Merge pull request #51 from spokenlanguage/50-constant-adam

50 constant adam

platalea/basic.py

@@ -96,6 +96,8 @@ def val_loss(net):
         scheduler = platalea.schedulers.cyclic(optimizer, len(data['train']), max_lr=config['max_lr'], min_lr=config['min_lr'])
     elif configured_scheduler == 'noam':
         scheduler = platalea.schedulers.noam(optimizer, config['d_model'])
+    elif configured_scheduler == 'constant':
+        scheduler = platalea.schedulers.constant(optimizer, config['constant_lr'])
     else:
         raise Exception("lr_scheduler config value " + configured_scheduler + " is invalid, use cyclic or noam")
     optimizer.zero_grad()

platalea/experiments/config.py

@@ -72,12 +72,14 @@ def __str__(self):
     '--downsampling-factor', default=None, type=float,
     dest='downsampling_factor',
     help='factor by which the dataset should be downsampled')
-args.add_argument('--lr_scheduler', default="cyclic", choices=['cyclic', 'noam'],
+args.add_argument('--lr_scheduler', default="cyclic", choices=['cyclic', 'noam', 'constant'],
                   help='The learning rate scheduler to use. WARNING: noam not yet implemented for most experiments!')
 args.add_argument('--cyclic_lr_max', default=2 * 1e-4, type=float,
                   help='Maximum learning rate for cyclic learning rate scheduler')
 args.add_argument('--cyclic_lr_min', default=1e-6, type=float,
                   help='Minimum learning rate for cyclic learning rate scheduler')
+args.add_argument('--constant_lr', default=1e-4, type=float,
+                  help='Learning rate for constant learning rate scheduler')
 args.add_argument('--device', type=str, default=None,
                   help="Device to train on. Can be passed on to platalea.hardware.device in experiments.")
 

platalea/experiments/flickr8k/transformer.py

@@ -11,7 +11,6 @@
 # Parsing arguments
 args.add_argument('--batch_size', default=32, type=int,
                   help='How many samples per batch to load.')
-
 args.add_argument('--trafo_d_model', default=512, type=int,
                   help='TRANSFORMER: The dimensionality of the transformer model.')
 args.add_argument('--trafo_encoder_layers', default=4, type=int,
@@ -77,7 +76,9 @@ def __new__(cls, value):
 logging.info('Building model')
 net = M.SpeechImage(config)
 run_config = dict(max_lr=args.cyclic_lr_max, min_lr=args.cyclic_lr_min, epochs=args.epochs, lr_scheduler=args.lr_scheduler,
-                  d_model=args.trafo_d_model, score_on_cpu=args.score_on_cpu, validate_on_cpu=args.validate_on_cpu)
+                  d_model=args.trafo_d_model, score_on_cpu=args.score_on_cpu, validate_on_cpu=args.validate_on_cpu,
+                  constant_lr=args.constant_lr,
+                  )
 
 logged_config = dict(run_config=run_config, encoder_config=config, speech_config=speech_config)
 logged_config['encoder_config'].pop('SpeechEncoder')  # Object info is redundant in log.

platalea/schedulers.py

@@ -34,3 +34,18 @@ def learning_rate(iteration):
 
     scheduler = lr_scheduler.LambdaLR(optimizer, learning_rate)
     return scheduler
+
+
+def constant(optimizer, lr):
+    """
+    Constant learning rate scheduler. The most trivial kind of scheduler, that keeps the learning rate constant.
+    :param optimizer:
+    :param lr:
+    :return:
+    """
+    logging.info("Using constant learning rate of {}".format(lr))
+
+    def learning_rate(_iteration):
+        return lr
+
+    return lr_scheduler.LambdaLR(optimizer, learning_rate, last_epoch=-1)