argmin to argmax

Kthyeon · May 11, 2021 · e86ebe3 · e86ebe3
1 parent f819848
commit e86ebe3
Show file tree

Hide file tree

Showing 2 changed files with 9 additions and 9 deletions.
diff --git a/dividemix/Train_cifar.py b/dividemix/Train_cifar.py
@@ -320,7 +320,7 @@ def fit_mixture(scores, labels, p_threshold=0.5):
 
         gmm.fit(feats_)
         prob = gmm.predict_proba(feats_)
-        prob = prob[:,gmm.means_.argmin()]         
+        prob = prob[:,gmm.means_.argmax()]         
 #         weights, means, covars = g.weights_, g.means_, g.covariances_
 
 #         # boundary? QDA!
@@ -336,7 +336,7 @@ def fit_mixture(scores, labels, p_threshold=0.5):
     return np.array(clean_labels, dtype=np.int64)    
 
 
-def fine(current_features, current_labels, fit = 'kmeans', prev_features=None, prev_labels=None, p_threshold=0.5):
+def fine(current_features, current_labels, fit = 'kmeans', prev_features=None, prev_labels=None, p_threshold=0.7):
     '''
     prev_features, prev_labels: data from the previous round
     current_features, current_labels: current round's data
@@ -358,7 +358,7 @@ def fine(current_features, current_labels, fit = 'kmeans', prev_features=None, p
         clean_labels = cleansing(scores, current_labels)
     elif 'gmm' in fit:
         # fit a two-component GMM to the loss
-        clean_labels = fit_mixture(scores, current_labels)
+        clean_labels = fit_mixture(scores, current_labels, p_threshold)
     else:
         raise NotImplemented
 
@@ -383,14 +383,14 @@ def cleansing(scores, labels):
 
     return np.array(clean_labels, dtype=np.int64)
 
-def extract_cleanidx(model, loader, mode='fine-kmeans'):
+def extract_cleanidx(model, loader, mode='fine-kmeans', p_threshold=0.6):
     model.eval()
     for params in model.parameters(): params.requires_grad = False
 
     # get teacher_idx
     if 'fine' in mode:
         features, labels = get_features(model, loader)
-        teacher_idx = fine(current_features=features, current_labels=labels, fit = 'fine-gmm')
+        teacher_idx = fine(current_features=features, current_labels=labels, fit = 'fine-gmm', p_threshold=p_threshold)
     else: # get teacher _idx via kmeans
         teacher_idx = get_loss_list(model, loader)
 
@@ -415,7 +415,7 @@ def extract_cleanidx(model, loader, mode='fine-kmeans'):
 
 
 if args.dataset=='cifar10':
-    warm_up = 10
+    warm_up = 0
 elif args.dataset=='cifar100':
     warm_up = 30
 
@@ -463,8 +463,8 @@ def extract_cleanidx(model, loader, mode='fine-kmeans'):
     root_dir=args.data_path,log=stats_log,noise_file='%s/%.1f_%s.json'%(args.data_path,args.r,args.noise_mode))
             all_loader = loader.run('warmup')
 
-            teacher_idx_1 = extract_cleanidx(net1, all_loader, mode=args.distill_mode)
-            teacher_idx_2 = extract_cleanidx(net2, all_loader, mode=args.distill_mode)
+            teacher_idx_1 = extract_cleanidx(net1, all_loader, mode=args.distill_mode, p_threshold=args.p_threshold)
+            teacher_idx_2 = extract_cleanidx(net2, all_loader, mode=args.distill_mode, p_threshold=args.p_threshold)
 
             pred1, prob1 = None, None
             pred2, prob2 = None, None

diff --git a/dynamic_selection/selection/gmm.py b/dynamic_selection/selection/gmm.py
@@ -31,7 +31,7 @@ def fit_mixture(scores, labels, p_threshold=0.5):
 
         gmm.fit(feats_)
         prob = gmm.predict_proba(feats_)
-        prob = prob[:,gmm.means_.argmin()]         
+        prob = prob[:,gmm.means_.argmax()]         
 #         weights, means, covars = g.weights_, g.means_, g.covariances_
 
 #         # boundary? QDA!