pyTorchAutoForge 0.3.1.dev0__tar.gz → 0.4.0__tar.gz

{pytorchautoforge-0.3.1.dev0 → pytorchautoforge-0.4.0}/.gitignore

@@ -170,3 +170,6 @@ cython_debug/
 #  and can be added to the global gitignore or merged into this file.  For a more nuclear
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
 #.idea/
+.vscode/browse.vc.db
+.vscode/browse.vc.db-shm
+.vscode/browse.vc.db-wal

{pytorchautoforge-0.3.1.dev0 → pytorchautoforge-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pyTorchAutoForge
-Version: 0.3.1.dev0
+Version: 0.4.0
 Summary: PyTorchAutoForge library is based on raw PyTorch and designed to automate DNN development, model tracking and deployment, tightly integrated with MLflow and Optuna. It supports Spiking networks libraries (WIP). Deployment can be performed using ONNX, pyTorch facilities or TensorRT (WIP). The library is designed to be compatible with Jetson Orin Nano Jetpack rev6.1, with bash script to automatically configure virtualenv.
 Author-email: "Pietro Califano (PC)" <petercalifano.gs@gmail.com>
 License: MIT
@@ -39,7 +39,6 @@ Requires-Dist: torch; platform_machine == 'x86_64'
 Requires-Dist: torchinfo
 Requires-Dist: torchvision; platform_machine == 'x86_64'
 Provides-Extra: cuda-all
-Requires-Dist: pycuda; (platform_machine == 'x86_64') and extra == 'cuda-all'
 Requires-Dist: pynvml; (platform_machine == 'x86_64') and extra == 'cuda-all'
 Description-Content-Type: text/markdown
 

{pytorchautoforge-0.3.1.dev0 → pytorchautoforge-0.4.0}/_version.py

@@ -28,7 +28,7 @@ version_tuple: VERSION_TUPLE
 commit_id: COMMIT_ID
 __commit_id__: COMMIT_ID
 
-__version__ = version = '0.3.1.dev0'
-__version_tuple__ = version_tuple = (0, 3, 1, 'dev0')
+__version__ = version = '0.4.0'
+__version_tuple__ = version_tuple = (0, 4, 0)
 
 __commit_id__ = commit_id = None

{pytorchautoforge-0.3.1.dev0 → pytorchautoforge-0.4.0}/environment.yml

@@ -197,7 +197,6 @@ dependencies:
       - pyasn1-modules==0.4.1
       - pybind11==2.13.6
       - pycparser==2.22
-      - pycuda==2025.1
       - pydantic==2.10.6
       - pydantic-core==2.27.2
       - pygments==2.19.1

{pytorchautoforge-0.3.1.dev0 → pytorchautoforge-0.4.0}/pyTorchAutoForge/api/tensorrt/TRTengineExporter.py

@@ -2,7 +2,6 @@ import tensorrt as trt
 import torch 
 import numpy as np
 import sys, os, shutil
-import pycuda as cuda
 import subprocess
 from enum import Enum
 from pyTorchAutoForge.api.onnx import ModelHandlerONNx
@@ -167,4 +166,4 @@ class TRTengineExporter:
             print("Engine built successfully and saved to:", output_engine_path)
 
         else:
-            raise ValueError("Invalid TRTengineExporterMode value. Please select TRTEXEC or PYTHON.")
+            raise ValueError("Invalid TRTengineExporterMode value. Please select TRTEXEC or PYTHON.")

{pytorchautoforge-0.3.1.dev0 → pytorchautoforge-0.4.0}/pyTorchAutoForge/datasets/LabelsClasses.py

@@ -30,6 +30,8 @@ class PTAF_Datakey(enum.Enum):
     PHASE_ANGLE = 9  # Phase angle of the scene
     CENTRE_OF_FIGURE = 10  # Centre of figure of the object
     APPARENT_SIZE = 11  # Apparent size of the object in pixels
+    SUN_DIRECTION_ANGLE_FROM_X = 12  # Angle of the sun direction from the x-axis in radians
+
 
