Merge pull request #59 from Vivswan/develop

v1.0.8

CHANGELOG.md

@@ -1,41 +1,38 @@
 # Changelog
 
+## 1.0.8
+* Removed redundant code from `reduce_precision`.
+* Added `types` argument to `PseudoParameter.parametrize_module` for better selection for Parameterising the Layers.
+
 ## 1.0.7
 * Fixed `GeLU` backward function equation.
 
 ## 1.0.6
-
 * `Model` is subclass of `BackwardModule` for additional functionality.
 * Using `inspect.isclass` to check if `backward_class` is a class in `Linear.set_backward_function`.
 * Repr using `self.__class__.__name__` in all classes.
 
 ## 1.0.5 (Patches for Pytorch 2.0.1)
-
 * Removed unnecessary `PseudoParameter.grad` property.
 * Patch for Pytorch 2.0.1, add filtering inputs in `BackwardGraph._calculate_gradients`.
 
 ## 1.0.4
-
 * Combined `PseudoParameter` and `PseudoParameterModule` for better visibility.
   * BugFix: fixed save and load of state_dict of `PseudoParameter` and transformation module.
 * Removed redundant class `analogvnn.parameter.Parameter`.
 
 ## 1.0.3
-
 * Added support for no loss function in `Model` class.
   * If no loss function is provided, the `Model` object will use outputs for gradient computation.
 * Added support for multiple loss outputs from loss function.
 
 ## 1.0.2
-
 * Bugfix: removed  `graph` from `Layer` class.
   * `graph` was causing issues with nested `Model` objects.
   * Now `_use_autograd_graph` is directly set while compiling the `Model` object.
 
 ## 1.0.1 (Patches for Pytorch 2.0.0)
-
 * added `grad.setter` to `PseudoParameterModule` class.
 
 ## 1.0.0
-
 * Public release.

CITATION.cff

@@ -11,10 +11,10 @@ preferred-citation:
     - family-names: Youngblood
       given-names: Nathan
       affiliation: University of Pittsburgh
-  doi: "10.48550/arXiv.2210.10048"
-  journal: "arXiv preprint arXiv:2210.10048"
+  doi: "10.1063/5.0134156"
+  journal: "APL Machine Learning"
   title: 'AnalogVNN: A fully modular framework for modeling and optimizing photonic neural networks'
-  year: 2022
+  year: 2023
 authors:
   - given-names: Vivswan
     family-names: Shah
@@ -25,7 +25,7 @@ authors:
     affiliation: University of Pittsburgh
 identifiers:
   - type: doi
-    value: 10.48550/arXiv.2210.10048
+    value: 10.1063/5.0134156
     description: >-
       The concept DOI for the collection containing
       all versions of the Citation File Format.

README.md

@@ -1,6 +1,7 @@
 # AnalogVNN
 
 [![arXiv](https://img.shields.io/badge/arXiv-2210.10048-orange.svg)](https://arxiv.org/abs/2210.10048)
+[![AML](https://img.shields.io/badge/AML-10.1063/5.0134156-orange.svg)](https://doi.org/10.1063/5.0134156)
 [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/Vivswan/AnalogVNN/blob/release/docs/_static/AnalogVNN_Demo.ipynb)
 
 [![PyPI version](https://badge.fury.io/py/analogvnn.svg)](https://badge.fury.io/py/analogvnn)
@@ -52,24 +53,28 @@ digital neural network models to their analog counterparts with just a few lines
 taking full advantage of the open-source optimization, deep learning, and GPU acceleration
 libraries available through PyTorch.
 
