def validate(values):
"""Validate the values in the input widget
Parameters
----------
values : list
inputs from layout widgets based on input index
"""
v = ["" for k in inputmap.keys()]
#v[0] = checkpath(values[0])
v[1] = checkpath(values[1])
v[2] = checkpath(values[2])
if values[3] not in ['.tif', '.nd2', '.lsm', '.czi']:
v[3] = "unknown file extension"
v[4] = checktype(values[4], int, "Can't convert to int")
if not v[4]:
if int(values[4]) <= 0:
v[4] = "Channels start at 1"
v[5] = checktype(values[5], int, "Can't convert to int")
v[5] = checktype(values[5], int, "Can't convert to int")
v[6] = checktype(values[6], int, "Can't convert to int")
v[7] = checktype(values[7], int, "Can't convert to int")
v[8] = checktype(values[8], int, "Can't convert to int")
v[9] = checktype(values[9], int, "Can't convert to int")
v[10] = checktype(values[10], int, "Can't convert to int")
v[11] = checktype(values[11], float, "Can't convert to float")
v[12] = checktype(values[12], float, "Can't convert to float")
v[13] = checktype(values[13], float, "Can't convert to float")
return v
def __init__(self, FLAGS):
self.FLAGS = FLAGS
self.model_out_dir = "../Model"
self.log_dir = "../Log"
utils.checkpath(self.model_out_dir)
utils.checkpath(self.log_dir)
self.callback = TensorBoard(self.log_dir)
shutil.copy('./bin/plotboard.py', self.log_dir)
shutil.copy('./bin/run.bat', self.log_dir)
def main(args):
# Create output directory
checkpath(args.out_path)
# list all files end with jpg, args.in_path should contain a folder with name imgs for jpg files
# and a folder with name pointlines for pkl files.
input_paths_train = sorted(glob(os.path.join(args.in_path, '{}/*.jpg'.format("imgs/train"))))
input_paths_test = sorted(glob(os.path.join(args.in_path, '{}/*.jpg'.format("imgs/test"))))
if len(input_paths_train) == 0 or len(input_paths_test) == 0:
raise Exception("No images are found in {}".format(args.in_path))
# filter out outdoor images labeled by us, this is optinal
outdoor_list_train = os.path.join(args.in_path, 'outdoor_list_train.txt')
outdoor_list_test = os.path.join(args.in_path, 'outdoor_list_test.txt')
with open(outdoor_list_train,'r') as fp:
outdoor_train = fp.readlines()
# the last one doesn't contain the '\n' so processed seperately
outdoor_names_train = [ p[:-2] for p in outdoor_train if len(p) > 9]
outdoor_names_train += [outdoor_train[-1][:-1]]
print(len(outdoor_names_train))
with open(outdoor_list_test,'r') as fp:
outdoor_test = fp.readlines()
outdoor_names_test = [ p[:-2] for p in outdoor_test if len(p) > 9]
outdoor_names_test += [outdoor_test[-1][:-1]]
data = []
data_train = []
data_test = []
# save training data
for fname in input_paths_train:
# retrive the index name of the image
basename = fname.split('/')[-1].split('.')[0]
if basename not in outdoor_names_train:
filename = os.path.join(args.in_path, 'pointlines/{}.pkl'.format(basename))
_ = process(args.uni_wf, args.img_size, args.out_path, filename, mode='Train')
item = (filename, 'Train')
data_train.append(item)
data.append(item)
# save test data
for fname in input_paths_test:
# retrieve the index name of the image
basename = fname.split('/')[-1].split('.')[0]
if basename not in outdoor_names_test:
filename = os.path.join(args.in_path, 'pointlines/{}.pkl'.format(basename))
_ = process(args.uni_wf, args.img_size, args.out_path, filename, mode='Test')
item = (filename, 'Test')
data_test.append(item)
data.append(item)
print("The length of the dataset is: {}".format(len(data)))
print("The length of the training set is: {}".format(len(data_train)))
print("The length of the test set is: {}".format(len(data_test)))