     def get_lbl_vector_size(self):
         # Define sizes for data keys
@@ -49,7 +51,8 @@ class PTAF_Datakey(enum.Enum):
             PTAF_Datakey.REFERENCE_SIZE: 1,
             PTAF_Datakey.PHASE_ANGLE: 1,
             PTAF_Datakey.CENTRE_OF_FIGURE: 2,  # x, y coordinates of the centre of figure
-            PTAF_Datakey.APPARENT_SIZE: 1  # Apparent size in pixels
+            PTAF_Datakey.APPARENT_SIZE: 1,  # Apparent size in pixels
+            PTAF_Datakey.SUN_DIRECTION_ANGLE_FROM_X: 1  # Angle of the sun direction from the x-axis in radians
         }
 
         return sizes.get(self, None)
@@ -235,7 +238,7 @@ class AuxiliaryLabels(BaseLabelsContainer):
         default=-1.0, metadata={'yaml': 'dPhaseAngleInDeg'})
     
     light_direction_rad_angle_from_x: float = field(
-        default=0.0, metadata={'yaml': 'dLightDirectionRadAngleFromX'})
+        default=-1.0, metadata={'yaml': 'dLightDirectionRadAngleFromX'})
     
     object_shape_matrix_cam_frame: list[list[float]] = field(
         default_factory=list, metadata={'yaml': 'dObjectShapeMatrix_CamFrame'})
@@ -311,6 +314,9 @@ class LabelsContainer(BaseLabelsContainer):
 
         elif item == PTAF_Datakey.APPARENT_SIZE:
             return float(self.geometric.obj_apparent_size_in_pix)
+        
+        elif item == PTAF_Datakey.SUN_DIRECTION_ANGLE_FROM_X:
+            return float(self.auxiliary.light_direction_rad_angle_from_x)
 
         else:
             # Raise AttributeError to maintain standard behavior

{pytorchautoforge-0.3.1.dev0 → pytorchautoforge-0.4.0}/pyTorchAutoForge/model_building/backbones/base_backbones.py

@@ -25,6 +25,7 @@ class FeatureExtractorConfig(BaseConfigClass):
     # Dimension of the final linear layer (if you want to add a linear layer)
     output_size: int | None = None
     remove_classifier: bool = True
+    remove_gap_layer: bool = False
     device: torch.device | str | None = None
     input_channels: int = 3 # Placeholder value
 

{pytorchautoforge-0.3.1.dev0 → pytorchautoforge-0.4.0}/pyTorchAutoForge/model_building/backbones/efficient_net.py

@@ -5,6 +5,7 @@ from .base_backbones import EfficientNetConfig, FeatureExtractorFactory
 
 import torch
 import torch.nn.functional as F
+import warnings
 from pyTorchAutoForge.model_building.backbones.spatial_features_operators import SpatialKptFeatureSoftmaxLocator
 from pyTorchAutoForge.model_building.poolingBlocks import CustomAdaptiveMaxPool2d
 
@@ -22,7 +23,13 @@ class EfficientNetBackbone(nn.Module):
 
         # Extract the “features” part all children except the final classifier/sequential
         modules = list(model.children())[:-1]
+
         if cfg.output_type == 'last':
+
+            if hasattr(cfg, 'remove_gap_layer'):
+                # Remove the final global average pooling layer (AdaptiveAvgPool2d) if specified in config
+                modules = modules[0] if isinstance(modules[-1], nn.AdaptiveAvgPool2d) and cfg.remove_gap_layer else modules
+
             # Wrap as a single ModuleList so that forward is simple
             self.feature_extractor = nn.ModuleList([nn.Sequential(*modules)])
 
@@ -35,8 +42,10 @@ class EfficientNetBackbone(nn.Module):
             self.feature_extractor = nn.ModuleList(
                 list(feature_extractor_modules.children()))
 
-            # Add last layer (global adaptive pooling) from modules
-            self.feature_extractor.append(modules[1])
+            # Add last layer (global adaptive pooling) from modules if not removed
+            if hasattr(cfg, 'remove_gap_layer'):
+                if isinstance(modules[-1], nn.AdaptiveAvgPool2d) and not(cfg.remove_gap_layer):
+                    self.feature_extractor.append(modules[-1])
 
             # Build average pooling layer
             self.feature_spill_preprocessor = nn.ModuleDict()

{pytorchautoforge-0.3.1.dev0 → pytorchautoforge-0.4.0}/pyproject.toml

@@ -56,7 +56,6 @@ classifiers=[
 [project.optional-dependencies]
 cuda_all = [
     "pynvml; platform_machine == 'x86_64'", 
-    "pycuda; platform_machine == 'x86_64'", 
 ]
 
 [tool.setuptools.packages.find]