Merge pull request #869 from marrlab/erm_hyper_init

matteo change erm

domainlab/algos/trainers/train_irm.py

@@ -66,5 +66,6 @@ def _cal_reg_loss(self, tensor_x, tensor_y, tensor_d, others=None):
         loss_2 = F.cross_entropy(phi[1::2] * dummy_w_scale, y[1::2])
         grad_1 = autograd.grad(loss_1, [dummy_w_scale], create_graph=True)[0]
         grad_2 = autograd.grad(loss_2, [dummy_w_scale], create_graph=True)[0]
-        loss_irm = torch.sum(grad_1 * grad_2)
-        return [loss_irm], [self.aconf.gamma_reg]
+        loss_irm_scalar = torch.sum(grad_1 * grad_2)  # scalar
+        loss_irm_tensor = loss_irm_scalar.expand(tensor_x.shape[0])
+        return [loss_irm_tensor], [self.aconf.gamma_reg]

domainlab/models/model_erm.py

@@ -10,7 +10,6 @@
 except:
     backpack = None
 
-
 def mk_erm(parent_class=AModelClassif, **kwargs):
     """
     Instantiate a Deepall (ERM) model
@@ -53,4 +52,30 @@ def convert4backpack(self):
             """
             self._net_invar_feat = extend(self._net_invar_feat, use_converter=True)
             self.net_classifier = extend(self.net_classifier,  use_converter=True)
+
+        def hyper_update(self, epoch, fun_scheduler): # pylint: disable=unused-argument
+            """
+            Method necessary to combine with hyperparameter scheduler
+
+            :param epoch:
+            :param fun_scheduler:
+            """
+
+        def hyper_init(self, functor_scheduler, trainer=None):
+            """
+            initiate a scheduler object via class name and things inside this model
+
+            :param functor_scheduler: the class name of the scheduler
+            """
+            return functor_scheduler(
+                trainer=trainer
+            )
+
+        @property
+        def list_str_multiplier_na(self):
+            """
+            list of multipliers which match the order in cal_reg_loss
+            """
+            return []
+
     return ModelERM

tests/test_irm.py

@@ -13,6 +13,17 @@ def test_irm():
     utils_test_algo(args)
 
 
+def test_irm_scheduler():
+    """
+    train with Invariant Risk Minimization
+    """
+    args = "--te_d=caltech --task=mini_vlcs --debug --bs=2 --model=erm \
+        --trainer=hyperscheduler_irm --nname=alexnet"
+    utils_test_algo(args)
+
+
+
+
 def test_irm_mnist():
     """
     train with Invariant Risk Minimization