Make logging output more relevant for modeling.

theanets/graph.py

@@ -260,7 +260,6 @@ def set_loss(self, *args, **kwargs):
         All parameters and keyword arguments are passed to :func:`add_loss`
         after clearing the current losses.
         '''
-        logging.info('resetting model losses!')
         self.losses = []
         self.add_loss(*args, **kwargs)
 
@@ -441,9 +440,13 @@ def build_graph(self, regularizers=()):
             A list of updates that should be performed by a theano function that
             computes something using this graph.
         '''
-        regularizers = tuple(regularizers)
         key = self._hash(regularizers)
         if key not in self._graphs:
+            logging.info('building computation graph')
+            for loss in self.losses:
+                loss.log()
+            for reg in regularizers:
+                reg.log()
             outputs = {i.name: i for i in self.inputs}
             updates = []
             for layer in self.layers:
@@ -543,6 +546,7 @@ def feed_forward(self, x, **kwargs):
         if key not in self._functions:
             outputs, updates = self.build_graph(regs)
             labels, exprs = list(outputs.keys()), list(outputs.values())
+            logging.info('compiling feed_forward function')
             self._functions[key] = (labels, theano.function(
                 self.inputs, exprs, updates=updates))
         labels, f = self._functions[key]

theanets/losses.py

@@ -53,8 +53,6 @@ def __init__(self, target, weight=1., weighted=False, output_name='out'):
         if ':' not in self.output_name:
             self.output_name += ':out'
 
-        self.log()
-
     @property
     def variables(self):
         '''A list of Theano variables used in this loss.'''

theanets/regularizers.py

@@ -152,7 +152,6 @@ class Regularizer(util.Registrar(str('Base'), (), {})):
     def __init__(self, pattern=None, weight=0.):
         self.pattern = pattern
         self.weight = weight
-        self.log()
 
     def log(self):
         '''Log some diagnostic info about this regularizer.'''
@@ -396,7 +395,6 @@ def loss(self, layers, outputs):
             pattern = ns[0] if len(ns) == 1 else '{' + ','.join(ns) + '}'
         matches = util.outputs_matching(outputs, pattern)
         hiddens = [expr for _, expr in matches]
-        logging.info('found %s matching %s', hiddens, pattern)
         if not hiddens:
             return 0
         return sum(abs(h).mean() for h in hiddens) / len(hiddens)
@@ -490,9 +488,7 @@ def loss(self, layers_, outputs):
             ns = [l.output_name() for l in layers_[1:-1]]
             pattern = ns[0] if len(ns) == 1 else '{' + ','.join(ns) + '}'
         targets = [expr for _, expr in util.outputs_matching(outputs, pattern)]
-        logging.info('found %s matching %s', targets, pattern)
         if not targets:
-            raise ValueError
             return 0
         wrt = [l.input for l in layers_
                if isinstance(l, layers.Input) and