print('finished preprocessing data')
def process(uni_wf, out_size, out_path, filename, mode):
# get output paths for preprocessed imgs and wireframes
basename = filename.split('/')[-1].split('.')[0]
wf_dir_train = os.path.join(out_path, 'wireframes/train')
checkpath(wf_dir_train)
img_dir_train = os.path.join(out_path, 'images/train')
checkpath(img_dir_train)
wf_dir_test = os.path.join(out_path, 'wireframes/test')
checkpath(wf_dir_test)
img_dir_test = os.path.join(out_path, 'images/test')
checkpath(img_dir_test)
with open(filename, 'rb') as f:
target = pickle.load(f, encoding='latin1')
img = target['img']
h, w, _ = img.shape
img_size = np.array((w, h))
img = cv2.resize(img, (out_size, out_size), interpolation=cv2.INTER_AREA)
if mode == 'Train':
img_dir = os.path.join(img_dir_train, '{}.png'.format(basename))
elif mode == 'Test':
img_dir = os.path.join(img_dir_test, '{}.png'.format(basename))
cv2.imwrite(img_dir, img)
points = target['points']
lines = target['lines']
wf = np.zeros((out_size, out_size))
for i, j in lines:
start = np.array( points[i] ) * out_size / img_size
end = np.array( points[j] ) * out_size / img_size
if uni_wf:
# use unified intensity
dist = 1
else:
# different intensity represents different dists, optional and haven't been thoroughly tested
dist = np.linalg.norm(end - start) / (out_size * np.sqrt(2))
if dist < 0.1:
dist = 0.2
elif dist > 0.5:
dist = 1
else:
dist = dist * 2
# haven't experimented with antialiased lines, user can optinally try lineType=cv2.LINE_AA
wf = cv2.line(wf, intx(start, out_size), intx(end, out_size), 255 * dist, 1, lineType=cv2.LINE_8)
if mode == 'Train':
save_dir = os.path.join(wf_dir_train, '{}.png'.format(basename))
elif mode == 'Test':
save_dir = os.path.join(wf_dir_test, '{}.png'.format(basename))
cv2.imwrite(save_dir, wf)
return wf
def __init__(self, FLAGS, LOGGER):
self.input_size = (60, 60) # image (input the GAN) size, must be times of 2^4
self.class_num = 36
self.FLAGS = FLAGS
self.val_ratio = FLAGS.val_ratio
self.batch_size = FLAGS.batch_size
#%% input dirs
self.data_dir = os.path.join('..', 'Data')
self.train_dir = os.path.join('..', 'Data', 'DataTrain')
# self.test_dir = os.path.join('..', 'Data', 'DataTest')
#%% output dirs
self.split_record_dir = os.path.join(LOGGER.log_dir, 'SplitRecord')
utils.checkpath(self.split_record_dir)
#%% record data list and split
self._read_data_list()
#%% implement batch fetcher
self.train_batch_fetcher = TrainBatchFetcher(self.train_names, self.batch_size)
def __init__(self, FLAGS):
self.FLAGS = FLAGS
self.img_out_dir = "../TrainResult"
self.model_out_dir = "../Model"
self.log_dir = "../Log"
utils.checkpath(self.img_out_dir)
utils.checkpath(self.model_out_dir)
utils.checkpath(self.log_dir)
self.callback = TensorBoard(self.log_dir)
import numpy as np
from PIL import Image, ImageDraw, ImageFont
from skimage import filters
from skimage.morphology import disk,square
from skimage.measure import regionprops
import skimage.morphology as sm
import utils
import time
time_start=time.time()
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
#%%
inputdir = os.path.join('.', 'TestData')
outputdir = os.path.join('.', 'TestResults')
utils.checkpath(outputdir)
filenames = utils.all_files_under(inputdir)
#%% load model
modelfile = os.path.join('.','Pretrained','Model.json')
weightsfile = os.path.join('.','Pretrained','Weights.h5')
model = utils.loadmodel(modelfile, weightsfile)
all_num = len(filenames)
pre_num = 0
acc_num = 0
setFont = ImageFont.truetype('C:/windows/fonts/Arial.ttf', 60)
fillColor = "#000000"
