PyDLLib 1.0.2__tar.gz → 1.0.3__tar.gz

{pydllib-1.0.2 → pydllib-1.0.3}/DLL/DeepLearning/Layers/_Conv2D.py

@@ -46,12 +46,13 @@ class Conv2D(BaseLayer):
         self.input_depth = input_depth
         self.output_shape = (self.output_depth, input_height - self.kernel_size + 1, input_width - self.kernel_size + 1)
         self.kernels_shape = (self.output_depth, input_depth, self.kernel_size, self.kernel_size)
-        self.nparams = np.prod(self.kernels_shape) + np.prod(self.output_shape)
+        self.nparams = np.prod(self.kernels_shape) + self.output_depth
         
         super().initialise_layer(input_shape, data_type, device)
 
         self.kernels = self.initialiser.initialise(self.kernels_shape, data_type=self.data_type, device=self.device)
-        self.biases = torch.zeros(self.output_shape)
+        # self.biases = torch.zeros(self.output_shape)
+        self.biases = torch.zeros((self.output_depth,), dtype=self.data_type, device=self.device)
         
         if self.activation:
             self.activation.initialise_layer(self.output_shape, self.data_type, self.device)
@@ -87,11 +88,12 @@ class Conv2D(BaseLayer):
 
         batch_size = input.shape[0]
         self.input = input
-        self.output = self.biases.clone().repeat(batch_size, 1, 1, 1)
+        self.output = torch.zeros((batch_size, *self.output_shape), dtype=self.data_type, device=self.device)
         for i in range(self.output_depth):
             for j in range(self.input_depth):
                 conv_output = F.conv2d(input[:, j:j+1, :, :], self.kernels[i:i+1, j:j+1, :, :], padding="valid")
                 self.output[:, i, :, :] += conv_output[:, 0, :, :]
+        self.output += self.biases.view(1, -1, 1, 1)
                 
         if self.normalisation: self.output = self.normalisation.forward(self.output, training=training)
         if self.activation: self.output = self.activation.forward(self.output)
@@ -122,7 +124,7 @@ class Conv2D(BaseLayer):
                 kernel_gradient[i, j] = F.conv2d(self.input[:, j:j+1, :, :], dCdy[:, i:i+1, :, :], padding="valid")[0, 0, :, :]
                 dCdx[:, j] += F.conv2d(dCdy[:, i:i+1, :, :], torch.flip(self.kernels[i:i+1, j:j+1, :, :], dims=(2, 3)), padding=[self.kernel_size - 1, self.kernel_size - 1])[0, 0, :, :]
                 
-        self.biases.grad += dCdy.mean(dim=0)
+        self.biases.grad += dCdy.mean(dim=(0, 2, 3))
         self.kernels.grad += kernel_gradient / batch_size
         return dCdx
     

{pydllib-1.0.2 → pydllib-1.0.3}/DLL/MachineLearning/SupervisedLearning/GaussianProcesses/_GaussianProcessClassifier.py

@@ -228,6 +228,8 @@ class GaussianProcessClassifier:
         history = {"log marginal likelihood": torch.zeros(floor(epochs / callback_frequency))}
 
         for epoch in range(epochs):
+            optimiser.zero_grad()
+
             # calculate the derivatives
             self.covariance_function.update(self._derivative, self.X)
 

{pydllib-1.0.2 → pydllib-1.0.3}/DLL/MachineLearning/SupervisedLearning/GaussianProcesses/_GaussianProcessRegressor.py

@@ -157,6 +157,8 @@ class GaussianProcessRegressor:
         history = {"log marginal likelihood": torch.zeros(floor(epochs / callback_frequency))}
 
         for epoch in range(epochs):
+            optimiser.zero_grad()
+
             # calculate the derivatives
             self.covariance_function.update(self._derivative, self.X)
 

{pydllib-1.0.2 → pydllib-1.0.3}/DLL/MachineLearning/SupervisedLearning/Kernels.py

@@ -185,8 +185,8 @@ class RBF(_Base):
         derivative_covariance_sigma = self.derivative_sigma(X, X)
         derivative_covariance_corr_len = self.derivative_corr_len(X, X)
         
-        self.sigma.grad = derivative_function(derivative_covariance_sigma).to(dtype=self.sigma.dtype)
-        self.correlation_length.grad = derivative_function(derivative_covariance_corr_len).to(dtype=self.correlation_length.dtype)
+        self.sigma.grad += derivative_function(derivative_covariance_sigma).to(dtype=self.sigma.dtype)
+        self.correlation_length.grad += derivative_function(derivative_covariance_corr_len).to(dtype=self.correlation_length.dtype)
 
     def parameters(self):
         """
@@ -311,8 +311,8 @@ class Linear(_Base):
         derivative_covariance_sigma = self.derivative_sigma(X, X)
         derivative_covariance_sigma_bias = self.derivative_sigma_bias(X, X)
         
