Use click instead of climate.

scripts/theanets-char-rnn

@@ -5,97 +5,86 @@ import logging
 import numpy as np
 import theanets
 
-g = climate.add_group('Data')
-g.add_argument('-d', '--data', nargs='+', metavar='FILE',
-               help='load text from FILE')
-g.add_argument('-t', '--time', default=100, type=int, metavar='T',
-               help='train on sequences of T characters')
-g.add_argument('-a', '--alphabet', default='', metavar='CHARS',
-               help='use CHARS for alphabet; defaults to all chars in text')
-g.add_argument('-A', '--exclude-alphabet', default='', metavar='CHARS',
-               help='discard CHARS from alphabet')
-
-g = climate.add_group('Architecture')
-g.add_argument('-l', '--layers', nargs='+', type=int, default=[100], metavar='N',
-               help='construct a network with layers of size N1, N2, ...')
-g.add_argument('-L', '--layer-type', default='lstm', metavar='{rnn|gru|lstm|clockwork}',
-               help='construct a network with this RNN layer type')
-g.add_argument('-g', '--activation', default='relu', metavar='FUNC',
-               help='function for hidden unit activations')
-
-g = climate.add_group('Training')
-g.add_argument('-O', '--algorithm', default=['nag'], nargs='+', metavar='ALGO',
+@click.command()
+@click.option('-d', '--data', multiple=True, metavar='FILE',
+              help='load text from FILE')
+@click.option('-t', '--time', default=100, type=int, metavar='T',
+              help='train on sequences of T characters')
+@click.option('-a', '--alphabet', default='', metavar='CHARS',
+              help='use CHARS for alphabet; defaults to all chars in text')
+@click.option('-A', '--exclude-alphabet', default='', metavar='CHARS',
+              help='discard CHARS from alphabet')
+@click.option('-l', '--layers', multiple=True, type=int, default=[100], metavar='N',
+              help='construct a network with layers of size N1, N2, ...')
+@click.option('-L', '--layer-type', default='lstm', metavar='{rnn|gru|lstm|clockwork}',
+              help='construct a network with this RNN layer type')
+@click.option('-g', '--activation', default='relu', metavar='FUNC',
+              help='function for hidden unit activations')
+@click.option('-O', '--algorithm', default=['nag'], multiple=True, metavar='ALGO',
                help='train with the given optimization algorithm(s)')
-g.add_argument('-p', '--patience', type=int, default=4, metavar='N',
+@click.option('-p', '--patience', type=int, default=4, metavar='N',
                help='stop training if less than --min-improvement for N validations')
-g.add_argument('-v', '--validate-every', type=int, default=10, metavar='N',
+@click.option('-v', '--validate-every', type=int, default=10, metavar='N',
                help='validate the model every N updates')
-g.add_argument('-b', '--batch-size', type=int, default=64, metavar='N',
+@click.option('-b', '--batch-size', type=int, default=64, metavar='N',
                help='train with mini-batches of size N')
-g.add_argument('-B', '--train-batches', type=int, metavar='N',
+@click.option('-B', '--train-batches', type=int, metavar='N',
                help='use at most N batches during gradient computations')
-g.add_argument('-V', '--valid-batches', type=int, metavar='N',
+@click.option('-V', '--valid-batches', type=int, metavar='N',
                help='use at most N batches during validation')
-g.add_argument('-i', '--min-improvement', type=float, default=0, metavar='R',
+@click.option('-i', '--min-improvement', type=float, default=0, metavar='R',
                help='train until relative improvement is less than R')
-g.add_argument('-x', '--max-gradient-norm', type=float, default=1, metavar='V',
+@click.option('-x', '--max-gradient-norm', type=float, default=1, metavar='V',
                help='clip gradient norm to the interval [0, V]')
-g.add_argument('-r', '--learning-rate', type=float, default=1e-4, metavar='V',
+@click.option('-r', '--learning-rate', type=float, default=1e-4, metavar='V',
                help='train the network with a learning rate of V')
-g.add_argument('-m', '--momentum', type=float, default=0.9, metavar='V',
+@click.option('-m', '--momentum', type=float, default=0.9, metavar='V',
                help='train the network with momentum of V')
-g.add_argument('-n', '--nesterov', action='store_true',
-               help='use Nesterov momentum')
-g.add_argument('-s', '--save-progress', metavar='FILE',
+@click.option('-n', '--nesterov/--no-nesterov', help='use Nesterov momentum')
+@click.option('-s', '--save-progress', metavar='FILE',
                help='save the model periodically to FILE')
-g.add_argument('-S', '--save-every', type=float, default=0, metavar='N',
+@click.option('-S', '--save-every', type=float, default=0, metavar='N',
                help='save the model every N iterations or -N minutes')
-
-g = climate.add_group('Regularization')
-g.add_argument('--input-noise', type=float, default=0, metavar='S',
+@click.option('--input-noise', type=float, default=0, metavar='S',
                help='add noise to network inputs drawn from N(0, S)')
-g.add_argument('--input-dropouts', type=float, default=0, metavar='R',
+@click.option('--input-dropouts', type=float, default=0, metavar='R',
                help='randomly set fraction R of input activations to 0')
-g.add_argument('--hidden-noise', type=float, default=0, metavar='S',
+@click.option('--hidden-noise', type=float, default=0, metavar='S',
                help='add noise to hidden activations drawn from N(0, S)')
-g.add_argument('--hidden-dropouts', type=float, default=0, metavar='R',
+@click.option('--hidden-dropouts', type=float, default=0, metavar='R',
                help='randomly set fraction R of hidden activations to 0')
-g.add_argument('--hidden-l1', type=float, default=0, metavar='K',
+@click.option('--hidden-l1', type=float, default=0, metavar='K',
                help='regularize hidden activity with K on the L1 term')
-g.add_argument('--hidden-l2', type=float, default=0, metavar='K',
+@click.option('--hidden-l2', type=float, default=0, metavar='K',
                help='regularize hidden activity with K on the L2 term')
-g.add_argument('--weight-l1', type=float, default=0, metavar='K',
+@click.option('--weight-l1', type=float, default=0, metavar='K',
                help='regularize network weights with K on the L1 term')
-g.add_argument('--weight-l2', type=float, default=0, metavar='K',
+@click.option('--weight-l2', type=float, default=0, metavar='K',
                help='regularize network weights with K on the L2 term')
-
-g = climate.add_group('RmsProp Optimization')
-g.add_argument('--rms-halflife', type=float, default=5, metavar='N',
+@click.option('--rms-halflife', type=float, default=5, metavar='N',
                help='use a half-life of N for RMS exponential moving averages')
-g.add_argument('--rms-regularizer', type=float, default=1e-8, metavar='N',
+@click.option('--rms-regularizer', type=float, default=1e-8, metavar='N',
                help='regularize RMS exponential moving averages by N')
-
-
-def main(args):
+def main(**kwargs):
     corpus = []
-    for f in args.data:
+    for f in kwargs['data']:
         corpus.append(open(f).read())
         logging.info('%s: loaded training document', f)
     logging.info('loaded %d training documents', len(corpus))
 
