Still debugging local search on DARTS.

archai/algos/local_search_darts/local_search_darts_reg.py

@@ -1,5 +1,6 @@
 import os
 import random
+import math as ma
 from overrides.overrides import overrides
 from typing import List, Tuple, Optional, Dict
 from archai.nas.discrete_search_space import DiscreteSearchSpace
@@ -78,6 +79,19 @@ class LocalSearchDartsReg(LocalSearch):
         if self.local_minima:
             best_minimum = max(self.local_minima, key=lambda x:x[1])
             return best_minimum
+        else:
+            # if no local minima encountered, return the best 
+            # encountered so far
+            logger.info('No local minima encountered! Returning best of visited.')
+            max_train_top1 = -ma.inf
+            argmax_arch = None
+            for _, arch in self.eval_cache.items():
+                this_train_top1 = arch.metadata['train_top1']
+                if this_train_top1 > max_train_top1:
+                    max_train_top1 = this_train_top1
+                    argmax_arch = arch
+            return argmax_arch, max_train_top1
+            
 
 
     @overrides
@@ -106,9 +120,9 @@ class LocalSearchDartsReg(LocalSearch):
         train_top1 = trainer_metrics.best_train_top1()
         arch.metadata['train_top1'] = train_top1
 
-        # DEBUG: simulate architecture evaluation
-        train_top1 = random.random()
-        arch.metadata['train_top1'] = train_top1
+        # # DEBUG: simulate architecture evaluation
+        # train_top1 = random.random()
+        # arch.metadata['train_top1'] = train_top1
 
         # cache it
         self.eval_cache[arch_flat_rep] = arch

archai/search_spaces/discrete_search_spaces/darts_search_spaces/discrete_search_space_darts.py

@@ -160,20 +160,21 @@ class DiscreteSearchSpaceDARTS(DiscreteSearchSpace):
         op_nbrs = op_nbrs_regular + op_nbrs_reduction
         assert len(op_nbrs) == 96
 
-        # now create the edge neighbors where the 
-        # only difference is in one of the input edges
-        # there should be 24 of them
-        edge_nbrs_regular = self._get_edge_neighbors(central_regular_cell, central_desc)
-        edge_nbrs_reduction = self._get_edge_neighbors(central_reduction_cell, central_desc)
+        # # now create the edge neighbors where the 
+        # # only difference is in one of the input edges
+        # # there should be 24 of them
+        # edge_nbrs_regular = self._get_edge_neighbors(central_regular_cell, central_desc)
+        # edge_nbrs_reduction = self._get_edge_neighbors(central_reduction_cell, central_desc)
 
-        assert len(edge_nbrs_regular) == 12
-        assert len(edge_nbrs_reduction) == 12
+        # assert len(edge_nbrs_regular) == 12
+        # assert len(edge_nbrs_reduction) == 12
 
-        edge_nbrs = edge_nbrs_regular + edge_nbrs_reduction                                           
-        assert len(edge_nbrs) == 24
+        # edge_nbrs = edge_nbrs_regular + edge_nbrs_reduction                                           
+        # assert len(edge_nbrs) == 24
 
         # Now convert all the model descs to actual Models
-        all_nbrs = op_nbrs + edge_nbrs
+        # all_nbrs = op_nbrs + edge_nbrs
+        all_nbrs = op_nbrs
         all_models = [Model(nbr_desc, self.drop_path_prob, affine=True) for nbr_desc in all_nbrs]
         
         all_arch_meta = []

confs/algos/local_search_darts_reg.yaml

@@ -9,7 +9,7 @@ nas:
     loader:
       val_ratio: 0.0 # don't need val for local search
     trainer:
-      epochs: 2
+      epochs: 1
 
   eval:
     model_desc: