def main():
args = parse_args()
vol = load_volume(args.input, stop=args.depth, ext='.bmp')
with h5py.File('downsampled_histogram.h5', 'r') as f:
hist = f['hist'][:]
hist = np.cumsum(hist)
vol = np.floor(255 * hist[vol])
vol[vol > 255] = 255
vol[vol < 0] = 0
vol = vol.astype(np.uint8)
if not os.path.isdir(args.output):
os.makedirs(args.output)
sums, counts = summate(integral_image(vol), np.asarray(args.window))
for t in np.arange(0, 1, 0.01):
seg = threshold_bradley_nd(vol,
t=t,
s=np.asarray(args.window),
sums=sums,
counts=counts)
dirname = os.path.join(args.output, '{0:.2f}'.format(t))
save_imgs(seg * 255, dirname)
def test(cfg: Namespace) -> None:
assert cfg.checkpoint not in [None, ""]
assert cfg.device == "cpu" or (cfg.device == "cuda"
and T.cuda.is_available())
exp_dir = ROOT_EXP_DIR / cfg.exp_name
os.makedirs(exp_dir / "out", exist_ok=True)
cfg.to_file(exp_dir / "test_config.json")
logger.info(f"[exp dir={exp_dir}]")
model = CAE()
model.load_state_dict(T.load(cfg.checkpoint))
model.eval()
if cfg.device == "cuda":
model.cuda()
logger.info(f"[model={cfg.checkpoint}] on {cfg.device}")
dataloader = DataLoader(dataset=ImageFolder720p(cfg.dataset_path),
batch_size=1,
shuffle=cfg.shuffle)
logger.info(f"[dataset={cfg.dataset_path}]")
loss_criterion = nn.MSELoss()
for batch_idx, data in enumerate(dataloader, start=1):
img, patches, _ = data
if cfg.device == "cuda":
patches = patches.cuda()
if batch_idx % cfg.batch_every == 0:
pass
out = T.zeros(6, 10, 3, 128, 128)
avg_loss = 0
for i in range(6):
for j in range(10):
x = patches[:, :, i, j, :, :].cuda()
y = model(x)
out[i, j] = y.data
loss = loss_criterion(y, x)
avg_loss += (1 / 60) * loss.item()
logger.debug("[%5d/%5d] avg_loss: %f", batch_idx, len(dataloader),
avg_loss)
# save output
out = np.transpose(out, (0, 3, 1, 4, 2))
out = np.reshape(out, (768, 1280, 3))
out = np.transpose(out, (2, 0, 1))
y = T.cat((img[0], out), dim=2)
save_imgs(
imgs=y.unsqueeze(0),
to_size=(3, 768, 2 * 1280),
name=exp_dir / f"out/test_{batch_idx}.png",
)
def train(args):
model = AutoencoderConv().cuda()
if args.load:
model.load_state_dict(torch.load(args.chkpt))
print("Loaded model from", args.chkpt)
model.train()
print("Done setup model")
dataset = ImageFolder720p(args.dataset_path)
dataloader = DataLoader(
dataset, batch_size=args.batch_size, shuffle=args.shuffle, num_workers=args.num_workers
)
print("Done setup dataloader: {len} batches of size {batch_size}".format(len=len(dataloader), batch_size=args.batch_size))
mse_loss = nn.MSELoss()
adam = torch.optim.Adam(model.parameters(), lr=args.learning_rate, weight_decay=1e-5)
sgd = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
optimizer = adam
for ei in range(args.resume_epoch, args.num_epochs):
for bi, (img, patches, _) in enumerate(dataloader):
avg_loss = 0
for i in range(6):
for j in range(10):
x = Variable(patches[:, :, i, j, :, :]).cuda()
y = model(x)
loss = mse_loss(y, x)
avg_loss += 0.6 * loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
print('[%3d/%3d][%5d/%5d] loss: %f' % (ei, args.num_epochs, bi, len(dataloader), avg_loss))
# save img
if bi % args.out_every == 0:
out = torch.zeros(6, 10, 3, 128, 128)
for i in range(6):
for j in range(10):
x = Variable(patches[0, :, i, j, :, :].unsqueeze(0)).cuda()
out[i, j] = model(x).cpu().data
out = np.transpose(out, (0, 3, 1, 4, 2))
out = np.reshape(out, (768, 1280, 3))
out = np.transpose(out, (2, 0, 1))
y = torch.cat((img[0], out), dim=2).unsqueeze(0)
save_imgs(imgs=y, to_size=(3, 768, 2 * 1280), name="out/{exp_name}/out_{ei}_{bi}.png".format(exp_name=args.exp_name, ei=ei, bi=bi))