-    alpha = set(args.alphabet)
+    alpha = set(kwargs['alphabet'])
     if not alpha:
         for c in corpus:
             alpha |= set(c)
-    alpha -= set(args.exclude_alphabet)
+    alpha -= set(kwargs['exclude_alphabet'])
     alpha = sorted(alpha)
     logging.info('character alphabet: %s', alpha)
 
     # encode document chars as integer alphabet index values.
     encoded = [np.array([alpha.index(c) for c in doc]) for doc in corpus]
 
     def batch():
-        T, B = args.time, args.batch_size
+        T, B = kwargs['time'], kwargs['batch_size']
         inputs = np.zeros((T, B, len(alpha)), 'f')
         outputs = np.zeros((T, B), 'i')
         enc = np.random.choice(encoded)
@@ -106,40 +95,41 @@ def main(args):
         return [inputs, outputs]
 
     layers = [len(alpha)]
-    for l in args.layers:
-        layers.append(
-            dict(size=l, form=args.layer_type, activation=args.activation))
+    for l in kwargs['layers']:
+        layers.append(dict(size=l,
+                           form=kwargs['layer_type'],
+                           activation=kwargs['activation']))
     layers.append(len(alpha))
 
     exp = theanets.Experiment(theanets.recurrent.Classifier, layers=layers)
 
     exp.train(
         batch,
-        algo=args.algorithm,
-        patience=args.patience,
-        min_improvement=args.min_improvement,
-        validate_every=args.validate_every,
-        batch_size=args.batch_size,
-        train_batches=args.train_batches,
-        valid_batches=args.valid_batches,
-        learning_rate=args.learning_rate,
-        momentum=args.momentum,
-        nesterov=args.nesterov,
-        save_progress=args.save_progress,
-        save_every=args.save_every,
-        weight_l1=args.weight_l1,
-        weight_l2=args.weight_l2,
-        hidden_l2=args.hidden_l2,
-        hidden_l1=args.hidden_l1,
-        input_noise=args.input_noise,
-        input_dropouts=args.input_dropouts,
-        hidden_noise=args.hidden_noise,
-        hidden_dropouts=args.hidden_dropouts,
+        algo=kwargs['algorithm'],
+        patience=kwargs['patience'],
+        min_improvement=kwargs['min_improvement'],
+        validate_every=kwargs['validate_every'],
+        batch_size=kwargs['batch_size'],
+        train_batches=kwargs['train_batches'],
+        valid_batches=kwargs['valid_batches'],
+        learning_rate=kwargs['learning_rate'],
+        momentum=kwargs['momentum'],
+        nesterov=kwargs['nesterov'],
+        save_progress=kwargs['save_progress'],
+        save_every=kwargs['save_every'],
+        weight_l1=kwargs['weight_l1'],
+        weight_l2=kwargs['weight_l2'],
+        hidden_l2=kwargs['hidden_l2'],
+        hidden_l1=kwargs['hidden_l1'],
+        input_noise=kwargs['input_noise'],
+        input_dropouts=kwargs['input_dropouts'],
+        hidden_noise=kwargs['hidden_noise'],
+        hidden_dropouts=kwargs['hidden_dropouts'],
     )
 
-    if args.save_progress:
-        exp.save(args.save_progress)
+    if kwargs['save_progress']:
+        exp.save(kwargs['save_progress'])
 
 
 if __name__ == '__main__':
-   climate.call(main)
+   main()