improve transport; add logger

.gitignore

@@ -1,3 +1,5 @@
+# Files ignored in GIT
+
 *.pdf
 *.pth
 *.dat
@@ -8,6 +10,7 @@
 *.pt
 *.zip
 *.text
+*.log
 #*.json
 #*.txt
 

configs/zee/models.yml

@@ -176,11 +176,13 @@ iceboost_swd:
 
   # Sliced Wasserstein distance [use with custom:binary_cross_entropy and custom:sliced_wasserstein]
   SWD_param:
-    beta: 5.0e-3
+    beta: 3.0e-2
     p: 1                 # p-norm (1,2, ...)
     num_slices: 500      # Number of MC projections (Higher the better) 
     mode: 'SWD'          # 'SWD' (basic), 'EBSW' (see icefit/transport.py)
     max_N: 400000        # Max events limit (400k & 500 slices works with 32 GB Nvidia V100)
+    vars: ['.*']         # Use all variables in SWD
+    #vars: ['fixedGridRhoAll', 'probe_eta', 'probe_pt'] # Use specific ones
 
 
 # ICEBOOST with an additional re-weighting in-the-loop regularization
@@ -249,7 +251,7 @@ lzmlp0: &LZMLP
     SWD_beta: 5.0e-3                  # Sliced Wasserstein [reweighting regularization]
     SWD_p: 1                          # p-norm (1,2,..), 1 perhaps more robust
     SWD_num_slices: 10000             # Number of MC projections (higher the better)
-    SWD_mode: 'SWD'                   # 'SWD' (basic), 'EBSW' (see icefit/transport.py)
+    SWD_mode: 'SWD'                   # 'SWD' (basic)
 
     lipschitz_beta:    5.0e-5         # lipschitz regularization (use with 'lzmlp')
     #logit_L1_beta: 1.0e-2            # logit norm reg. ~ beta * torch.sum(|logits|)
@@ -337,7 +339,7 @@ fastkan0: &FASTKAN
     SWD_beta: 5.0e-3                  # Sliced Wasserstein [reweighting regularization]
     SWD_p: 1                          # p-norm (1,2,..), 1 perhaps more robust
     SWD_num_slices: 10000             # Number of MC projections (higher the better)
-    SWD_mode: 'SWD'                   # 'SWD' (basic), 'EBSW' (see icefit/transport.py)
+    SWD_mode: 'SWD'                   # 'SWD' (basic)
 
     #lipshitz_beta: 1.0e-4            # Lipshitz regularization (use with 'lzmlp')
     #logit_L1_beta: 1.0e-2            # logit norm reg. ~ beta * torch.sum(|logits|)
@@ -429,7 +431,7 @@ dmlp0: &DMLP
     SWD_beta: 5.0e-3                  # Sliced Wasserstein [reweighting regularization]
     SWD_p: 1                          # p-norm (1,2,..), 1 perhaps more robust
     SWD_num_slices: 10000             # Number of MC projections (higher the better)
-    SWD_mode: 'SWD'                   # 'SWD' (basic), 'EBSW' (see icefit/transport.py)
+    SWD_mode: 'SWD'                   # 'SWD' (basic)
 
     #logit_L1_beta: 1.0e-2            # logit norm reg. ~ lambda * torch.sum(|logits|)
     logit_L2_beta: 5.0e-3             # logit norm reg. ~ lambda * torch.sum(logits**2)