Add checkpoints used for preemption. (#3789)

references/detection/train.py

@@ -188,13 +188,19 @@ def main(args):
         train_one_epoch(model, optimizer, data_loader, device, epoch, args.print_freq)
         lr_scheduler.step()
         if args.output_dir:
-            utils.save_on_master({
+            checkpoint = {
                 'model': model_without_ddp.state_dict(),
                 'optimizer': optimizer.state_dict(),
                 'lr_scheduler': lr_scheduler.state_dict(),
                 'args': args,
-                'epoch': epoch},
+                'epoch': epoch
+            }
+            utils.save_on_master(
+                checkpoint,
                 os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))
+            utils.save_on_master(
+                checkpoint,
+                os.path.join(args.output_dir, 'checkpoint.pth'))
 
         # evaluate after every epoch
         evaluate(model, data_loader_test, device=device)

references/segmentation/train.py

@@ -157,15 +157,19 @@ def main(args):
         train_one_epoch(model, criterion, optimizer, data_loader, lr_scheduler, device, epoch, args.print_freq)
         confmat = evaluate(model, data_loader_test, device=device, num_classes=num_classes)
         print(confmat)
-        utils.save_on_master(
-            {
+        checkpoint = {
             'model': model_without_ddp.state_dict(),
             'optimizer': optimizer.state_dict(),
             'lr_scheduler': lr_scheduler.state_dict(),
             'epoch': epoch,
             'args': args
-            },
+        }
+        utils.save_on_master(
+            checkpoint,
             os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))
+        utils.save_on_master(
+            checkpoint,
+            os.path.join(args.output_dir, 'checkpoint.pth'))
 
     total_time = time.time() - start_time
     total_time_str = str(datetime.timedelta(seconds=int(total_time)))