[Update] Release semilearn 0.3.1.

.github/workflows/blank.yml

@@ -33,4 +33,4 @@ jobs:
       - name: Run a multi-line script
         run: |
           echo Add other actions to build,
-          echo test, and deploy your project.
+          echo test, and deploy your project.

.github/workflows/issue.yml

@@ -0,0 +1,17 @@
+name: 'Close stale issues and PR'
+on:
+  schedule:
+    - cron: '30 1 * * *'
+
+jobs:
+  stale:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/stale@v8
+        with:
+          stale-issue-message: 'This issue is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.'
+          stale-pr-message: 'This PR is stale because it has been open 45 days with no activity. Remove stale label or comment or this will be closed in 10 days.'
+          close-issue-message: 'This issue was closed because it has been stalled for 5 days with no activity.'
+          days-before-stale: 30
+          days-before-close: 5
+          days-before-pr-close: -1

requirements.txt

@@ -16,9 +16,6 @@ joblib==1.2.0
 kiwisolver==1.4.3
 Markdown==3.3.7
 matplotlib==3.5.2
-mkl-fft>=1.3.0
-mkl-random==1.2.2
-mkl-service==2.4.0
 networkx==2.6.3
 numpy
 oauthlib==3.2.1
@@ -51,7 +48,7 @@ tifffile==2021.11.2
 timm==0.5.4
 tokenizers==0.12.1
 torch>=1.12.0
-torchaudio==0.12.0
+torchaudio>=0.12.0
 torchvision>=0.13.0
 tqdm==4.64.0
 transformers==4.20.1

semilearn/algorithms/meanteacher/meanteacher.py

@@ -52,8 +52,8 @@ def train_step(self, x_lb, y_lb, x_ulb_w, x_ulb_s):
 
             self.bn_controller.freeze_bn(self.model)
             outs_x_ulb_s = self.model(x_ulb_s)
-            logits_x_ulb_s = outs_x_ulb_w['logits']
-            feats_x_ulb_s = outs_x_ulb_w['feat']
+            logits_x_ulb_s = outs_x_ulb_s['logits']
+            feats_x_ulb_s = outs_x_ulb_s['feat']
             self.bn_controller.unfreeze_bn(self.model)
 
             feat_dict = {'x_lb':feats_x_lb, 'x_ulb_w':feats_x_ulb_w, 'x_ulb_s':feats_x_ulb_s}

semilearn/datasets/utils.py

@@ -52,10 +52,6 @@ def split_ssl_data(args, data, targets, num_classes,
     return data[lb_idx], targets[lb_idx], data[ulb_idx], targets[ulb_idx]
 
 
-def sample_labeled_data():
-    pass
-
-
 def sample_labeled_unlabeled_data(args, data, target, num_classes,
                                   lb_num_labels, ulb_num_labels=None,
                                   lb_imbalance_ratio=1.0, ulb_imbalance_ratio=1.0,
@@ -88,7 +84,7 @@ def sample_labeled_unlabeled_data(args, data, target, num_classes,
     if ulb_imbalance_ratio == 1.0:
         # balanced setting
         if ulb_num_labels is None or ulb_num_labels == 'None':
-            pass # ulb_samples_per_class = [int(len(data) / num_classes) - lb_samples_per_class[c] for c in range(num_classes)] # [int(len(data) / num_classes) - int(lb_num_labels / num_classes)] * num_classes
+            ulb_samples_per_class = None
         else:
             assert ulb_num_labels % num_classes == 0, "ulb_num_labels must be dividable by num_classes in balanced setting"
             ulb_samples_per_class = [int(ulb_num_labels / num_classes)] * num_classes
@@ -104,7 +100,7 @@ def sample_labeled_unlabeled_data(args, data, target, num_classes,
         idx = np.where(target == c)[0]
         np.random.shuffle(idx)
         lb_idx.extend(idx[:lb_samples_per_class[c]])
-        if ulb_num_labels is None or ulb_num_labels == 'None':
+        if ulb_samples_per_class is None:
             ulb_idx.extend(idx[lb_samples_per_class[c]:])
         else:
             ulb_idx.extend(idx[lb_samples_per_class[c]:lb_samples_per_class[c]+ulb_samples_per_class[c]])

