Merge branch 'master' of gitlab.doc.ic.ac.uk:yw21218/metarl

benchmark/pickles/04_22_cf_ln_gru.txt

+[(('ShearY', 0.2, 5), ('Rotate', 0.6, 6)),
+ (('TranslateX', 0.8, 3), ('Posterize', 0.1, 3)),
+ (('TranslateY', 0.0, 8), ('Equalize', 0.7, None)),
+ (('Equalize', 0.3, None), ('Contrast', 0.2, 0)),
+ (('ShearX', 0.4, 5), ('Contrast', 0.2, 8)),
+ (('TranslateX', 0.9, 3), ('Solarize', 0.4, 5)),
+ (('Color', 0.2, 4), ('Solarize', 0.6, 8)),
+ (('ShearX', 0.1, 8), ('Equalize', 0.4, None)),
+ (('Posterize', 0.7, 5), ('Solarize', 1.0, 4))][0.6056999564170837, 0.6329999566078186, 0.6171000003814697, 0.62909996509552]original small policys accuracies: [0.6236000061035156, 0.6187999844551086, 0.617900013923645]
\ No newline at end of file

benchmark/pickles/04_22_fm_sn_gru.txt

+[(('Color', 0.9, 3), ('Contrast', 0.8, 3)),
+ (('Sharpness', 0.9, 0), ('Solarize', 0.3, 7)),
+ (('Color', 0.0, 6), ('Solarize', 0.4, 3)),
+ (('Brightness', 0.1, 3), ('Brightness', 0.5, 9)),
+ (('Solarize', 0.9, 6), ('Rotate', 0.6, 1)),
+ (('Contrast', 0.7, 3), ('Posterize', 0.9, 4)),
+ (('Solarize', 0.6, 2), ('Contrast', 0.5, 6)),
+ (('TranslateX', 0.0, 4), ('AutoContrast', 0.3, None)),
+ (('Equalize', 0.0, None), ('Brightness', 0.8, 1))][0.7490999698638916, 0.8359999656677246, 0.8394999504089355]original small policys accuracies: [0.8380999565124512, 0.8376999497413635, 0.8376999497413635]
\ No newline at end of file

benchmark/pickles/04_22_fm_sn_rs.txt

+[(('ShearX', 1.0, 0), ('Color', 0.3, 2)),
+ (('AutoContrast', 0.0, None), ('Brightness', 0.7, 2)),
+ (('Invert', 0.1, None), ('Contrast', 0.1, 6)),
+ (('Solarize', 0.4, 2), ('Contrast', 0.9, 2)),
+ (('Equalize', 0.0, None), ('Contrast', 0.0, 2)),
+ (('Rotate', 0.4, 0), ('Posterize', 0.5, 9)),
+ (('Posterize', 0.7, 3), ('Invert', 0.1, None)),
+ (('Solarize', 0.6, 1), ('Contrast', 0.0, 0)),
+ (('Color', 0.2, 6), ('Posterize', 0.4, 7))][0.6222999691963196, 0.6868000030517578, 0.8374999761581421, 0.8370999693870544, 0.6934999823570251]original small policys accuracies: [0.8431999683380127, 0.8393999934196472, 0.8377999663352966]
\ No newline at end of file

benchmark/scripts/04_22_fm_gru.py

@@ -40,6 +40,7 @@ run_benchmark(
     child_network_architecture=child_network_architecture,
     agent_arch=aal.gru_learner,
     config=config,
+    total_iter=144
     )
 
 rerun_best_policy(
@@ -48,5 +49,6 @@ rerun_best_policy(
     train_dataset=train_dataset,
     test_dataset=test_dataset,
     child_network_architecture=child_network_architecture,
+    config=config,
     repeat_num=5
     )
\ No newline at end of file

benchmark/scripts/util_04_22.py

@@ -6,7 +6,7 @@ import torch
 import MetaAugment.child_networks as cn
 import MetaAugment.autoaugment_learners as aal
 
-from pprint import pprint
+import pprint
 
 """
 testing gru_learner and randomsearch_learner on
@@ -75,16 +75,21 @@ def get_mega_policy(history, n):
         assert len(history) >= n
 
         # agent.history is a list of (policy(list), val_accuracy(float)) tuples 
-        sorted_history = sorted(history, key=lambda x:x[1]) # sort wrt acc
+        sorted_history = sorted(history, key=lambda x:x[1], reverse=True) # sort wrt acc
 
         best_history = sorted_history[:n]
 
         megapolicy = []
+        # we also want to keep track of how good the best policies were
+        # maybe if we add them all up, they'll become worse! Hopefully better tho
+        orig_accs = []
+
         for policy,acc in best_history:
             for subpolicy in policy:
                 megapolicy.append(subpolicy)
+            orig_accs.append(acc)
         
-        return megapolicy
+        return megapolicy, orig_accs
 
 
 def rerun_best_policy(
@@ -93,25 +98,30 @@ def rerun_best_policy(
     train_dataset,
     test_dataset,
     child_network_architecture,
+    config,
     repeat_num
     ):
 
     with open(agent_pickle, 'rb') as f:
-        agent = torch.load(f, map_location=device)
+        agent = torch.load(f)
     
-    megapol = get_mega_policy(agent.history)
+    megapol, orig_accs = get_mega_policy(agent.history,3)
     print('mega policy to be tested:')
-    pprint(megapol)
-    
+    pprint.pprint(megapol)
+    print(orig_accs)
+
     accs=[]
     for _ in range(repeat_num):
         print(f'{_}/{repeat_num}')
+        temp_agent = aal.aa_learner(**config)
         accs.append(
-                agent.test_autoaugment_policy(megapol,
+                temp_agent.test_autoaugment_policy(megapol,
                                     child_network_architecture,
                                     train_dataset,
                                     test_dataset,
                                     logging=False)
                     )
         with open(accs_txt, 'w') as f:
+            f.write(pprint.pformat(megapol))
             f.write(str(accs))
+            f.write(f'original small policys accuracies: {orig_accs}')