♻️ remove input/embedding dropout rates

saber/config.ini

@@ -34,7 +34,7 @@ decay = 0.0
 # Three dropout values must be specified (separated by a comma), corresponding to the dropout rate
 # to apply to the input, output and recurrent connections respectively. Must be a value between 0.0
 # and 1.0.
-dropout_rate = 0.3, 0.3, 0.1
+dropout_rate = 0.1
 
 batch_size = 32
 

saber/config.py

@@ -156,7 +156,7 @@ def _parse_config_args(self, config):
             args['learning_rate'] = config['training'].getfloat('learning_rate')
             args['decay'] = config['training'].getfloat('decay')
             args['grad_norm'] = config['training'].getfloat('grad_norm')
-            args['dropout_rate'] = config['training']['dropout_rate'].split(',')
+            args['dropout_rate'] = config['training'].getfloat('dropout_rate')
             args['batch_size'] = config['training'].getint('batch_size')
             args['validation_split'] = config['training'].getfloat('validation_split')
             args['k_folds'] = config['training'].getint('k_folds')
@@ -206,12 +206,6 @@ def _post_process_config_args(self, args):
         args['output_folder'] = generic_utils.clean_path(args['output_folder'])
         if args['pretrained_model'] and args['pretrained_model'] not in PRETRAINED_MODELS:
             args['pretrained_model'] = generic_utils.clean_path(args['pretrained_model'])
-        # build dictionary for dropout rates
-        args['dropout_rate'] = {
-            'input': float(args['dropout_rate'][0]),
-            'output': float(args['dropout_rate'][1]),
-            'recurrent': float(args['dropout_rate'][2]),
-        }
 
         return args
 
@@ -254,11 +248,8 @@ def _parse_cli_args(self):
         parser.add_argument('--decay', required=False, type=float,
                             help=('float >= 0. Learning rate decay over each update. Note that for '
                                   'certain optimizers this value is ignored. Defaults to 0.'))
-        parser.add_argument('--dropout_rate', required=False, nargs=3, type=float,
-                            metavar=('input', 'output', 'recurrent'),
-                            help=('Expects three values, separated by a space, which specify the '
-                                  'fraction of units to drop for input, output and recurrent '
-                                  'connections respectively. Values must be between 0 and 1.'))
+        parser.add_argument('--dropout_rate', required=False, type=float,
+                            help='float between 0 and 1. Fraction of the input units to drop.')
         parser.add_argument('--grad_norm', required=False, type=float,
                             help='Tau threshold value for gradient normalization, defaults to 1.')
         parser.add_argument('--k_folds', required=False, type=int,

saber/constants.py

@@ -133,7 +133,7 @@
                        'pubmed&id=')
 # CONFIG
 CONFIG_ARGS = ['model_name', 'save_model', 'dataset_folder', 'output_folder',
-               'pretrained_model', 'word_embed_dim', 'char_embed_dim', 'optimizer', 'activation',
-               'learning_rate', 'decay', 'grad_norm', 'dropout_rate', 'batch_size', 'k_folds',
-               'epochs', 'criteria', 'verbose', 'debug', 'save_all_weights', 'tensorboard',
-               'replace_rare_tokens', 'variational_dropout', 'validation_split', 'dataset_reader']
+               'pretrained_model', 'optimizer', 'activation', 'learning_rate', 'decay', 'grad_norm',
+               'dropout_rate', 'batch_size', 'k_folds', 'epochs', 'criteria', 'verbose', 'debug',
+               'save_all_weights', 'tensorboard', 'replace_rare_tokens', 'validation_split',
+               'dataset_reader']

saber/tests/resources/constants.py

@@ -115,7 +115,7 @@
                             'learning_rate': '0.0',
                             'grad_norm': '1.0',
                             'decay': '0.0',
-                            'dropout_rate': '0.3, 0.3, 0.1',
+                            'dropout_rate': '0.1',
                             'batch_size': '32',
                             'validation_split': '0.0',
                             'k_folds': '2',
@@ -143,7 +143,7 @@
                           'learning_rate': 0.0,
                           'decay': 0.0,
                           'grad_norm': 1.0,
-                          'dropout_rate': {'input': 0.3, 'output': 0.3, 'recurrent': 0.1},
+                          'dropout_rate': 0.1,
                           'batch_size': 32,
                           'validation_split': 0.0,
                           'k_folds': 2,
@@ -162,7 +162,7 @@
                            'grad_norm': 1.0,
                            'learning_rate': 0.05,
                            'decay': 0.5,
-                           'dropout_rate': [0.6, 0.6, 0.2],
+                           'dropout_rate': 0.6,
                            # the datasets are used for test purposes so they must
                            # point to the correct resources, this can be ensured by passing their
                            # paths here
@@ -181,7 +181,7 @@
                             'learning_rate': 0.05,
                             'decay': 0.5,
                             'grad_norm': 1.0,
-                            'dropout_rate': {'input': 0.6, 'output': 0.6, 'recurrent': 0.2},
+                            'dropout_rate': 0.6,
                             'batch_size': 32,
                             'validation_split': 0.0,
                             'k_folds': 2,

saber/tests/resources/dummy_config.ini

@@ -34,7 +34,7 @@ decay = 0.0
 # Three dropout values must be specified (separated by a comma), corresponding to the dropout rate
 # to apply to the input, output and recurrent connections respectively. Must be a value between 0.0
 # and 1.0.
-dropout_rate = 0.3, 0.3, 0.1
+dropout_rate = 0.1
 
 batch_size = 32