-        self.sigma.grad = derivative_function(derivative_covariance_sigma).to(dtype=self.sigma.dtype)
-        self.sigma_bias.grad = derivative_function(derivative_covariance_sigma_bias).to(dtype=self.sigma_bias.dtype)
+        self.sigma.grad += derivative_function(derivative_covariance_sigma).to(dtype=self.sigma.dtype)
+        self.sigma_bias.grad += derivative_function(derivative_covariance_sigma_bias).to(dtype=self.sigma_bias.dtype)
 
     def parameters(self):
         """
@@ -405,7 +405,7 @@ class WhiteGaussian(_Base):
         """
         derivative_covariance_sigma = self.derivative_sigma(X, X)
         
-        self.sigma.grad = derivative_function(derivative_covariance_sigma).to(dtype=self.sigma.dtype)
+        self.sigma.grad += derivative_function(derivative_covariance_sigma).to(dtype=self.sigma.dtype)
 
     def parameters(self):
         """
@@ -550,9 +550,9 @@ class Periodic(_Base):
         derivative_covariance_corr_len = self.derivative_corr_len(X, X)
         derivative_covariance_period = self.derivative_period(X, X)
 
-        self.sigma.grad = derivative_function(derivative_covariance_sigma).to(dtype=self.sigma.dtype)
-        self.correlation_length.grad = derivative_function(derivative_covariance_corr_len).to(dtype=self.correlation_length.dtype)
-        self.period.grad = derivative_function(derivative_covariance_period).to(dtype=self.period.dtype)
+        self.sigma.grad += derivative_function(derivative_covariance_sigma).to(dtype=self.sigma.dtype)
+        self.correlation_length.grad += derivative_function(derivative_covariance_corr_len).to(dtype=self.correlation_length.dtype)
+        self.period.grad += derivative_function(derivative_covariance_period).to(dtype=self.period.dtype)
 
     def parameters(self):
         """
@@ -697,9 +697,9 @@ class RationalQuadratic(_Base):
         derivative_covariance_corr_len = self.derivative_corr_len(X, X)
         derivative_covariance_alpha = self.derivative_alpha(X, X)
 
-        self.sigma.grad = derivative_function(derivative_covariance_sigma).to(dtype=self.sigma.dtype)
-        self.correlation_length.grad = derivative_function(derivative_covariance_corr_len).to(dtype=self.correlation_length.dtype)
-        self.alpha.grad = derivative_function(derivative_covariance_alpha).to(dtype=self.alpha.dtype)
+        self.sigma.grad += derivative_function(derivative_covariance_sigma).to(dtype=self.sigma.dtype)
+        self.correlation_length.grad += derivative_function(derivative_covariance_corr_len).to(dtype=self.correlation_length.dtype)
+        self.alpha.grad += derivative_function(derivative_covariance_alpha).to(dtype=self.alpha.dtype)
 
     def parameters(self):
         """
@@ -838,8 +838,8 @@ class Matern(_Base):
         derivative_covariance_sigma = self.derivative_sigma(X, X)
         derivative_covariance_corr_len = self.derivative_corr_len(X, X)
 
-        self.sigma.grad = derivative_function(derivative_covariance_sigma).to(dtype=self.sigma.dtype)
-        self.correlation_length.grad = derivative_function(derivative_covariance_corr_len).to(dtype=self.correlation_length.dtype)
+        self.sigma.grad += derivative_function(derivative_covariance_sigma).to(dtype=self.sigma.dtype)
+        self.correlation_length.grad += derivative_function(derivative_covariance_corr_len).to(dtype=self.correlation_length.dtype)
 
     def parameters(self):
         """

{pydllib-1.0.2 → pydllib-1.0.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: PyDLLib
-Version: 1.0.2
+Version: 1.0.3
 Summary: A deep and machine learning library. Made for learning purposes and ease of understanding.
 Author-email: Aatu Selkee <aatu.selkee@gmail.com>
 Project-URL: Homepage, https://naapeli.github.io/deep-learning-library/index.html

{pydllib-1.0.2 → pydllib-1.0.3}/PyDLLib.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: PyDLLib
-Version: 1.0.2
+Version: 1.0.3
 Summary: A deep and machine learning library. Made for learning purposes and ease of understanding.
 Author-email: Aatu Selkee <aatu.selkee@gmail.com>
 Project-URL: Homepage, https://naapeli.github.io/deep-learning-library/index.html

{pydllib-1.0.2 → pydllib-1.0.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "PyDLLib"
-version = "1.0.2"
+version = "1.0.3"
 authors = [
     {name = "Aatu Selkee", email = "aatu.selkee@gmail.com"},
 ]