add instructions

data/vctk/prep_vctk.py

@@ -14,17 +14,27 @@
 
 parser = argparse.ArgumentParser()
 
-parser.add_argument('--file-list')
-parser.add_argument('--out')
-parser.add_argument('--in-dir', default='~/')
-parser.add_argument('--interpolate', action='store_true')
-parser.add_argument('--low-pass', action='store_true')
-parser.add_argument('--dimension', type=int, default=6400)
-parser.add_argument('--stride', type=int, default=3200)
-parser.add_argument('--scale', type=int, default=2)
-parser.add_argument('--batch-size', type=int, default=128)
+parser.add_argument('--file-list',
+  help='list of input wav files to process')
+parser.add_argument('--in-dir', default='~/',
+  help='folder where input files are located')
+parser.add_argument('--out',
+  help='path to output h5 archive')
+parser.add_argument('--scale', type=int, default=2,
+  help='scaling factor')
+parser.add_argument('--dimension', type=int, default=8192,
+  help='dimension of patches')
+parser.add_argument('--stride', type=int, default=3200,
+  help='stride when extracting patches')
+parser.add_argument('--interpolate', action='store_true', 
+  help='interpolate low-res patches with cubic splines')
+parser.add_argument('--low-pass', action='store_true',
+  help='apply low-pass filter when generating low-res patches')
+parser.add_argument('--batch-size', type=int, default=128,
+  help='we produce # of patches that is a multiple of batch size')
 parser.add_argument('--sr', type=int, default=16000, help='audio sampling rate')
-parser.add_argument('--sam', type=float, default=1.)
+parser.add_argument('--sam', type=float, default=1.,
+  help='subsampling factor for the data')
 
 args = parser.parse_args()
 

src/run.py

@@ -23,28 +23,36 @@ def make_parser():
   train_parser = subparsers.add_parser('train')
   train_parser.set_defaults(func=train)
 
-  train_parser.add_argument('--train', required=True)
-  train_parser.add_argument('--val', required=True)
-  train_parser.add_argument('-e', '--epochs', type=int, default=100)
-  train_parser.add_argument('--batch-size', type=int, default=128)
-  train_parser.add_argument('--load', help='preload-existing model')
-  train_parser.add_argument('--logname', default='tmp-run')
-  train_parser.add_argument('--layers', default=4, type=int)
-  train_parser.add_argument('--alg', default='adam')
-  train_parser.add_argument('--lr', default=1e-3, type=float)
-  train_parser.add_argument('--sr', help='sampling rate for the wav', 
-                                    type=int, default=16000)
+  train_parser.add_argument('--train', required=True,
+    help='path to h5 archive of training patches')
+  train_parser.add_argument('--val', required=True,
+    help='path to h5 archive of validation set patches')
+  train_parser.add_argument('-e', '--epochs', type=int, default=100,
+    help='number of epochs to train')
+  train_parser.add_argument('--batch-size', type=int, default=128,
+    help='training batch size')
+  train_parser.add_argument('--logname', default='tmp-run',
+    help='folder where logs will be stored')
+  train_parser.add_argument('--layers', default=4, type=int,
+    help='number of layers in each of the D and U halves of the network')
+  train_parser.add_argument('--alg', default='adam',
+    help='optimization algorithm')
+  train_parser.add_argument('--lr', default=1e-3, type=float,
+    help='learning rate')
 
   # eval
 
   eval_parser = subparsers.add_parser('eval')
   eval_parser.set_defaults(func=eval)
 
-  eval_parser.add_argument('--logname', required=True)
-  eval_parser.add_argument('--out-label', default='')
-  eval_parser.add_argument('--wav-file-list', help='list of audio files')
+  eval_parser.add_argument('--logname', required=True,
+    help='path to training checkpoint')
+  eval_parser.add_argument('--out-label', default='',
+    help='append label to output samples')
+  eval_parser.add_argument('--wav-file-list', 
+    help='list of audio files for evaluation')
   eval_parser.add_argument('--r', help='upscaling factor', type=int)
-  eval_parser.add_argument('--sr', help='sampling rate', 
+  eval_parser.add_argument('--sr', help='high-res sampling rate', 
                                    type=int, default=16000)
 
   return parser
@@ -61,21 +69,8 @@ def train(args):
   r = Y_train[0].shape[1] / X_train[0].shape[1]
   assert n_chan == 1
 
-  # # determine super-resolution level
-  # n_chan, n_dim = Y_train[0].shape
-  # r = Y_train[0].shape[1] / X_train[0].shape[1]
-  # assert n_chan == 1
-
-  # # transponse to match tensorflow (batch, height, width, chan) format
-  # X_train, X_val = X_train.transpose([0,2,1]), X_val.transpose([0,2,1])
-  # Y_train, Y_val = Y_train.transpose([0,2,1]), Y_val.transpose([0,2,1])
-
-  # load model
-  from_ckpt = True if args.load is not None else False
-  model = get_model(args, n_dim, r, from_ckpt=from_ckpt, train=True)
-
-  # load checkpoint
-  if from_ckpt: model.load(args.load)
+  # create model
+  model = get_model(args, n_dim, r, from_ckpt=False, train=True)
 
   # train model
   model.fit(X_train, Y_train, X_val, Y_val, n_epoch=args.epochs)