def train(name,
epoch=20000,
debug=False,
batch_size=128,
sample_iter=100,
model_save_iter=1000,
variation=True,
recover_dir=None,
test=True
):
experiment_dir = None
if not debug:
experiment_dir = EXPERIMENT_ROOT / name
experiment_dir.mkdir(parents=True, exist_ok=True)
prepareLogger(experiment_dir/'train_log.txt')
print('train')
else:
model_save_iter = -1
sample_iter = 1
epoch = 2
batch_size = 2
test=True
print('debug train')
loader = load_ferg()
pprl = Pprl(loader, variation=variation, debug=debug)
if recover_dir is not None:
pprl.load_weights(recover_dir)
pprl.train(experiment_dir=experiment_dir,
total_iter=epoch,
batch_size=batch_size,
sample_iter=sample_iter,
model_save_iter=model_save_iter)
if test:
pprl.evaluate(batch_size=batch_size)
def train(name,
epoch=20000,
debug=False,
batch_size=128,
sample_iter=100,
model_save_iter=1000,
variation=True,
recover_dir=None,
test=True
):
name = 'ae_gan' + name
experiment_dir = EXPERIMENT_ROOT / name
experiment_dir.mkdir(parents=True, exist_ok=True)
prepareLogger(experiment_dir/'train_log.txt')
print('train')
print(f'{name} aegan without reconstruction loss')
loader = load_ferg()
ae_gan = AeGan(loader, variation=variation, debug=debug)
if recover_dir is not None:
ae_gan.load_weights(recover_dir)
ae_gan.train(experiment_dir, total_iter=epoch, batch_size=batch_size, sample_iter=sample_iter, model_save_iter=model_save_iter)
if test:
acc_y_on_x = ae_gan.evaluate_y_on_x()
acc_p_on_x = ae_gan.evaluate_p_on_x()
print(f'acc_y_on_x = {acc_y_on_x}, acc_p_on_x={acc_p_on_x}')
def test(name, iter_no=100, debug=False, batch_size=128, epoch=20):
experiment_dir = EXPERIMENT_ROOT / name
prepareLogger()
print('test')
loader = load_ferg()
pprl = Pprl(loader, debug=debug)
pprl.load_weights(experiment_dir=experiment_dir, iter_no=iter_no)
pprl.evaluate(batch_size=batch_size, num_epochs=epoch)
def main(gpu, epoch, debug):
os.environ["CUDA_VISIBLE_DEVICES"] = gpu
loader = load_ferg()
ae = Ae(loader, debug=debug)
ae.train(num_epochs=epoch)
acc_y_on_x = ae.evaluate_y_on_x()
acc_p_on_x = ae.evaluate_p_on_x()
print(f'acc_y_on_x = {acc_y_on_x}, acc_p_on_x={acc_p_on_x}')
def test(name, iter_no=100, debug=False, num_epochs=20, batch_size=128):
experiment_dir = EXPERIMENT_ROOT / name
prepareLogger()
print('test')
loader = load_ferg()
aegan_qp = AeganQp(loader, debug=debug)
aegan_qp.load_weights(experiment_dir=experiment_dir, iter_no=iter_no)
aegan_qp.evaluate(batch_size=batch_size, num_epochs=num_epochs)
def test(name, iter_no=100, debug=False, num_epochs=20):
name = 'gan_qp' + name
experiment_dir = EXPERIMENT_ROOT / name
prepareLogger()
print('test')
loader = load_ferg()
gan_qp = GanQp(loader, debug=debug)
gan_qp.load_weights(experiment_dir=experiment_dir, iter_no=iter_no)
acc_y_on_x = gan_qp.evaluate_y_on_x(num_epochs=num_epochs)
acc_p_on_x = gan_qp.evaluate_p_on_x(num_epochs=num_epochs)
print(f'acc_y_on_x = {acc_y_on_x}, acc_p_on_x={acc_p_on_x}')
def train(name,
epoch=20000,
debug=False,
batch_size=128,
sample_iter=100,
model_save_iter=1000,
variation=True,
recover_dir=None,
test=True):
experiment_dir = None
if not debug:
experiment_dir = EXPERIMENT_ROOT / name
experiment_dir.mkdir(parents=True, exist_ok=True)
prepareLogger(experiment_dir / 'train_log.txt')
print('train')
else:
model_save_iter = -1
sample_iter = -1
epoch = 2
batch_size = 2
test = False
print('debug train')
loader = load_ferg()
gan_qp = GanQp(loader, variation=variation, debug=debug)
if recover_dir is not None:
gan_qp.load_weights(recover_dir)
gan_qp.train(experiment_dir=experiment_dir,
total_iter=epoch,
batch_size=batch_size,
sample_iter=sample_iter,
model_save_iter=model_save_iter)
if test:
acc_y_on_x = gan_qp.evaluate_y_on_x()
acc_p_on_x = gan_qp.evaluate_p_on_x()
print(f'acc_y_on_x = {acc_y_on_x},\
acc_p_on_x={acc_p_on_x}')
def show():
print('show')
loader = load_ferg()
gan_qp = GanQp(loader)
gan_qp.summary()
def show():
print('show')
loader = load_ferg()
pprl = Pprl(loader)
pprl.summary()
def show():
print('show')
loader = load_ferg()
ae_gan = AeGan(loader)
ae_gan.summary()
