def main(args):
exec_time = datetime.now().strftime('%Y%m%d-%H%M%S')
tensorboard = TensorBoard(log_dir=f'log/vae/{exec_time}',
update_freq='batch')
new_model = args.new_model
epochs = int(args.epochs)
steps = int(args.steps)
# instantiate VAE
vae = VAE() if new_model else load_vae(WEIGHT_FILE_NAME)
# get training set and validation set generators
t_gen, v_gen = get_generators(INPUT_DIR_NAME, TV_RATIO)
# start training!
vae.train(t_gen,
v_gen,
epochs=epochs,
steps_per_epoch=steps,
validation_steps=int(steps * TV_RATIO),
workers=10,
callbacks=[tensorboard])
# save model weights
vae.save_weights(WEIGHT_FILE_NAME)
def main(args):
new_model = args.new_model
N = int(args.N)
epochs = int(args.epochs)
vae = VAE()
if not new_model:
try:
vae.set_weights('./vae/weights.h5')
except:
print("Either set --new_model or ensure ./vae/weights.h5 exists")
raise
try:
data, N = import_data(N)
except:
print('NO DATA FOUND')
raise
print('DATA SHAPE = {}'.format(data.shape))
for epoch in range(epochs):
print('EPOCH ' + str(epoch))
vae.train(data)
vae.save_weights('./vae/weights.h5')
def main(args):
new_model = args.new_model
S = int(args.S)
N = int(args.N)
batch = int(args.batch)
model_name = str(args.model_name)
print(args.alpha)
alpha = float(args.alpha)
vae = VAE()
if not new_model:
try:
vae.set_weights('./vae/' + model_name + '/' + model_name +
'_weights.h5')
except:
print("Either set --new_model or ensure ./vae/weights.h5 exists")
raise
else:
if os.path.isdir('./vae/' + model_name):
print("A model with this name already exists")
else:
os.mkdir('./vae/' + model_name)
os.mkdir('./vae/' + model_name + '/log/')
filelist = os.listdir(DIR_NAME)
filelist = [x for x in filelist if x != '.DS_Store' and x != '.gitignore']
filelist.sort()
for i in range(round(float(N - S) / batch)):
data = import_data(S + i * batch, S + (i + 1) * batch, filelist)
dataS = []
dataB = []
for d in data:
beta = alpha + np.random.rand() * (1 - alpha)
dataS.append(
cv2.resize(crop(d, alpha * beta),
dsize=(SCREEN_SIZE_X, SCREEN_SIZE_Y),
interpolation=cv2.INTER_CUBIC))
dataB.append(
cv2.resize(crop(d, beta),
dsize=(SCREEN_SIZE_X, SCREEN_SIZE_Y),
interpolation=cv2.INTER_CUBIC))
dataS = np.asarray(dataS)
dataB = np.asarray(dataB)
vae.train(dataS, dataB, model_name
) # uncomment this to train augmenting VAE, simple RNN (2)
#vae.train(np.vstack([dataS, dataB]), np.vstack([dataS, dataB]), model_name) # uncomment this to train simple VAE, RNN (1)
vae.save_weights('./vae/' + model_name + '/' + model_name +
'_weights.h5')
print('Imported {} / {}'.format(S + (i + 1) * batch, N))
def main(args):
new_model = args.new_model
S = int(args.S)
N = int(args.N)
batch = int(args.batch)
epochs = int(args.epochs)
model_name = str(args.model_name)
vae = VAE()
if not new_model:
try:
vae.set_weights('./vae/' + model_name + '_weights.h5')
except:
print("Either set --new_model or ensure ./vae/" + model_name +
"_weights.h5 exists")
raise
elif not os.path.isdir('./vae/' + model_name):
os.mkdir('./vae/' + model_name)
os.mkdir('./vae/' + model_name + '/log/')
filelist = os.listdir(DIR_NAME + model_name)
filelist = [x for x in filelist if x != '.DS_Store' and x != '.gitignore']
filelist.sort()
N = max(N, len(filelist))
for i in range(int(round(float(N - S) / batch) + 1)):
dataS, dataB = import_data(S + i * batch, S + (i + 1) * batch,
filelist, model_name)
for epoch in range(epochs):
vae.train(
dataS, dataB, model_name
) # uncomment this to train augmenting VAE, simple RNN (2)
#vae.train(np.vstack([dataS, dataB]), np.vstack([dataS, dataB]), model_name) # uncomment this to train simple VAE, RNN (1)
vae.save_weights('./vae/' + model_name + '/' + model_name +
'_weights.h5')
print('Imported {} / {}'.format(S + (i + 1) * batch, N))