REFSTR = '0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ'
def fit_mlh(model, problems, data, name,
fixed={}, fitting={}, niter=5,
outdir='.', method='Nelder-Mead', save=True, quiet=False):
"""Use maximum likelihood to fit CHASE model"""
sim_id = sim_id_str(name, fixed, fitting)
checkpath(outdir)
cols = ['iteration', 'success', 'nllh', 'k', 'N', 'bic']
# get range of thetas for grid search
thetas = filter(lambda k: k.count('theta') > 0, fitting.keys())
if len(thetas) > 0:
theta_min, theta_max = fitting['theta']
theta_prod = map(list, list(product(range(theta_min, theta_max + 1), repeat=len(thetas))))
cols += thetas
else:
theta_prod = [[fixed['theta']]]
cols += ['theta']
rest = filter(lambda p: p.count('theta')==0, fitting.keys())
rest.sort()
cols += rest
# determine number of parameters and observations
k = len(fitting)
N = data.shape[0]
# create fit table
arr = []
for i in range(niter):
for th in theta_prod:
arr.append([i, np.nan, np.nan, k, N, np.nan] + th + [np.nan for _ in range(k - len(thetas))])
fitdf = pd.DataFrame(arr, columns=cols)
# iterate through parameter combinations
for i, row in fitdf.iterrows():
# update pars with current values of theta
pars = deepcopy(fixed)
for th in thetas:
pars[th] = row[th]
pars['fitting'] = OrderedDict([(p, fitting[p]) for p in rest])
# if theta=1, can't fit tau
if len(thetas)==1 and row[th]==1 and 'tau' in pars['fitting'] and 'stepsize' not in fixed:
del pars['fitting']['tau']
init = []
for p in pars['fitting']:
# if fitting normal stopping distribution, initialize at mean
if p=='mu':
init.append(data.samplesize.mean())
else:
init.append(uniform(fitting[p][0], fitting[p][1]))
# fit!
f = minimize(model.nloglik_opt, init, (problems, data, pars,),
method=method, options={'ftol': .001})
fitdf.ix[i,'success'] = f['success']
fitdf.ix[i,'nllh'] = f['fun']
fitdf.ix[i,'bic'] = bic(f['fun'], k, N)
for v, p in enumerate(pars['fitting'].keys()):
fitdf.ix[i,p] = f['x'][v]
if not quiet:
print sim_id
print '%s/%s' % (i, fitdf.shape[0])
print '%s: %s' % (thetas, row[thetas].values)
print fitdf.ix[i]
# save the table
if save: fitdf.to_csv('%s/%s.csv' % (outdir, sim_id))
return fitdf
def main(args):
# set which gpu(s) to use, should set PCI_BUS_ID first
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
num_gpus = (len(args.gpu) + 1) // 2
# create model directories
checkpath(args.modelG_path)
checkpath(args.modelD_path)
# tensorboard writer
checkpath(args.log_path)
writer = SummaryWriter(args.log_path)
# load data
data_loader, num_train = get_loader(args,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
training=True)
data_loader_val, num_test = get_loader(args,
batch_size=args.val_bs,
shuffle=False,
num_workers=args.num_workers,
training=False)
print('Finished data loading')
print("The length of the train set is: {}".format(num_train))
print("The length of the test set is: {}".format(num_test))
colorguide = True
if args.nocolor:
colorguide = False
# loss multipliers
lambdas = [
args.lambda_imgl1, args.lambda_wfl1, args.lambda_ssim,
args.lambda_color
]
lambda_perceptual = args.lambda_perceptual
# Generator
netG = Generator(lambdas=lambdas,
colorguide=colorguide,
input_nc=1,
output_nc=1)
if num_gpus > 1:
# multi-gpu training with synchonized batchnormalization
# make sure enough number of gpus are available
assert (torch.cuda.device_count() >= num_gpus)
# since we have set CUDA_VISIBLE_DEVICES to avoid some invalid device id issues
netG = DataParallelWithCallback(
netG, device_ids=[i for i in range(num_gpus)])
netG_single = netG.module
else:
# single gpu training
netG_single = netG
# Discriminator
netD = NLayerDiscriminator(input_nc=4, n_layers=4)
