Merge pull request #73 from decile-team/selcon_sahasra

Selcon sahasra

.gitignore

@@ -9,4 +9,5 @@ __pycache__
 /data/
 /log/
 /RayLogs/
-.vscode
+*.tar.gz
+.vscode

configs/SL/config_glister-warm_cifar10.py

@@ -39,7 +39,7 @@
                                 selection_type='Supervised',
                                 greedy='Stochastic'),
 
-              train_args=dict(num_epochs=300,
+              train_args=dict(num_epochs=100,
                               device="cuda",
                               print_every=10,
                               results_dir='results/',

configs/SL/config_selcon_lawschool.py

@@ -0,0 +1,50 @@
+# Learning setting
+config = dict(setting="SL",
+
+              dataset=dict(name="LawSchool_selcon",
+                           datadir="../data",
+                           feature="dss",
+                           type="pre-defined"),
+
+              dataloader=dict(shuffle=True,
+                              batch_size=100,
+                              pin_memory=False),
+
+              model=dict(architecture='RegressionNet',
+                         type='pre-defined',
+                         input_dim=10,
+                         numclasses=10),
+
+              ckpt=dict(is_load=False,
+                        is_save=True,
+                        dir='results/',
+                        save_every=20),
+
+              loss=dict(type='MeanSquaredLoss',
+                        use_sigmoid=False),
+
+              optimizer=dict(type="adam", 
+                             lr=0.01),
+
+              scheduler=dict(type="StepLR", # added this new scheduler type
+                             step_size=1,
+                             gamma=0.1),
+
+              dss_args=dict(type="SELCON",
+                                fraction=0.01,
+                                select_every=35,
+                                kappa=0,
+                                delta=0.04,
+                                linear_layer=False,
+                                lam=1e-5,
+                                batch_sampler='sequential',
+                                selection_type='Supervised'),
+
+              train_args=dict(num_epochs=200,
+                              device_selcon="cuda",
+                              print_every=1,
+                              results_dir='results/',
+                              print_args=["val_loss", "tst_loss", "trn_loss", "time"],
+                              return_args=[]
+                              )
+              )

cords/selectionstrategies/SL/__init__.py

@@ -5,4 +5,5 @@
 from .submodularselectionstrategy import SubmodularSelectionStrategy
 from .gradmatchstrategy import GradMatchStrategy
 from .fixedweightstrategy import FixedWeightStrategy
-from .adapweightsstrategy import AdapWeightsStrategy
+from .selconstrategy import SELCONstrategy
+from .adapweightsstrategy import AdapWeightsStrategy

cords/selectionstrategies/SL/glisterstrategy.py

@@ -122,6 +122,8 @@ def _update_grads_val(self, grads_curr=None, first_init=False):
         if first_init:
             for batch_idx, (inputs, targets) in enumerate(valloader):
                 inputs, targets = inputs.to(self.device), targets.to(self.device, non_blocking=True)
+                print(inputs.shape, targets.shape)
+                exit(0) 
                 if batch_idx == 0:
                     out, l1 = self.model(inputs, last=True, freeze=True)
                     loss = self.loss(out, targets).sum()