def train(self):
error = tensor.neq(self.y.flatten(), self.y_hat.flatten() > 0.5).mean()
error.name = 'error'
self.error = error
experiment = Experiment(self.params['model_name'], self.train_stream)
experiment.cost = self.cost
experiment.set_adam(self.params['learning_rate'])
experiment.add_printing(after_epoch=True)
experiment.monitor_f_score(self.y,
self.y_hat,
average='macro',
threshold=self.params['threshold'])
experiment.monitor_auc_score(self.y, self.y_hat, average='macro')
experiment.add_timing()
experiment.extensions.append(
EarlyStopping('dev_f_score',
epochs=self.params['n_epochs'],
choose_best=max))
weights = VariableFilter(theano_name='W')(experiment.cg.variables)
experiment.regularize_max_norm(self.params['max_norms'], weights)
experiment.apply_dropout(self.params['dropout'])
experiment.track_best('dev_f_score',
save_path=self.params['model_name'] + '.tar',
choose_best=max)
experiment.track_best('dev_cost',
save_path=self.params['model_name'] +
'_cost.tar')
experiment.plot_channels(channels=[
['tra_f_score', 'dev_f_score'],
['tra_cost', 'dev_cost'],
],
url_bokeh='http://localhost:5006/',
before_first_epoch=True,
after_epoch=True)
experiment.add_monitored_vars([error])
experiment.add_norm_grads_vars()
experiment.monitor_stream(self.train_stream,
prefix='tra',
after_epoch=True)
experiment.monitor_stream(self.dev_stream, prefix='dev')
self.experiment = experiment
print('# of params for the model: {0}'.format(
experiment.get_num_params()))
main_loop = experiment.get_main_loop()
if not os.path.isfile(self.params['model_name'] + '.tar'):
main_loop.run()
with open(self.params['model_name'] + '.tar', "rb") as f:
print('loading saved model...')
main_loop.model.set_parameter_values(load_parameters(f))
