Merge pull request #176 from automl/load_labeled_fix

Fix for loading labeled architecture in optimizers
automl · Sep 21, 2023 · fc40ee5 · fc40ee5
2 parents dfa2e67 + 927efe5
commit fc40ee5
Show file tree

Hide file tree

Showing 3 changed files with 67 additions and 19 deletions.
diff --git a/naslib/optimizers/discrete/bananas/optimizer.py b/naslib/optimizers/discrete/bananas/optimizer.py
@@ -60,6 +60,9 @@ def __init__(self, config, zc_api=None):
             config.search, 'zc_names') else None
         self.zc_only = config.search.zc_only if hasattr(
             config.search, 'zc_only') else False
+
+        self.load_labeled = config.search.load_labeled if hasattr(
+            config.search, 'load_labeled') else False
 
     def adapt_search_space(self, search_space, scope=None, dataset_api=None):
         assert (
@@ -119,7 +122,7 @@ def _sample_new_model(self):
         model = torch.nn.Module()
         model.arch = self.search_space.clone()
         model.arch.sample_random_architecture(
-            dataset_api=self.dataset_api, load_labeled=self.use_zc_api)
+            dataset_api=self.dataset_api, load_labeled=self.load_labeled)
         model.arch_hash = model.arch.get_hash()
 
         if self.search_space.instantiate_model == True:

diff --git a/naslib/optimizers/discrete/npenas/optimizer.py b/naslib/optimizers/discrete/npenas/optimizer.py
@@ -59,6 +59,8 @@ def __init__(self, config, zc_api=None):
             config.search, 'zc_names') else None 
         self.zc_only = config.search.zc_only if hasattr(
             config.search, 'zc_only') else False
+        self.load_labeled = config.search.load_labeled if hasattr(
+            config.search, 'load_labeled') else False
 
     def adapt_search_space(self, search_space, scope=None, dataset_api=None):
         assert (
@@ -119,7 +121,7 @@ def _sample_new_model(self):
         model = torch.nn.Module()
         model.arch = self.search_space.clone()
         model.arch.sample_random_architecture(
-            dataset_api=self.dataset_api, load_labeled=self.use_zc_api)
+            dataset_api=self.dataset_api, load_labeled=self.load_labeled)
         model.arch_hash = model.arch.get_hash()
 
         if self.search_space.instantiate_model == True:

diff --git a/naslib/runners/zc/zc_config.yaml b/naslib/runners/zc/zc_config.yaml
@@ -6,23 +6,66 @@ cutout_prob: 1.0
 dataset: cifar10
 out_dir: run
 predictor: fisher
-search_space: nasbench101 #nasbench201 #nasbench301
-seed: 0
+search_space: nasbench201 #nasbench101  #nasbench301
 test_size: 200
 train_size: 400
-zc_ensemble: true
-zc_only: true
-zc_names:
-  - params
-  - flops
-  - jacov
-  - plain
-  - grasp
-  - snip
-  - fisher
-  - grad_norm
-  - epe_nas
-  - synflow
-  - l2_norm
 optimizer: npenas
-train_portion: 0.7
+train_portion: 0.7
+seed: 0
+
+search:
+  # for bohb
+  seed: 0
+  budgets: 50000000
+  checkpoint_freq: 1000
+  fidelity: 108 
+
+  # for all optimizers
+  epochs: 10
+
+  # for bananas and npenas, choose one predictor 
+  # out of the 16 model-based predictors
+  predictor_type: var_sparse_gp
+
+  # number of initial architectures
+  num_init: 10  
+
+  # NPENAS
+  k: 10
+  num_ensemble: 3
+  acq_fn_type: its
+  acq_fn_optimization: mutation
+  encoding_type: adjacency_one_hot
+  num_arches_to_mutate: 1 
+  max_mutations: 1
+  num_candidates: 50
+
+  # jacov data loader
+  batch_size: 256
+  data_size: 25000
+  cutout: False
+  cutout_length: 16
+  cutout_prob: 1.0
+  train_portion: 0.7
+
+  # other params
+  debug_predictor: False
+  sample_size: 10
+  population_size: 30
+
+  # zc parameters
+  use_zc_api: False
+  zc_ensemble: true
+  zc_names:
+    - params
+    - flops
+    - jacov
+    - plain
+    - grasp
+    - snip
+    - fisher
+    - grad_norm
+    - epe_nas
+    - synflow
+    - l2_norm
+  zc_only: true