def on_train_end(self, logs={}):
if self.reporting_freq:
self.report()
printflush('\nFFNN Training Finished! (%s Batches in total)\n'
% '{:,}'.format(self.latest_batch))
if self.latest_val_acc is None:
printflush('Training Accuracy (approx) = %s%%\n'
% '{:.1f}'.format(100. * self.approx_train_acc_in_latest_epoch))
else:
printflush('Best trained FFNN (with lowest Validation Loss) is from epoch #%s'
% '{:,}'.format(self.best_model_epoch))
printflush('Training Accuracy (approx) = %s%%, Validation Accuracy = %s%%\n'
% ('{:.1f}'.format(100. * self.best_model_train_acc),
'{:.1f}'.format(100. * self.latest_val_acc)))
def report(self, batch_in_epoch=None):
if batch_in_epoch:
batch_text = ' Batch ' + '{0:03}'.format(batch_in_epoch)
else:
batch_text = ''
if self.latest_val_acc is None:
val_acc_text = ''
else:
val_acc_text = 'ValidAcc (prev epoch)=' + '{:.1f}'.format(100. * self.latest_val_acc) + '%'
printflush('Epoch %s%s: TrainAcc (approx)=%s%%, %s'
% ('{:,}'.format(self.latest_epoch),
batch_text,
'{:.1f}'.format(100. * self.approx_train_acc_in_latest_epoch),
val_acc_text), end='\r')
self.train_losses_curve_data_source.data['x'] = self.batches
self.train_losses_curve_data_source.data['y'] = self.train_losses
self.valid_losses_curve_data_source.data['x'] = self.batches
self.valid_losses_curve_data_source.data['y'] = self.val_losses
self.bokeh_session.store_objects(self.train_losses_curve_data_source, self.valid_losses_curve_data_source)
def on_train_begin(self, logs={}):
printflush('\nFFNN Training Progress')
printflush('______________________')
def __init__(self, reporting_freq=False, plot_title='Neural Network Learning Curves', bokeh_output='server'):
super(Callback, self).__init__()
self.latest_epoch = -1
self.latest_batch = -1
self.batches = []
self.train_losses = []
self.approx_train_acc_in_latest_epoch = 0.
self.val_losses = []
self.latest_val_acc = None
self.min_val_loss = inf
self.best_model = None
self.best_model_epoch = None
self.best_model_train_acc = None
self.best_model_val_acc = None
self.reporting_freq = reporting_freq
printflush('\nConnecting to Bokeh Server for live Learning Curves plotting...\n')
try:
output_server('')
self.bokeh_session = cursession()
self.fig = figure(title=plot_title,
x_axis_label='# of Training Data Batches', y_axis_label='Loss',
plot_height=680, plot_width=880)
self.fig.line((), (), name='TrainLoss', legend='Training Loss')
self.fig.circle((), (), name='ValidLoss', legend='Validation Loss', color='red')
show(self.fig)
self.train_losses_curve_data_source = self.fig.select(dict(name='TrainLoss'))[0].data_source
self.valid_losses_curve_data_source = self.fig.select(dict(name='ValidLoss'))[0].data_source
printflush('\nConnecting to Bokeh Server for live Learning Curves plotting... done!\n')
except:
printflush('\nBokeh Server Connection *FAILED!*')
printflush('Please make sure Bokeh package is already installed in Python, and')
printflush('please open a new Command-Line Terminal window\n (separate from this Terminal window)')
printflush(' and run the following command firs to launch Bokeh Server:')
printflush(' bokeh-server --backend=memory\n')
_exit(0)
