updates

configs/experiments/cifar10/cifar10_5.yaml

@@ -41,7 +41,7 @@ LR:
 TEST:
   EMP_PARAMS: True
   SCORE: True
-  MAGNITUDE: 0.0024
+  MAGNITUDE: 0.00
   IN_FEATS: [64, 128, 128, 512]
 
 COMMENTS:

losses.py

@@ -9,7 +9,7 @@ def gaussian_log_p(x, mean, log_sd):
 
 
 class FlowConLoss:
-  def __init__(self, cfg, device, p_y=None):
+  def __init__(self, cfg, device, test=False):
     self.cfg = cfg
     self.device=device
     self.n_bins = cfg.FLOW.N_BINS
@@ -18,6 +18,7 @@ def __init__(self, cfg, device, p_y=None):
 
     self.tau = cfg.LOSS.TAU
     self.tau2 = cfg.LOSS.TAU2
+    self.test = test
 
 
     # RAF12
@@ -49,15 +50,24 @@ def nllLoss(self, z, logdet, mu, log_sd):
 
     logdet = logdet.mean()
     loss = self.init_loss + logdet + log_p_nll
+
+    if self.test:
+      score = loss / (log(2) * self.n_pixel), # CONVERTING LOGe to LOG2 |
+      log_p = log_p_nll / (log(2) * self.n_pixel)  #                     v
+      return (
+        score, 
+        log_p, 
+        (logdet / (log(2) * self.n_pixel)).mean(), 
+        (log_p_all/ (log(2) * self.n_pixel))
+      ) 
+    else:
+      return ( 
 
-    return ( 
-      # (-loss / (log(2) * self.n_pixel)).mean(), # CONVERTING LOGe to LOG2 |
-      # (log_p_nll / (log(2) * self.n_pixel)).mean(), #                     v
-      (loss / (log(2) * self.n_pixel)), # CONVERTING LOGe to LOG2 |
-      (log_p_nll / (log(2) * self.n_pixel)), #                     v
-      (logdet / (log(2) * self.n_pixel)).mean(), 
-      (log_p_all/ (log(2) * self.n_pixel))
-  )
+        (-loss / (log(2) * self.n_pixel)).mean(), # CONVERTING LOGe to LOG2 |
+        (log_p_nll / (log(2) * self.n_pixel)).mean(), #                     v
+        (logdet / (log(2) * self.n_pixel)).mean(), 
+        (log_p_all/ (log(2) * self.n_pixel))
+    )
 
 
   def conLoss(self, log_p_all, labels):

tester/emperical_params.py

@@ -69,7 +69,6 @@ def calc_emp_params(cfg, args, loader, pretrained, flow, dist_dir, device, label
   z = torch.cat(z_all, dim=0)
 
   plot_umap(cfg, z.cpu(), labels_all.cpu(), f"{args.config}", 2, "in_ood", labels_in_ood)
-  e()
 
   mu_k, std_k , z_k= [], [], []
   for cls in range(cfg.DATASET.N_CLASS):
@@ -276,7 +275,7 @@ def calc_score(cfg, args, loader, pretrained, flow, mu, log_sd, criterion, devic
   # ood_datasets = ['cifar10']
   result = {}
 
-  criterion = FlowConLoss(cfg, device)
+  criterion = FlowConLoss(cfg, device, test=True)
 
 
   loss_criterion = nn.CrossEntropyLoss()

utils.py

@@ -265,8 +265,6 @@ def plot_umap(cfg, X_lst_un, y_lst, name, dim, mode, labels_in_ood=None):
     df = pd.DataFrame(X_lst, columns=["x", "y"])
   df_color = pd.DataFrame(y_lst_label, columns=["class"])
   df['legend_group'] = df_color['class'].apply(lambda x: 'ID' if x in id_list else "OOD")
-  print(df_color.head())
-  # e()
   df = df.join(df_color)
   if dim == 3:
     fig = px.scatter_3d(df, x='x', y='y', z='z',color='class', title=f"{name}", \
@@ -276,11 +274,8 @@ def plot_umap(cfg, X_lst_un, y_lst, name, dim, mode, labels_in_ood=None):
                     #  color_discrete_sequence=colors_all
     )
 
-#   fig.update_traces(marker=dict(size=8))
-  # fig.update_traces(marker=dict(size=12, line=dict(width=2)))
-  # fig.update_traces(marker=dict(size=[12 if c == '100' else 8 for c in df['class']]))
   fig.for_each_trace(lambda t: t.update(showlegend=False) if t.name in id_list else t.update(name='OOD'))
-  fig.update_layout(title="CV vs Entropy",  margin=dict(l=0, r=0, t=0, b=0), showlegend=True)
+  fig.update_layout(title=None,  margin=dict(l=0, r=0, t=0, b=0), showlegend=True)
   fig.update_layout(
       legend=dict(
           yanchor="top",