if num_gpus > 1:
netD = DataParallelWithCallback(
netD, device_ids=[i for i in range(num_gpus)])
netD_single = netD.module
else:
netD_single = netD
# print(netG_single)
# print(netD_single)
if args.pretrained and args.netG_path != '' and args.netD_path != '':
netG_single.load_state_dict(torch.load(args.netG_path))
netD_single.load_state_dict(torch.load(args.netD_path))
# Right now we only support gpu training
if torch.cuda.is_available():
netG = netG.cuda()
netD = netD.cuda()
# define the perceptual loss, place outside the forward func in G for better multi-gpu training
Ploss = PNet()
if num_gpus > 1:
Ploss = DataParallelWithCallback(
Ploss, device_ids=[i for i in range(num_gpus)])
if torch.cuda.is_available():
Ploss = Ploss.cuda()
# setup optimizer
lr = args.learning_rate
optimizerD = optim.Adam(netD_single.parameters(),
lr=lr,
betas=(args.beta1, 0.999))
schedulerD = ReduceLROnPlateau(optimizerD,
factor=0.7,
patience=10,
mode='min',
min_lr=1e-06)
optimizerG = optim.Adam(netG_single.parameters(),
lr=lr,
betas=(args.beta1, 0.999))
schedulerG = ReduceLROnPlateau(optimizerG,
factor=0.7,
patience=10,
mode='min',
min_lr=1e-06)
for epoch in range(args.num_epochs):
# switch to train mode
netG.train()
netD.train()
for i, (img_real, wf_real, color_real) in enumerate(data_loader, 0):
img_real = img_real.cuda()
wf_real = wf_real.cuda()
color_real = color_real.cuda()
# Update D network, we freeze parameters in G to save memory
for p in netG_single.parameters():
p.requires_grad = False
for p in netD_single.parameters():
p.requires_grad = True
# if using TTUR, D can be trained multiple steps per G step
for _ in range(args.D_steps):
optimizerD.zero_grad()
# train with real
real_AB = torch.cat((img_real, wf_real), 1)
errD_real = 0.5 * netD(trainG=False,
trainReal=True,
real_AB=real_AB,
fake_AB=None).sum()
errD_real.backward()
# train with fake
img_fake, wf_fake, _, _, _, _, _ = netG(trainG=False,
img_real=None,
wf_real=wf_real,
color_real=color_real)
fake_AB = torch.cat((img_fake, wf_fake), 1)
errD_fake = 0.5 * netD(trainG=False,
trainReal=False,
real_AB=None,
fake_AB=fake_AB).sum()
errD_fake.backward()
errD = errD_real + errD_fake
optimizerD.step()
del img_fake, wf_fake, fake_AB, real_AB, errD_real, errD_fake
iterations_before_epoch = epoch * len(data_loader)
writer.add_scalar('D Loss', errD.item(),
iterations_before_epoch + i)
del errD
# Update G network, we freeze parameters in D to save memory
for p in netG.parameters():
p.requires_grad = True
for p in netD.parameters():
p.requires_grad = False
optimizerG.zero_grad()
img_fake, wf_fake, lossG, wf_ssim, img_l1, color_l1, wf_l1 = netG(
trainG=True,
img_real=img_real,
wf_real=wf_real,
color_real=color_real)
ploss = Ploss(img_fake, img_real.detach()).sum()
fake_AB = torch.cat((img_fake, wf_fake), 1)
lossD = netD(trainG=True,
trainReal=False,
real_AB=None,
fake_AB=fake_AB).sum()
errG = (lossG.sum() + lambda_perceptual * ploss + lossD)
errG.backward()
optimizerG.step()
del color_real, fake_AB, lossG, errG
if args.nocolor:
print(
'Epoch: [{}/{}] Iter: [{}/{}] PercLoss : {:.4f} ImageL1 : {:.6f} WfL1 : {:.6f} WfSSIM : {:.6f}'
.format(epoch, args.num_epochs, i, len(data_loader),
ploss.item(),
img_l1.sum().item(),
wf_l1.sum().item(),
num_gpus + wf_ssim.sum().item()))
else:
print(
'Epoch: [{}/{}] Iter: [{}/{}] PercLoss : {:.4f} ImageL1 : {:.6f} WfL1 : {:.6f} WfSSIM : {:.6f} ColorL1 : {:.6f}'
.format(epoch, args.num_epochs, i, len(data_loader),
ploss.item(),
img_l1.sum().item(),
wf_l1.sum().item(),
num_gpus + wf_ssim.sum().item(),
color_l1.sum().item()))
writer.add_scalar('Color Loss',
