Changed proxynas natsbench space to optionally use train or val error thresholds.

.vscode/launch.json

@@ -228,7 +228,7 @@
             "request": "launch",
             "program": "${cwd}/scripts/main.py",
             "console": "integratedTerminal",
-            "args": ["--full", "--algos", "proxynas_natsbench_space", "--datasets", "synthetic_cifar10"]
+            "args": ["--full", "--algos", "proxynas_natsbench_space", "--datasets", "ImageNet16-120"]
         },
         {
             "name": "Proxynas-Natsbench-Space-Toy",

archai/algos/proxynas/conditional_trainer.py

@@ -34,17 +34,21 @@ class ConditionalTrainer(ArchTrainer, EnforceOverrides):
         super().__init__(conf_train, model, checkpoint)
 
         # region config vars specific to freeze trainer
-        self._train_top1_acc_threshold = conf_train['train_top1_acc_threshold']
+        self._top1_acc_threshold = conf_train['top1_acc_threshold']
+        self._use_val = conf_train['use_val']
         # endregion
 
     @overrides
     def _should_terminate(self):
         # if current validation accuracy is above threshold
         # terminate training
-        best_train_top1_avg = self._metrics.best_train_top1()
+        if self._use_val:
+            best_top1_avg = self._metrics.best_val_top1()
+        else:
+            best_top1_avg = self._metrics.best_train_top1()
 
-        if best_train_top1_avg >= self._train_top1_acc_threshold:
+        if best_top1_avg >= self._top1_acc_threshold:
-            logger.info(f'terminating at {best_train_top1_avg}')
+            logger.info(f'terminating at {best_top1_avg}')
             logger.info('----------terminating regular training---------')
             return True
         else:

confs/algos/proxynas_natsbench_space.yaml

@@ -13,6 +13,7 @@ nas:
     model_desc:
       num_edges_to_sample: 2
     loader:
+      val_ratio: 0.2
       train_batch: 256
       aug: '' # in natsbench paper they use random flip and crop, which are part of the regular transforms
       naswotrain:
@@ -21,7 +22,8 @@ nas:
         train_batch: 1024 # batch size for freeze training. 2048 works reliably on V100 with cell13 onwards unfrozen
     trainer:
       plotsdir: ''
-      train_top1_acc_threshold: 0.15 # after some accuracy we will shift into training only the last 'n' layers
+      use_val: True
+      top1_acc_threshold: 0.2 # after some accuracy we will shift into training only the last 'n' layers
       apex:
         _copy: '/common/apex'
       aux_weight: '_copy: /nas/eval/model_desc/aux_weight'