eta=0.2,
k_eta=0.99998,
teaching=False)
print('out: ' + str(res_out))
print('goal: ' + str(res_goal))
res = [
1 if res_out[i] == res_goal[i] else 0 for i in range(0, mini_batch_size)
]
print(res)
res_out = net.SGD(training_data,
validation_data,
shared_image_data,
mini_batch_size,
epochs=1,
eta=0.2,
k_eta=0.99998,
teaching=False)
print('out: ' + str(res_out))
print('goal: ' + str(res_goal))
res = [
1 if res_out[i] == res_goal[i] else 0 for i in range(0, mini_batch_size)
]
print(res)
x = image_data[0]
num = 8
digit = x[num]
visual.show_image(digit)
visual.save_image_png(digit)
options = {'test_size': test_size, 'no_of_pipeline': no_of_pipeline}
return options
window1 = preestimator()
window2 = None # start off with 1 window open
while True: # Event Loop
window, event, values = sg.read_all_windows()
if event in (sg.WIN_CLOSED, 'Exit', 'Submit'):
window.close()
if window == window1:
break
elif window == window2:
window2 = None
if event == 'Estimator':
parsePreestimators()
window2 = estimator()
elif event == 'Submit':
parseEstimators()
#additional_param()
additional_parameters = additional_param()
path = main(additional_parameters)
pop = sg.Popup('Submit')
if pop == 'OK':
window.close
show_image(path)
def main(args, device):
# Initialize datasets and loaders
datasetGenerator = dataset.DatasetGenerator(
args.dataset, args.train_batch, args.test_batch, args.data_path,
args.trans_arg, args.transform, args.delta_path + args.noise + '.pt')
train_loader, validation_loader, test_loader = datasetGenerator.get_data_loaders(
)
# Initialize model
model = args.dataset if args.model is None else args.model
net = get_model(model).to(device)
if args.load_model is not None:
net.load_state_dict(
torch.load(args.model_path + args.load_model + '.pt'))
# Initialize attacks and trainer
trainer = Trainer(device, net)
pgd_max = attacks.PGDMax(device, args.epsilon, args.epsilon / 10,
args.iterations, args.restarts)
fgsm = attacks.FGSM(device, args.epsilon)
if args.todo == "train_nat":
# Train the network in a natural manner
trainer.train(train_loader, validation_loader, args.epoch,
args.model_path + args.save)
test_trained_model(net, args.model_path + args.save + '.pt', trainer,
pgd_max, fgsm, test_loader)
return
elif args.todo == "train_pgd":
# Train the network using adversarial training
trainer.train(train_loader,
validation_loader,
args.epoch,
args.model_path + args.save,
attack=pgd_max)
test_trained_model(net, args.model_path + args.save + '.pt', trainer,
pgd_max, fgsm, test_loader)
return
elif args.todo == "train_fgsm":
# Train the network using adversarial training
trainer.train(train_loader,
validation_loader,
args.epoch,
args.model_path + args.save,
attack=fgsm)
test_trained_model(net, args.model_path + args.save + '.pt', trainer,
pgd_max, fgsm, test_loader)
return
elif args.todo == "test":
# Test performance of trained model model
test_trained_model(net, args.model_path + args.load_model + '.pt',
trainer, pgd_max, fgsm, test_loader)
return
elif args.todo == "show_noise":
index = 32 # Index of image showed in the visualization of data augmentation techniques
images, _ = next(iter(train_loader))
noise = torch.load(args.delta_path + args.noise + '.pt')
# Multiply noise to ensure it is visible when plotted
for i in range(len(noise)):
noise[i] = noise[i].mul(100)
visual.show_images(noise, 10, 10, save=args.result_path + 'noise')
visual.show_images(images, 3, 3, save=args.result_path + 'images')
visual.show_image(images[index],
save=args.result_path + 'img_' + str(index))
visual.show_image(noise[index],
save=args.result_path + 'noise_' + str(index))
return
perturbation = NoiseGenerator(device, net, args.epsilon, args.iterations,
args.max_iter, args.train_step,
args.stop_error)
if args.todo == "sample_wise":
# Generate sample wise error minimizing noise
perturbation.sample_wise(train_loader,
save=args.delta_path + args.save)
return
elif args.todo == "class_wise":
# Generate class wise error minimizing noise
perturbation.class_wise(train_loader,
num_classes=10,
save=args.delta_path + args.save)
return
else:
raise NotImplementedError
from result_dir import make_dir_if_not_exists
from visual import show_image
show_image('RESULT/2021_05_01_22_27_31/PLOT')
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('code', type=str)
parser.add_argument('-i', '--image', action='store_true')
parser.add_argument('-p', '--preview', action='store_true')
parser.add_argument('-j', '--use-json', action='store_true')
args = parser.parse_args()
if args.use_json:
with open(args.code) as f:
code = json.load(f)
else:
code_parts = args.code.split('-')
code = {'N': int(code_parts[0]), # grid size NxN
'S': code_parts[1], # Ships
'T': code_parts[2]} # Attacks
ships = game.create_ships(code['S'])
attacks = game.create_attacks(code['T'])
result = game.eval_game(ships, attacks)
print(result)
if args.image or args.preview:
image = visual.draw_game((code['N'], code['N']), ships, attacks)
if args.image:
image.save('grid_battle_ship.jpg', quality=95)
if args.preview:
visual.show_image(image)