update places configuration files, unfreeze weights for the first sta…

…ge, and reduce the learning rates.
zhmiao · Feb 11, 2020 · d904d35 · d904d35
1 parent 4dd065b
commit d904d35
Show file tree

Hide file tree

Showing 2 changed files with 14 additions and 14 deletions.
diff --git a/config/Places_LT/stage_1.py b/config/Places_LT/stage_1.py
@@ -5,27 +5,27 @@
 training_opt['dataset'] = 'Places_LT'
 training_opt['log_dir'] = './logs/Places_LT/stage1'
 training_opt['num_classes'] = 365
-training_opt['batch_size'] = 256
+training_opt['batch_size'] = 128
 training_opt['num_workers'] = 4
 training_opt['num_epochs'] = 30
 training_opt['display_step'] = 10
 training_opt['feature_dim'] = 512
 training_opt['open_threshold'] = 0.1
 training_opt['sampler'] = None
-training_opt['scheduler_params'] = {'step_size':10, 'gamma':0.1}
+training_opt['scheduler_params'] = {'step_size': 10, 'gamma': 0.1}
 config['training_opt'] = training_opt
 
 networks = {}
-feature_param = {'use_modulatedatt':False, 'use_fc': True, 'dropout': None,
-			  'stage1_weights': False, 'dataset': training_opt['dataset'], 'caffe': True}
-feature_optim_param = {'lr': 0.01, 'momentum':0.9, 'weight_decay':0.0005}
+feature_param = {'use_modulatedatt': False, 'use_fc': True, 'dropout': None,
+                 'stage1_weights': False, 'dataset': training_opt['dataset'], 'caffe': True}
+feature_optim_param = {'lr': 0.001, 'momentum': 0.9, 'weight_decay': 0.0005}
 networks['feat_model'] = {'def_file': './models/ResNet152Feature.py',
                           'params': feature_param,
                           'optim_params': feature_optim_param,
-                          'fix': True}
+                          'fix': False}
 classifier_param = {'in_dim': training_opt['feature_dim'], 'num_classes': training_opt['num_classes'],
-				'stage1_weights': False, 'dataset': training_opt['dataset']}
-classifier_optim_param = {'lr': 0.1, 'momentum':0.9, 'weight_decay':0.0005}
+                    'stage1_weights': False, 'dataset': training_opt['dataset']}
+classifier_optim_param = {'lr': 0.1, 'momentum': 0.9, 'weight_decay': 0.0005}
 networks['classifier'] = {'def_file': './models/DotProductClassifier.py',
                           'params': classifier_param,
                           'optim_params': classifier_optim_param}

diff --git a/config/Places_LT/stage_2_meta_embedding.py b/config/Places_LT/stage_2_meta_embedding.py
@@ -5,7 +5,7 @@
 training_opt['dataset'] = 'Places_LT'
 training_opt['log_dir'] = './logs/Places_LT/meta_embedding'
 training_opt['num_classes'] = 365
-training_opt['batch_size'] = 256
+training_opt['batch_size'] = 128
 training_opt['num_workers'] = 4
 training_opt['num_epochs'] = 30
 training_opt['display_step'] = 10
@@ -18,15 +18,15 @@
 
 networks = {}
 feature_param = {'use_modulatedatt': True, 'use_fc': True, 'dropout': None, 
-			  'stage1_weights': True, 'dataset': training_opt['dataset'], 'caffe': False}
-feature_optim_param = {'lr': 0.01, 'momentum':0.9, 'weight_decay':0.0005}
+                 'stage1_weights': True, 'dataset': training_opt['dataset'], 'caffe': False}
+feature_optim_param = {'lr': 0.001, 'momentum': 0.9, 'weight_decay': 0.0005}
 networks['feat_model'] = {'def_file': './models/ResNet152Feature.py',
                           'params': feature_param,
                           'optim_params': feature_optim_param,
                           'fix': True}
-classifier_param = {'in_dim': training_opt['feature_dim'], 'num_classes': training_opt['num_classes'], 
-				'stage1_weights': True, 'dataset': training_opt['dataset']}
-classifier_optim_param = {'lr': 0.1, 'momentum':0.9, 'weight_decay':0.0005}
+classifier_param = {'in_dim': training_opt['feature_dim'], 'num_classes': training_opt['num_classes'],
+                    'stage1_weights': True, 'dataset': training_opt['dataset']}
+classifier_optim_param = {'lr': 0.01, 'momentum': 0.9, 'weight_decay': 0.0005}
 networks['classifier'] = {'def_file': './models/MetaEmbeddingClassifier.py',
                           'params': classifier_param,
                           'optim_params': classifier_optim_param}