adaptiveLR v3

ovotools/adaptive_lr.py

@@ -4,10 +4,10 @@ import ignite
 
 def create_adaptive_supervised_trainer(model, optimizer, loss_fn, metrics={},
                               device=None, non_blocking=False,
-                              prepare_batch=ignite.engine._prepare_batch, lr_scale = 1.1, warmup_iters = 50):
+                              prepare_batch=ignite.engine._prepare_batch, lr_scale = 1.1, warmup_iters = 50, ls_mult = 3):
     """
     Factory function for creating a trainer for supervised models.
-
+l
     Args:
         model (`torch.nn.Module`): the model to train.
         optimizer (`torch.optim.Optimizer`): the optimizer to use.
@@ -41,29 +41,38 @@ def create_adaptive_supervised_trainer(model, optimizer, loss_fn, metrics={},
         model.train()
 
         if engine.state.iteration > warmup_iters:
-            prev_k = 1
-            loss = None
-            new_ks_list = (1/lr_scale, lr_scale,)
+            if engine.state.iteration % 2:
+                new_k = 1 / lr_scale
+            else:
+                new_k =  lr_scale
+            for group in optimizer.param_groups:
+                group['lr'] *= new_k
+        else:
+            prev_k = new_k = 1
+
+        if engine.state.iteration > 1:
+            optimizer.step()
+
+        if engine.state.iteration > warmup_iters:
             with torch.no_grad():
-                for new_k in new_ks_list:
-                    correct_model(prev_k, new_k)
                 y_pred = model(x)
-                    loss0 = loss
-                    loss = loss_fn(y_pred, y)
+                loss0 = loss_fn(y_pred, y)
+                print('iter\t{}.{}'.format(engine.state.epoch, engine.state.iteration), 'lr * {:5.3}'.format(new_k), 'loss', loss0.item())
                 prev_k = new_k
-                    print('iter\t{}.{}'.format(engine.state.epoch, engine.state.iteration), 'lr',
-                          optimizer.param_groups[0]['lr'], '*', new_k, 'loss', loss.item())
-                if loss0 < loss or (loss0 == loss and engine.state.iteration % 2):
-                    new_k = new_ks_list[0]
+                new_k = 1/new_k
                 correct_model(prev_k, new_k)
-                for group in optimizer.param_groups:
-                    group['lr'] *= new_k
-
         optimizer.zero_grad()
         y_pred = model(x)
         loss = loss_fn(y_pred, y)
         loss.backward()
-        optimizer.step()
+
+        if engine.state.iteration > warmup_iters:
+            with torch.no_grad():
+                print('iter\t{}.{}'.format(engine.state.epoch, engine.state.iteration), 'lr * {:5.3}'.format(new_k), 'loss', loss.item())
+
+                if loss < loss0 or (loss == loss0 and engine.state.iteration % 2):
+                    for group in optimizer.param_groups:
+                        group['lr'] *= new_k/prev_k
 
                 print('iter\t{}.{}'.format(engine.state.epoch, engine.state.iteration), 'lr', optimizer.param_groups[0]['lr'], 'loss', loss.item())