l0-python 0.4.3__py3-none-any.whl → 0.5.0__py3-none-any.whl

l0/__init__.py

@@ -5,7 +5,7 @@ A PyTorch implementation of L0 regularization for neural network sparsification
 and intelligent sampling, based on Louizos, Welling, & Kingma (2017).
 """
 
-__version__ = "0.1.0"
+__version__ = "0.5.0"
 
 from .distributions import HardConcrete
 from .gates import FeatureGate, HybridGate, L0Gate, SampleGate

l0/calibration.py

@@ -60,6 +60,7 @@ class SparseCalibrationWeights(nn.Module):
         log_weight_jitter_sd: float = 0.0,
         log_alpha_jitter_sd: float = 0.01,
         device: str | torch.device = "cpu",
+        use_gates: bool = True,
     ):
         super().__init__()
         self.n_features = n_features
@@ -69,6 +70,7 @@ class SparseCalibrationWeights(nn.Module):
         self.log_weight_jitter_sd = log_weight_jitter_sd
         self.log_alpha_jitter_sd = log_alpha_jitter_sd
         self.device = torch.device(device)
+        self.use_gates = use_gates
 
         # Initialize weights (on original scale)
         if init_weights is None:
@@ -193,6 +195,10 @@ class SparseCalibrationWeights(nn.Module):
         torch.Tensor
             Positive calibration weights with L0 sparsity applied
         """
+        if not self.use_gates:
+            # No gates - just return positive weights
+            return torch.exp(self.log_weight)
+
         # Sample or get deterministic gates
         if deterministic:
             gates = self.get_deterministic_gates()
@@ -453,6 +459,46 @@ class SparseCalibrationWeights(nn.Module):
                         1 - active_info["count"] / self.n_features
                     )
 
+                    # Weight distribution diagnostic
+                    if active_weights.numel() > 0:
+                        w_tiny = (active_weights < 0.01).sum().item()
+                        w_small = (
+                            ((active_weights >= 0.01) & (active_weights < 0.1))
+                            .sum()
+                            .item()
+                        )
+                        w_med = (
+                            ((active_weights >= 0.1) & (active_weights < 1.0))
+                            .sum()
+                            .item()
+                        )
+                        w_normal = (
+                            ((active_weights >= 1.0) & (active_weights < 10.0))
+                            .sum()
+                            .item()
+                        )
+                        w_large = (
+                            (
+                                (active_weights >= 10.0)
+                                & (active_weights < 1000.0)
+                            )
+                            .sum()
+                            .item()
+                        )
+                        w_huge = (active_weights >= 1000.0).sum().item()
+
+                        total_active = active_weights.numel()
+                        weight_dist = (
+                            f"[<0.01: {100*w_tiny/total_active:.1f}%, "
+                            f"0.01-0.1: {100*w_small/total_active:.1f}%, "
+                            f"0.1-1: {100*w_med/total_active:.1f}%, "
+                            f"1-10: {100*w_normal/total_active:.1f}%, "
+                            f"10-1000: {100*w_large/total_active:.1f}%, "
+                            f">1000: {100*w_huge/total_active:.1f}%]"
+                        )
+                    else:
+                        weight_dist = "[no active weights]"
+
                     # Calculate components of the actual loss being minimized
                     actual_data_loss = data_loss.item()
                     actual_l0_loss = l0_loss.item()
@@ -464,7 +510,8 @@ class SparseCalibrationWeights(nn.Module):
                             f"mean_group_mare={mean_group_mare:.4%}, "
                             f"max_error={max_rel_err:.1%}, "
                             f"total_loss={actual_total_loss:.3f}, "
-                            f"active={active_info['count']:4d}/{self.n_features} ({sparsity_pct:.1f}% sparse)"
+                            f"active={active_info['count']:4d}/{self.n_features} ({sparsity_pct:.1f}% sparse)\n"
+                            f"         Weight dist: {weight_dist}"
                         )
                     else:
                         print(
@@ -472,7 +519,8 @@ class SparseCalibrationWeights(nn.Module):
                             f"mean_error={mean_rel_err:.4%}, "
                             f"max_error={max_rel_err:.1%}, "
                             f"total_loss={actual_total_loss:.3f}, "
-                            f"active={active_info['count']:4d}/{self.n_features} ({sparsity_pct:.1f}% sparse)"
+                            f"active={active_info['count']:4d}/{self.n_features} ({sparsity_pct:.1f}% sparse)\n"
+                            f"         Weight dist: {weight_dist}"
                         )
 
