PySimpleGUI/Demo_Machine_Learning.py

58 lines
3.2 KiB
Python

import PySimpleGUI as sg
def MachineLearningGUI():
sg.SetOptions(text_justification='right')
form = sg.FlexForm('Machine Learning Front End', font=("Helvetica", 12)) # begin with a blank form
layout = [[sg.Text('Machine Learning Command Line Parameters', font=('Helvetica', 16))],
[sg.Text('Passes', size=(15, 1)), sg.Spin(values=[i for i in range(1, 1000)], initial_value=20, size=(6, 1)),
sg.Text('Steps', size=(18, 1)), sg.Spin(values=[i for i in range(1, 1000)], initial_value=20, size=(6, 1))],
[sg.Text('ooa', size=(15, 1)), sg.In(default_text='6', size=(10, 1)), sg.Text('nn', size=(15, 1)), sg.In(default_text='10', size=(10, 1))],
[sg.Text('q', size=(15, 1)), sg.In(default_text='ff', size=(10, 1)), sg.Text('ngram', size=(15, 1)), sg.In(default_text='5', size=(10, 1))],
[sg.Text('l', size=(15, 1)), sg.In(default_text='0.4', size=(10, 1)), sg.Text('Layers', size=(15, 1)), sg.Drop(values=('BatchNorm', 'other'),auto_size_text=True)],
[sg.Text('_' * 100, size=(65, 1))],
[sg.Text('Flags', font=('Helvetica', 15), justification='left')],
[sg.Checkbox('Normalize', size=(12, 1), default=True), sg.Checkbox('Verbose', size=(20, 1))],
[sg.Checkbox('Cluster', size=(12, 1)), sg.Checkbox('Flush Output', size=(20, 1), default=True)],
[sg.Checkbox('Write Results', size=(12, 1)), sg.Checkbox('Keep Intermediate Data', size=(20, 1))],
[sg.Text('_' * 100, size=(65, 1))],
[sg.Text('Loss Functions', font=('Helvetica', 15), justification='left')],
[sg.Radio('Cross-Entropy', 'loss', size=(12, 1)), sg.Radio('Logistic', 'loss', default=True, size=(12, 1))],
[sg.Radio('Hinge', 'loss', size=(12, 1)), sg.Radio('Huber', 'loss', size=(12, 1))],
[sg.Radio('Kullerback', 'loss', size=(12, 1)), sg.Radio('MAE(L1)', 'loss', size=(12, 1))],
[sg.Radio('MSE(L2)', 'loss', size=(12, 1)), sg.Radio('MB(L0)', 'loss', size=(12, 1))],
[sg.Submit(), sg.Cancel()]]
button, values = form.LayoutAndRead(layout)
del(form)
sg.SetOptions(text_justification='left')
return button, values
def CustomMeter():
# create the progress bar element
progress_bar = sg.ProgressBar(10000, orientation='h', size=(20,20))
# layout the form
layout = [[sg.Text('A custom progress meter')],
[progress_bar],
[sg.Cancel()]]
# create the form
form = sg.FlexForm('Custom Progress Meter')
# display the form as a non-blocking form
form.LayoutAndRead(layout, non_blocking=True)
# loop that would normally do something useful
for i in range(10000):
# check to see if the cancel button was clicked and exit loop if clicked
button, values = form.ReadNonBlocking()
if button == 'Cancel' or values == None:
break
# update bar with loop value +1 so that bar eventually reaches the maximum
progress_bar.UpdateBar(i+1)
# done with loop... need to destroy the window as it's still open
form.CloseNonBlockingForm()
if __name__ == '__main__':
CustomMeter()
MachineLearningGUI()