color_l1.sum().item(),
iterations_before_epoch + i)
# tensorboard log
writer.add_scalar('G Loss', lossD.item(),
iterations_before_epoch + i)
writer.add_scalar('Image L1 Loss',
img_l1.sum().item(), iterations_before_epoch + i)
writer.add_scalar('Wireframe MSSSIM Loss',
num_gpus + wf_ssim.sum().item(),
iterations_before_epoch + i)
writer.add_scalar('Wireframe L1',
wf_l1.sum().item(), iterations_before_epoch + i)
writer.add_scalar('Image Perceptual Loss', ploss.item(),
iterations_before_epoch + i)
del wf_ssim, ploss, img_l1, color_l1, wf_l1, lossD
with torch.no_grad():
# show generated tarining images in tensorboard
if i % args.val_freq == 0:
real_img = vutils.make_grid(
img_real.detach()[:args.val_size],
normalize=True,
scale_each=True)
writer.add_image('Real Image', real_img,
(iterations_before_epoch + i) //
args.val_freq)
real_wf = vutils.make_grid(
wf_real.detach()[:args.val_size],
normalize=True,
scale_each=True)
writer.add_image('Real Wireframe', real_wf,
(iterations_before_epoch + i) //
args.val_freq)
fake_img = vutils.make_grid(
img_fake.detach()[:args.val_size],
normalize=True,
scale_each=True)
writer.add_image('Fake Image', fake_img,
(iterations_before_epoch + i) //
args.val_freq)
fake_wf = vutils.make_grid(
wf_fake.detach()[:args.val_size],
normalize=True,
scale_each=True)
writer.add_image('Fake Wireframe', fake_wf,
(iterations_before_epoch + i) //
args.val_freq)
del real_img, real_wf, fake_img, fake_wf
del img_real, wf_real, img_fake, wf_fake
# do checkpointing
if epoch % args.save_freq == 0 and epoch > 0:
torch.save(netG_single.state_dict(),
'{}/netG_epoch_{}.pth'.format(args.modelG_path, epoch))
torch.save(netD_single.state_dict(),
'{}/netD_epoch_{}.pth'.format(args.modelD_path, epoch))
# validation
with torch.no_grad():
netG_single.eval()
# since we use a realtively large validation batchsize, we don't go through the who test set
(img_real, wf_real, color_real) = next(iter(data_loader_val))
img_real = img_real.cuda()
wf_real = wf_real.cuda()
color_real = color_real.cuda()
img_fake, wf_fake, _, _, _, _, _ = netG_single(
trainG=False,
img_real=None,
wf_real=wf_real,
color_real=color_real)
# update lr based on the validation perceptual loss
val_score = Ploss(img_fake.detach(), img_real.detach()).sum()
schedulerG.step(val_score)
schedulerD.step(val_score)
print('Current lr: {:.6f}'.format(
optimizerG.param_groups[0]['lr']))
real_img = vutils.make_grid(img_real.detach()[:args.val_size],
normalize=True,
scale_each=True)
writer.add_image('Test: Real Image', real_img, epoch)
real_wf = vutils.make_grid(wf_real.detach()[:args.val_size],
normalize=True,
scale_each=True)
writer.add_image('Test: Real Wireframe', real_wf, epoch)
fake_img = vutils.make_grid(img_fake.detach()[:args.val_size],
normalize=True,
scale_each=True)
writer.add_image('Test: Fake Image', fake_img, epoch)
fake_wf = vutils.make_grid(wf_fake.detach()[:args.val_size],
normalize=True,
scale_each=True)
writer.add_image('Test: Fake Wireframe', fake_wf, epoch)
netG_single.train()
del img_real, real_img, wf_real, real_wf, img_fake, fake_img, wf_fake, fake_wf
# close tb writer
writer.close()
def single_ac_train(env,
actor,
critic,
store_path='./',
batch_size=32,
epsilon=0.01,
save_interval=1000,
update_interval=1000,
learning_starts=200,
memory_size=50000,
max_epoch=100000,
max_iter=10000):
event_path = os.path.join(store_path, 'actor_events')
actor_model_path = os.path.join(store_path, 'actor_models')
critic_model_path = os.path.join(store_path, 'critic_models')
checkpath(event_path)
checkpath(actor_model_path)
checkpath(critic_model_path)
actor.load_model(actor_model_path)
