[::-1]:
error = layer.backprop(error) error = layer.back(error)
return error return error
def backprop(self) -> np.ndarray:
for layer in self.layers:
layer.backprop()