numpy manual seed

zaeemzadeh · Apr 8, 2020 · d9e3998 · d9e3998
1 parent 44d223b
commit d9e3998
Show file tree

Hide file tree

Showing 3 changed files with 17 additions and 19 deletions.
diff --git a/acquisition.py b/acquisition.py
@@ -23,18 +23,18 @@ def acquisition(pool_loader, train_loader, model, opts):
 
     # setting the number of the samples to be pooled
     if len(train_loader.dataset.indices) == 0: # initial acquisition
-        print 'initial selection: n_pool = ', opts.init_train_size
+        print('initial selection: n_pool = ', opts.init_train_size)
         n_pool = opts.init_train_size
     else:
-        print 'n_pool = ', opts.n_pool
+        print('n_pool = ', opts.n_pool)
         n_pool = opts.n_pool
 
-    print 'extracting features of the training dataset. '
+    print('extracting features of the training dataset. ')
     train_features, train_labels = extract_features(train_loader_noshuffle, model)
-    print 'extracting features of the pooling dataset.'
+    print('extracting features of the pooling dataset.')
     pool_features, pool_labels = extract_features(pool_loader_noshuffle, model)
 
-    print 'selecting samples.'
+    print('selecting samples.')
     pooled_idx = ipm(train_features, pool_features, n_pool)
 
     pooled_idx_set = set([pool_loader_noshuffle.dataset.indices[i] for i in pooled_idx])
@@ -47,7 +47,7 @@ def acquisition(pool_loader, train_loader, model, opts):
 
 def extract_features(data_loader, model, label_only=False):
     if not label_only:
-        feature_extractor = nn.Sequential(*list(model.module.children())[:-2])
+        feature_extractor = nn.Sequential(*list(model.module.children())[:-1])
         feature_extractor = feature_extractor.cuda()
         feature_extractor = nn.DataParallel(feature_extractor, device_ids=None)
         feature_extractor.eval()
@@ -62,7 +62,6 @@ def extract_features(data_loader, model, label_only=False):
                 inputs = Variable(inputs)
                 batch_features = feature_extractor(inputs).data.view(inputs.size(0), -1)
 
-            # TODO: convert to numpy more efficiently
             features.extend(batch_features.cpu().numpy())
 
         if i % 100 == 0:

diff --git a/main.py b/main.py
@@ -50,6 +50,7 @@
         json.dump(vars(opt), opt_file)
 
     torch.manual_seed(opt.manual_seed)
+    np.random.seed(opt.manual_seed)
 
     model, parameters = generate_model(opt)
 
@@ -93,12 +94,12 @@
             pool_idx_set = set(range(len(labeled_data)))
         elif opt.init_selection == 'random':
             # split data randomly
-            print 'initial data selection: random'
+            print('initial data selection: random')
             training_idx_set = set(np.random.permutation(range(len(labeled_data)))[:opt.init_train_size])
             pool_idx_set = set(range(len(labeled_data))) - training_idx_set
         elif opt.init_selection == 'uniform_random':
             # select balanced dataset randomly
-            print 'initial data selection: uniform_random'
+            print('initial data selection: uniform_random')
             labeled_data_loader = torch.utils.data.DataLoader(
                 labeled_data,
                 batch_size=256,
@@ -187,18 +188,18 @@
     # initial selection if necessary
     clusters = []
     if opt.init_selection == 'same':
-        print 'initial data selection: same'
+        print('initial data selection: same')
         acquisition(pool_loader, train_loader, model, opt)
 
     cycle_val_acc = []
     cycle_test_acc = []
 
     while len(training_data) <= opt.max_train_size:
         print('=========================================')
-        print 'train dataset size: ', len(training_data)
-        print 'pool  dataset size: ', len(pool_data)
-        print 'max train dataset size: ', opt.max_train_size
-        print '# pooled data per cycle: ', opt.n_pool
+        print('train dataset size: ', len(training_data))
+        print('pool  dataset size: ', len(pool_data))
+        print('max train dataset size: ', opt.max_train_size)
+        print('# pooled data per cycle: ', opt.n_pool)
 
         print('Training')
         max_val_acc = 0
@@ -216,7 +217,7 @@
                 scheduler.step(validation_loss)
 
         cycle_val_acc.append(max_val_acc)
-        print cycle_val_acc
+        print(cycle_val_acc)
 
         if opt.test:
             spatial_transform = Compose([
@@ -242,7 +243,7 @@
                                        subset='validation', verbose=True, top_k=1)
             ucf101.evaluate()
             cycle_test_acc.append(ucf101.hit_at_k)
-            print cycle_test_acc
+            print(cycle_test_acc)
 
         # pool new labeled data
         if len(training_data) + opt.n_pool > opt.max_train_size:

diff --git a/opts.py b/opts.py
@@ -221,7 +221,7 @@ def parse_opts():
         type=int,
         help='ResNeXt cardinality')
     parser.add_argument(
-        '--manual_seed', default=0, type=int, help='Manually set random seed')
+        '--manual_seed', default=1, type=int, help='Manually set random seed')
 
      # acqusition args
     parser.add_argument('--init_train_size', type=int, default=101)
@@ -231,7 +231,5 @@ def parse_opts():
     parser.add_argument('--n_pool', type=int, default=101)
     parser.add_argument('--max_train_size', type=int, default=404)
 
-
-
     args = parser.parse_args()
     return args