critic.load_model(critic_model_path)
summary_writer = SummaryWriter(event_path)
memory_buffer = Memory(memory_size)
results_buffer = ResultsBuffer()
states = env.reset()
for i in range(max_epoch):
states = env.reset()
episode_buffer = Episode_Record()
episode_buffer.append('state', states)
while True:
actions = actor.get_action(states, epsilon)
next_states, rewards, dones, info = env.step(actions)
episode_buffer.append('reward', rewards)
episode_buffer.append('action', actions)
if dones:
state_batch, reward_batch, action_batch = episode_buffer.dump()
score_batch = critic.get_target(state_batch)
target_batch = np.zeros_like(reward_batch)
target_batch[-1] = reward_batch[-1]
for idx in range(len(reward_batch) - 2, -1, -1):
target_batch[idx] = reward_batch[idx] + \
0.95 * target_batch[idx + 1]
global_step, critic_summary, advantage_batch = critic.update(
state_batch, target_batch)
# advantage_batch = np.zeros_like(reward_batch)
# R = 0.0
# for idx in range(len(reward_batch) - 1, -1, -1):
# R = R * 0.95 + reward_batch[idx]
# advantage_batch[idx] = R
# advantage_batch -= np.mean(advantage_batch)
# advantage_batch /= np.std(advantage_batch)
actor_summary = actor.update(state_batch, action_batch,
advantage_batch)
# results_buffer.add_summary(summary_writer, global_step)
actor.update_target()
critic.update_target()
# actor.save_model(actor_model_path, global_step)
# critic.save_model(critic_model_path, global_step)
print("Epoch {} earns a reward of {}.".format(
i, np.sum(reward_batch)))
break
else:
episode_buffer.append('state', next_states)
states = next_states
def main(args):
# by default we only consider single gpu inference
assert (len(args.gpu) == 1)
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
# load data
data_loader_val, num_test = get_loader(args,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
training=False)
print('finished data loading')
# Generator
colorguide = True
if args.nocolor:
colorguide = False
netG = Generator(lambdas=None,
colorguide=colorguide,
input_nc=1,
output_nc=1)
netG.load_state_dict(torch.load(args.model_path))
if torch.cuda.is_available():
netG = netG.cuda()
out_path = args.out_path
checkpath(out_path)
predictions_fid_real = []
predictions_fid_fake = []
fid_model = InceptionV3().cuda()
fid_model.eval()
Perceptual = PNet().cuda()
avg_ssim = 0
lpips = 0
# validate on test set, TODO: test with single color guide image
with torch.no_grad():
netG.eval()
for i, (img_real, wf_real,
color_real) in enumerate(data_loader_val, 0):
img_real = img_real.cuda()
wf_real = wf_real.cuda()
if colorguide:
color_real = color_real.cuda()
# in case we are in the last interation
batch_size = img_real.size(0)
img_fake, wf_fake, _, _, _, _, _ = netG(trainG=False,
img_real=None,
wf_real=wf_real,
color_real=color_real)
ssim_score = ssim(img_real, img_fake).item() * batch_size
avg_ssim += ssim_score
lpips += Perceptual(img_real, img_fake) * batch_size
# TODO: save generated wireframes
save_singleimages(img_fake, out_path, i * args.batch_size,
args.img_size)
pred_fid_real = fid_model(img_real)[0]
pred_fid_fake = fid_model(img_fake)[0]
predictions_fid_real.append(
pred_fid_real.data.cpu().numpy().reshape(batch_size, -1))
predictions_fid_fake.append(
pred_fid_fake.data.cpu().numpy().reshape(batch_size, -1))
print('SSIM: {:6f}'.format(avg_ssim / num_test))
print('LPIPS: {:6f}'.format(lpips / num_test))
predictions_fid_real = np.concatenate(predictions_fid_real, 0)
predictions_fid_fake = np.concatenate(predictions_fid_fake, 0)
fid = compute_fid_score(predictions_fid_fake, predictions_fid_real)
print('FID: {:6f}'.format(fid))