-AnalogVNN Paper: [https://arxiv.org/abs/2210.10048](https://arxiv.org/abs/2210.10048)
+AnalogVNN Paper: [https://doi.org/10.1063/5.0134156](https://doi.org/10.1063/5.0134156)
 
 ## Citing AnalogVNN
 
 We would appreciate if you cite the following paper in your publications for which you used AnalogVNN:
 
 ```bibtex
-@article{shah2022analogvnn,
+@article{shah2023analogvnn,
   title={AnalogVNN: A fully modular framework for modeling and optimizing photonic neural networks},
   author={Shah, Vivswan and Youngblood, Nathan},
-  journal={arXiv preprint arXiv:2210.10048},
-  year={2022}
+  journal={APL Machine Learning},
+  volume={1},
+  number={2},
+  year={2023},
+  publisher={AIP Publishing}
 }
 ```
 
 Or in textual form:
 
 ```text
 Vivswan Shah, and Nathan Youngblood. "AnalogVNN: A fully modular framework for modeling 
-and optimizing photonic neural networks." arXiv preprint arXiv:2210.10048 (2022).
+and optimizing photonic neural networks." APL Machine Learning 1.2 (2023).
+DOI: 10.1063/5.0134156
 ```

analogvnn/fn/reduce_precision.py

@@ -12,19 +12,16 @@ def reduce_precision(x: TENSOR_OPERABLE, precision: TENSOR_OPERABLE, divide: TEN
     Args:
       x (TENSOR_OPERABLE): Tensor
       precision (TENSOR_OPERABLE): the precision of the quantization.
-      divide (TENSOR_OPERABLE): the number of bits to be reduced
+      divide (TENSOR_OPERABLE): the rounding value that is if divide is 0.5,
+         then 0.6 will be rounded to 1.0 and 0.4 will be rounded to 0.0.
 
     Returns:
       TENSOR_OPERABLE: TENSOR_OPERABLE with the same shape as x, but with values rounded to the nearest
       multiple of precision.
     """
 
     x = x if isinstance(x, Tensor) else torch.tensor(x, requires_grad=False)
-    g: Tensor = x * precision
-    f = torch.sign(g) * torch.maximum(
-        torch.floor(torch.abs(g)),
-        torch.ceil(torch.abs(g) - divide)
-    ) * (1 / precision)
+    f = torch.sign(x) * torch.ceil(torch.abs(x * precision) - divide) * (1 / precision)
     return f
 
 

analogvnn/parameter/PseudoParameter.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from typing import Callable, Any
+from typing import Callable, Any, Optional, Union, Tuple
 
 import torch
 import torch.nn as nn
@@ -219,13 +219,20 @@ def parameterize(cls, module: nn.Module, param_name: str, transformation: Callab
         return new_param
 
     @classmethod
-    def parametrize_module(cls, module: nn.Module, transformation: Callable, requires_grad: bool = True):
+    def parametrize_module(
+            cls,
+            module: nn.Module,
+            transformation: Callable,
+            requires_grad: bool = True,
+            types: Optional[Union[type, Tuple[type]]] = None,
+    ):
         """Parametrize all parameters of a module.
 
         Args:
             module (nn.Module): the module parameters to parametrize.
             transformation (Callable): the transformation.
             requires_grad (bool): if True, only parametrized parameters that require gradients.
+            types (Union[type, Tuple[type]]): the type or tuple of types to parametrize.
         """
 
         with torch.no_grad():
@@ -236,6 +243,9 @@ def parametrize_module(cls, module: nn.Module, transformation: Callable, require
                 if requires_grad and not parameter.requires_grad:
                     continue
 
+                if types is not None and not isinstance(parameter, types):
+                    continue
+
                 cls.parameterize(module=module, param_name=name, transformation=transformation)
 
             for sub_module in module.children():

pyproject.toml

@@ -19,7 +19,7 @@ where = ["analogvnn"]
 [project]
 # $ pip install analogvnn
 name = "analogvnn"
-version = "1.0.7"
+version = "1.0.8"
 description = "A fully modular framework for modeling and optimizing analog/photonic neural networks"
 readme = "README.md"
 requires-python = ">=3.7"

requirements.txt

@@ -2,6 +2,7 @@
 torch
 torchvision
 torchaudio
+dataclasses
 numpy>=1.22.2
 scipy
 networkx