anton/mlpractical
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Added ability to add additional monitoring channels during training.

Browse Source
This commit is contained in:
Matt Graham 2016-09-19 12:18:44 +01:00
parent d55311e501
commit 03f27cab75
1 changed files with 16 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
mlp/trainers.py
View File

@ -13,13 +13,16 @@ logger = logging.getLogger(__name__)
class Trainer(object): class Trainer(object):
def __init__(self, model, cost, learning_rule, train_dataset, def __init__(self, model, cost, learning_rule, train_dataset,
valid_dataset=None): valid_dataset=None, data_monitors=None):
self.model = model self.model = model
self.cost = cost self.cost = cost
self.learning_rule = learning_rule self.learning_rule = learning_rule
self.learning_rule.initialise(self.model.params) self.learning_rule.initialise(self.model.params)
self.train_dataset = train_dataset self.train_dataset = train_dataset
self.valid_dataset = valid_dataset self.valid_dataset = valid_dataset
self.data_monitors = OrderedDict([('cost', cost)])
if data_monitors is not None:
self.data_monitors.update(data_monitors)
def do_training_epoch(self): def do_training_epoch(self):
for inputs_batch, targets_batch in self.train_dataset: for inputs_batch, targets_batch in self.train_dataset:
@ -29,19 +32,23 @@ class Trainer(object):
activations, grads_wrt_outputs) activations, grads_wrt_outputs)
self.learning_rule.update_params(grads_wrt_params) self.learning_rule.update_params(grads_wrt_params)
def data_cost(self, dataset): def monitors(self, dataset, label):
cost = 0. data_mon_vals = OrderedDict([(key + label, 0.) for key
in self.data_monitors.keys()])
for inputs_batch, targets_batch in dataset: for inputs_batch, targets_batch in dataset:
activations = self.model.fprop(inputs_batch) activations = self.model.fprop(inputs_batch)
cost += self.cost(activations[-1], targets_batch) for key, data_monitor in self.data_monitors.items():
cost /= dataset.num_batches data_mon_vals[key + label] += data_monitor(
return cost activations[-1], targets_batch)
for key, data_monitor in self.data_monitors.items():
data_mon_vals[key + label] /= dataset.num_batches
return data_mon_vals
def get_epoch_stats(self): def get_epoch_stats(self):
epoch_stats = OrderedDict() epoch_stats = OrderedDict()
epoch_stats['cost(train)'] = self.data_cost(self.train_dataset) epoch_stats.update(self.monitors(self.train_dataset, '(train)'))
if self.valid_dataset is not None: if self.valid_dataset is not None:
epoch_stats['cost(valid)'] = self.data_cost(self.valid_dataset) epoch_stats.update(self.monitors(self.valid_dataset, '(valid)'))
epoch_stats['cost(param)'] = self.model.params_cost() epoch_stats['cost(param)'] = self.model.params_cost()
return epoch_stats return epoch_stats
@ -61,4 +68,4 @@ class Trainer(object):
stats = self.get_epoch_stats() stats = self.get_epoch_stats()
self.log_stats(epoch, epoch_time, stats) self.log_stats(epoch, epoch_time, stats)
run_stats.append(stats.values()) run_stats.append(stats.values())
return np.array(run_stats), stats.keys() return np.array(run_stats), {k: i for i, k in enumerate(stats.keys())}