bayesian generator bugfix, use best total loss

heidelberg-hepml · Mar 29, 2023 · c64dc23 · c64dc23
1 parent eb173ab
commit c64dc23
Show file tree

Hide file tree

Showing 2 changed files with 17 additions and 11 deletions.
diff --git a/src/loaders/prec_inn.py b/src/loaders/prec_inn.py
@@ -32,8 +32,11 @@ def load(params: dict) -> list[DiscriminatorData]:
     generated = pd.read_hdf(params["generated_file"]).to_numpy()
     fake_momenta = generated[:,:20].reshape(-1, 5, 4)
     fake_log_weights = generated[:,20:] if generated.shape[1] > 20 else None
-    true_momenta = true_momenta[np.all(np.isfinite(true_momenta), axis=(1,2))]
-    fake_momenta = fake_momenta[np.all(np.isfinite(fake_momenta), axis=(1,2))]
+    true_mask = np.all(np.isfinite(true_momenta), axis=(1,2))
+    fake_mask = np.all(np.isfinite(fake_momenta), axis=(1,2))
+    true_momenta = true_momenta[true_mask]
+    fake_momenta = fake_momenta[fake_mask]
+    fake_log_weights = fake_log_weights[fake_mask] if fake_log_weights is not None else None
     multiplicity_true = np.sum(true_momenta[:,:,0] != 0., axis=1)
     multiplicity_fake = np.sum(fake_momenta[:,:,0] != 0., axis=1)
 
@@ -47,7 +50,6 @@ def load(params: dict) -> list[DiscriminatorData]:
         mult = subset["multiplicity"]
         subset_true = true_momenta[multiplicity_true == mult][:,:mult]
         subset_fake = fake_momenta[multiplicity_fake == mult][:,:mult]
-        subset_logw = fake_log_weights[multiplicity_fake == mult]
         train_true, test_true, val_true = split_data(
             subset_true,
             params["train_split"],
@@ -58,11 +60,15 @@ def load(params: dict) -> list[DiscriminatorData]:
             params["train_split"],
             params["test_split"]
         )
-        train_logw, test_logw, val_logw = split_data(
-            subset_logw,
-            params["train_split"],
-            params["test_split"]
-        )
+        if fake_log_weights is None:
+            test_logw = None
+        else:
+            subset_logw = fake_log_weights[multiplicity_fake == mult]
+            train_logw, test_logw, val_logw = split_data(
+                subset_logw,
+                params["train_split"],
+                params["test_split"]
+            )
         preproc_kwargs = {
             "norm": {},
             "include_momenta": params.get("include_momenta", True),
@@ -133,7 +139,7 @@ def compute_preprocessing(
         input_obs.append(mass)
 
     if append_delta_r:
-        drinv = lambda x: np.minimum(1/(x+1e-7), 20)
+        drinv = lambda x: x #np.minimum(1/(x+1e-7), 20)
         if mult > 3:
             input_obs.append(drinv(dr(obs.phi[:,2], obs.phi[:,3], obs.eta[:,2], obs.eta[:,3])))
         if mult > 4:

diff --git a/src/train.py b/src/train.py
@@ -202,8 +202,8 @@ def train(self):
                 print(f"    Epoch {epoch:3d}: train loss {train_loss:.6f}, " +
                       f"val loss {val_loss:.6f}, LR {epoch_lr:.3e}", flush=True)
 
-            if val_bce_loss < best_val_loss:
-                best_val_loss = val_bce_loss
+            if val_loss < best_val_loss:
+                best_val_loss = val_loss
                 self.save("best")
 
             if checkpoint_interval is not None and (epoch+1) % checkpoint_interval == 0: