NeuralEngine 0.2.3__tar.gz → 0.2.4__tar.gz

{neuralengine-0.2.3 → neuralengine-0.2.4/NeuralEngine.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: NeuralEngine
-Version: 0.2.3
+Version: 0.2.4
 Summary: A framework/library for building and training neural networks.
 Home-page: https://github.com/Prajjwal2404/NeuralEngine
 Author: Prajjwal Pratap Shah
@@ -202,6 +202,7 @@ NeuralEngine offers the following core capabilities:
 - `ne.Embedding(embed_size, vocab_size, timesteps=None)`: Embedding layer for mapping indices to dense vectors, with optional positional encoding.
 - `ne.LayerNorm(num_feat, eps=1e-7)`: Layer normalization for stabilizing training.
 - `ne.Dropout(prob=0.5)`: Dropout regularization for reducing overfitting.
+- `ne.Layer.freezed = True/False`: Freeze or unfreeze layer parameters during training.
 - All layers inherit from a common base and support extensibility for custom architectures.
 
 ### Activations

{neuralengine-0.2.3/NeuralEngine.egg-info → neuralengine-0.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: NeuralEngine
-Version: 0.2.3
+Version: 0.2.4
 Summary: A framework/library for building and training neural networks.
 Home-page: https://github.com/Prajjwal2404/NeuralEngine
 Author: Prajjwal Pratap Shah
@@ -202,6 +202,7 @@ NeuralEngine offers the following core capabilities:
 - `ne.Embedding(embed_size, vocab_size, timesteps=None)`: Embedding layer for mapping indices to dense vectors, with optional positional encoding.
 - `ne.LayerNorm(num_feat, eps=1e-7)`: Layer normalization for stabilizing training.
 - `ne.Dropout(prob=0.5)`: Dropout regularization for reducing overfitting.
+- `ne.Layer.freezed = True/False`: Freeze or unfreeze layer parameters during training.
 - All layers inherit from a common base and support extensibility for custom architectures.
 
 ### Activations

{neuralengine-0.2.3 → neuralengine-0.2.4}/README.md

@@ -154,6 +154,7 @@ NeuralEngine offers the following core capabilities:
 - `ne.Embedding(embed_size, vocab_size, timesteps=None)`: Embedding layer for mapping indices to dense vectors, with optional positional encoding.
 - `ne.LayerNorm(num_feat, eps=1e-7)`: Layer normalization for stabilizing training.
 - `ne.Dropout(prob=0.5)`: Dropout regularization for reducing overfitting.
+- `ne.Layer.freezed = True/False`: Freeze or unfreeze layer parameters during training.
 - All layers inherit from a common base and support extensibility for custom architectures.
 
 ### Activations

{neuralengine-0.2.3 → neuralengine-0.2.4}/neuralengine/nn/layers.py

@@ -14,6 +14,7 @@ class Layer:
     def __init__(self):
         self._mode: Mode = Mode.TRAIN
         self._in_size: tuple | None = None
+        self._freezed: bool = False
 
     def __call__(self, x, *args, **kwargs) -> Tensor:
         """
@@ -61,6 +62,20 @@ class Layer:
         """Performs the forward pass of the layer. To be implemented by subclasses."""
         raise NotImplementedError("forward() must be implemented in subclasses")
 
+    @property
+    def freezed(self) -> bool:
+        """Whether the layer's parameters are frozen (not trainable)."""
+        return self._freezed
+    
+    @freezed.setter
+    def freezed(self, freeze: bool) -> None:
+        self._freezed = freeze
+        for _, attr in self.__dict__.items():
+            if isinstance(attr, Layer):
+                attr.freezed = freeze
+            elif isinstance(attr, Tensor):
+                attr.requires_grad = not freeze
+
     def parameters(self):
         """Yields all trainable parameters for the layer."""
         for _, attr in self.__dict__.items():
@@ -156,7 +171,7 @@ class LSTM(Layer):
         if self.attention:
             self.attention.in_size = self.enc_size if self.enc_size else self.out_size
 
-    def forward(self, x: Tensor, c: Tensor = None, h: Tensor = None, enc_seq: Tensor = None):
+    def forward(self, x: Tensor, c: Tensor = None, h: Tensor = None, enc_seq: Tensor = None) -> list[Tensor]:
         # LSTM forward pass
         c = c if c else zeros((x.shape[0], self.out_size))
         h = h if h else zeros((x.shape[0], self.out_size + self.ctx_size))
@@ -310,8 +325,8 @@ class LayerNorm(Layer):
         super().__init__()
         self.eps = eps
         if num_feat:
-            self.gamma = ones((1, *num_feat), requires_grad=True)
-            self.beta = zeros((1, *num_feat), requires_grad=True)
+            self.gamma = ones(num_feat, requires_grad=True)
+            self.beta = zeros(num_feat, requires_grad=True)
 
     def forward(self, x: Tensor) -> Tensor:
         # z = (x - μ) / (σ + ε)

{neuralengine-0.2.3 → neuralengine-0.2.4}/neuralengine/tensor.py

@@ -2,7 +2,7 @@ import neuralengine.config as cf
 from itertools import accumulate
 
 
-grad_enabled: bool = True
+autograd_enabled: bool = True
 
 class Tensor:
     """Core Tensor class for autograd and computation."""
@@ -12,11 +12,11 @@ class Tensor:
         @param requires_grad: Whether to track gradients for this tensor.
         @param dtype: Data type of the tensor.
         """
-        global grad_enabled
+        global autograd_enabled
         self.data = array(data, dtype=dtype)
-        self.requires_grad = requires_grad if grad_enabled else False
+        self.requires_grad = requires_grad if autograd_enabled else False
         self.shape = self.data.shape
-        self.dtype = dtype.__name__
+        self.dtype = dtype
         self._operation = _operation
         if self.requires_grad:
             self.grad = cf.nu.zeros_like(self.data)
@@ -33,7 +33,7 @@ class Tensor:
 
     def __repr__(self) -> str:
         """String representation of the tensor."""
-        return f"Tensor(data={self.data}, dtype={self.dtype}, requires_grad={self.requires_grad})"
+        return f"Tensor(data={self.data}, dtype={self.dtype.__name__}, requires_grad={self.requires_grad})"
 
     def __add__(self, other):
         """New = A + B"""
@@ -742,13 +742,13 @@ class SetSlice:
 class NoGrad:
     """Context manager to disable gradient tracking."""
     def __enter__(self):
-        global grad_enabled
-        self.prev = grad_enabled
-        grad_enabled = False
+        global autograd_enabled
+        self.prev = autograd_enabled
+        autograd_enabled = False
 
     def __exit__(self, *_):
-        global grad_enabled
-        grad_enabled = self.prev
+        global autograd_enabled
+        autograd_enabled = self.prev
 
 
 

{neuralengine-0.2.3 → neuralengine-0.2.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "NeuralEngine"
-version = "0.2.3"
+version = "0.2.4"
 description = "A framework/library for building and training neural networks."
 authors = [
     { name = "Prajjwal Pratap Shah", email = "prajjwalpratapshah@outlook.com" }

{neuralengine-0.2.3 → neuralengine-0.2.4}/setup.py

@@ -11,7 +11,7 @@ def read_requirements():
 
 setup(
     name="NeuralEngine",
-    version="0.2.3",
+    version="0.2.4",
     author="Prajjwal Pratap Shah",
     author_email="prajjwalpratapshah@outlook.com",
     maintainer="Prajjwal Pratap Shah",