semilearn/imb_algorithms/debiaspl/debiaspl.py

@@ -49,7 +49,7 @@ def set_hooks(self):
     def compute_prob(self, logits):
         # update p_hat
         probs = super().compute_prob(logits)
-        delta_p = probs.mean()
+        delta_p = probs.mean(dim=0)
         self.p_hat = self.ema_m * self.p_hat + (1 - self.ema_p) * delta_p
         return super().compute_prob(logits - self.tau * torch.log(self.p_hat))
 

semilearn/lighting/config.py

@@ -69,7 +69,7 @@ def get_config(config):
     ## core algorithm setting
     parser.add_argument('-alg', '--algorithm', type=str, default='fixmatch', help='ssl algorithm')
     parser.add_argument('--use_cat', type=str2bool, default=True, help='use cat operation in algorithms')
-    parser.add_argument('--use_amp', type=str2bool, default=False, help='use mixed precision training or not')
+    parser.add_argument('--amp', type=str2bool, default=False, help='use mixed precision training or not')
     parser.add_argument('--clip_grad', type=float, default=0)
 
     ## imbalance algorithm setting

semilearn/lighting/trainer.py

@@ -28,16 +28,17 @@ def __init__(self, config, algorithm, verbose=0):
 
 
     def fit(self, train_lb_loader, train_ulb_loader, eval_loader):
+        self.algorithm.loader_dict = {
+            'trian_lb': train_lb_loader,
+            'train_ulb': train_ulb_loader,
+            'eval': eval_loader
+        }
         self.algorithm.model.train()
-
-        # EMA Init
-        self.algorithm.ema = EMA(self.algorithm.model, self.algorithm.ema_m)
-        self.algorithm.ema.register()
-
         # train
         self.algorithm.it = 0
         self.algorithm.best_eval_acc = 0.0
         self.algorithm.best_epoch = 0
+        self.algorithm.call_hook("before_run")
 
         for epoch in range(self.config.epoch):
             self.algorithm.epoch = epoch
@@ -48,19 +49,26 @@ def fit(self, train_lb_loader, train_ulb_loader, eval_loader):
             bar = Bar('Processing', max=len(train_lb_loader))
 
             self.algorithm.model.train()
+            self.algorithm.call_hook("before_train_epoch")
 
             for data_lb, data_ulb in zip(train_lb_loader, train_ulb_loader):
 
                 if self.algorithm.it > self.config.num_train_iter:
                     break
 
-                result = self.algorithm.train_step(**self.algorithm.process_batch(**data_lb, **data_ulb))
-
+                self.algorithm.call_hook("before_train_step")
+                out_dict, log_dict = self.algorithm.train_step(**self.algorithm.process_batch(**data_lb, **data_ulb))
+                self.algorithm.out_dict = out_dict
+                self.algorithm.log_dict = log_dict
+                self.algorithm.call_hook("after_train_step")
+
                 bar.suffix = ("Iter: {batch:4}/{iter:4}.".format(batch=self.algorithm.it, iter=len(train_lb_loader)))
                 bar.next()
                 self.algorithm.it += 1
             bar.finish()
 
+            self.algorithm.call_hook("after_train_epoch")
+
             # validate
             result = self.evaluate(eval_loader)
 

setup.py

@@ -26,7 +26,7 @@
 
 setup(
     name='semilearn',
-    version='0.3.0',
+    version='0.3.1',
     description='Unfied Semi-Supervised Learning Benchmark',
     long_description=long_description,
     long_description_content_type='text/markdown',
@@ -41,5 +41,5 @@
     include_package_data=True,
     # install_requires=['torch >= 1.8', 'torchvision', 'torchaudio', 'transformers', 'timm', 'progress', 'ruamel.yaml', 'scikit-image', 'scikit-learn', 'tensorflow', ''],
     install_requires=install_requires,
-    python_requires='>=3.8',
+    python_requires='>=3.9',
 )