         return self

{l0_python-0.4.3.dist-info → l0_python-0.5.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: l0-python
-Version: 0.4.3
+Version: 0.5.0
 Summary: L0 regularization for sparse neural networks and intelligent sampling
 Author-email: PolicyEngine <hello@policyengine.org>
 License: MIT License

l0_python-0.5.0.dist-info/RECORD

@@ -0,0 +1,12 @@
+l0/__init__.py,sha256=ImRgZfVg-RJhLoqNaLc8ysjddZ7CYGHJPJbRUyfSoHY,1192
+l0/calibration.py,sha256=1-C6B6O47l6qlEcQl3N8waAPAyj5PFlzBHl17KbXqGo,20018
+l0/distributions.py,sha256=GvVoDasdxoDlPPdzt33lFzkuJgQsfkEn9e-abxru3nM,6325
+l0/gates.py,sha256=54OORMg9qSsdbIyzYIL9t9CCHWsKbumSKl_d4-YFuNU,11443
+l0/layers.py,sha256=j_DPEFdfUc37ojtoK2ufGzVYzZ_DwusqTcJXCVxZS68,13785
+l0/penalties.py,sha256=okqXgX2eGUmg5prcDclicTQALPPPF9FA5eQP09V4GTE,9569
+l0/sparse.py,sha256=UVpkrrJBnQ9iSNDoWK5F_sr43tQareWXzYHyMq_cKss,9726
+l0_python-0.5.0.dist-info/licenses/LICENSE,sha256=LoyFJ8qmpwWPp5RmxFGHhdBxsemvp_BqGyiCgNAZHlk,1069
+l0_python-0.5.0.dist-info/METADATA,sha256=zoUDdaSbSAWeWtWW-K_5Wgl6RbTqGHfGIyfYobvxEBQ,7095
+l0_python-0.5.0.dist-info/WHEEL,sha256=qELbo2s1Yzl39ZmrAibXA2jjPLUYfnVhUNTlyF1rq0Y,92
+l0_python-0.5.0.dist-info/top_level.txt,sha256=brP2SoElgmGhNJk-qRFEQZ5UE-d1e87Tr60hSSh8n34,3
+l0_python-0.5.0.dist-info/RECORD,,

{l0_python-0.4.3.dist-info → l0_python-0.5.0.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (80.9.0)
+Generator: setuptools (80.10.1)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

l0_python-0.4.3.dist-info/RECORD

@@ -1,12 +0,0 @@
-l0/__init__.py,sha256=ezGVLGbdU_jAElAJSdLKjSWCnZRILAZi8qMBPjcHwQc,1192
-l0/calibration.py,sha256=3VM-28M1yp-DM6AlHX7sqZCaE7OJevpY9Ey-syjSPik,17838
-l0/distributions.py,sha256=GvVoDasdxoDlPPdzt33lFzkuJgQsfkEn9e-abxru3nM,6325
-l0/gates.py,sha256=54OORMg9qSsdbIyzYIL9t9CCHWsKbumSKl_d4-YFuNU,11443
-l0/layers.py,sha256=j_DPEFdfUc37ojtoK2ufGzVYzZ_DwusqTcJXCVxZS68,13785
-l0/penalties.py,sha256=okqXgX2eGUmg5prcDclicTQALPPPF9FA5eQP09V4GTE,9569
-l0/sparse.py,sha256=UVpkrrJBnQ9iSNDoWK5F_sr43tQareWXzYHyMq_cKss,9726
-l0_python-0.4.3.dist-info/licenses/LICENSE,sha256=LoyFJ8qmpwWPp5RmxFGHhdBxsemvp_BqGyiCgNAZHlk,1069
-l0_python-0.4.3.dist-info/METADATA,sha256=gN2ufd7zkVDielkTS2iGbUeKkf9bIMTiGZtvCkpc960,7095
-l0_python-0.4.3.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-l0_python-0.4.3.dist-info/top_level.txt,sha256=brP2SoElgmGhNJk-qRFEQZ5UE-d1e87Tr60hSSh8n34,3
-l0_python-0.4.3.dist-info/RECORD,,