# save model
if bi % args.save_every == args.save_every - 1:
torch.save(model.state_dict(), "checkpoints/{exp_name}/model_{ei}_{bi}.state".format(exp_name=args.exp_name, ei=ei, bi=bi))
torch.save(model.state_dict(), "checkpoints/{exp_name}/model_final.state".format(exp_name=args.exp_name))
def test(cfg: Namespace) -> None:
logger.info("=== Testing ===")
# initial setup
prologue(cfg)
model = CAE()
model.load_state_dict(torch.load(cfg.chkpt))
model.eval()
if cfg.device == "cuda":
model.cuda()
logger.info("Loaded model")
dataset = ImageFolder720p(cfg.dataset_path)
dataloader = DataLoader(dataset, batch_size=1, shuffle=cfg.shuffle)
logger.info("Loaded data")
loss_criterion = nn.MSELoss()
for batch_idx, data in enumerate(dataloader, start=1):
img, patches, _ = data
if cfg.device == 'cuda':
patches = patches.cuda()
if batch_idx % cfg.batch_every == 0:
pass
out = torch.zeros(6, 10, 3, 128, 128)
# enc = torch.zeros(6, 10, 16, 8, 8)
avg_loss = 0
for i in range(6):
for j in range(10):
x = Variable(patches[:, :, i, j, :, :]).cuda()
y = model(x)
out[i, j] = y.data
loss = loss_criterion(y, x)
avg_loss += (1 / 60) * loss.item()
logger.debug('[%5d/%5d] avg_loss: %f' %
(batch_idx, len(dataloader), avg_loss))
# save output
out = np.transpose(out, (0, 3, 1, 4, 2))
out = np.reshape(out, (768, 1280, 3))
out = np.transpose(out, (2, 0, 1))
y = torch.cat((img[0], out), dim=2)
save_imgs(
imgs=y.unsqueeze(0),
to_size=(3, 768, 2 * 1280),
name=f"../experiments/{cfg.exp_name}/out/test_{batch_idx}.png")
# final setup
epilogue(cfg)
def test(self, epoch):
self.net.eval()
pd_num = 0
path = os.path.join(self.pred, f'epoch_{epoch}')
if not os.path.exists(path):
os.mkdir(path)
with torch.no_grad():
for data in tqdm(self.testset):
res = self.net(data)
save_imgs(path, res, pd_num)
pd_num += 1
def main():
args = parse_args()
start = time.time()
print("Loading image volume")
vol = load_volume(args.input)
if args.show:
plt.imshow(vol[0], cmap='Greys_r')
plt.show()
print("Prepping threshold subvolume shape")
if len(args.shape) < len(vol.shape):
shape = list(vol.shape[:len(vol.shape) - len(args.shape)])
shape.extend(args.shape)
else:
shape = args.shape
if len(args.step) < len(shape):
step = list(vol.shape[:len(vol.shape) - len(args.step)])
step.extend(args.step)
else:
step = args.step
print("Thresholding subvolumes")
step = args.step
thresh = np.zeros(vol.shape)
for i in range(0, vol.shape[0], step[0] if step else vol.shape[0]):
endi = i + shape[0] if i + shape[0] < vol.shape[0] else vol.shape[0]
for j in range(0, vol.shape[1], step[1] if step else vol.shape[1]):
endj = j + shape[1] if j + shape[1] < vol.shape[1] else vol.shape[1]
for k in range(0, vol.shape[2], step[2] if step else vol.shape[2]):
endk = k + shape[2] if k + shape[2] < vol.shape[
2] else vol.shape[2]
subvol = np.copy(vol[i:endi, j:endj, k:endk])
subvol = threshold_bradley_nd(subvol,
s=(4, shape[1], shape[2]),
t=args.threshold)
subvol = np.abs(1 - subvol) if np.max(subvol) > 0 else subvol
subvol = binary_opening(subvol)
subvol = binary_fill_holes(subvol)
subvol[subvol > 0] = 1
thresh[i:endi, j:endj, k:endk] += subvol
if args.show:
plt.imshow(thresh[0], cmap='Greys_r')
plt.show()
thresh[thresh > 0] = 255
print("Saving segmentation")
save_imgs(thresh, args.output)
print("Running Time: {}s".format(time.time() - start))
def test(args):
model = AutoencoderConv().cuda()
model.load_state_dict(torch.load(args.chkpt))
model.eval()
print("Loaded model from", args.chkpt)
dataset = ImageFolder720p(args.dataset_path)
dataloader = DataLoader(dataset, batch_size=1, shuffle=args.shuffle)
print(f"Done setup dataloader: {len(dataloader)}")
mse_loss = nn.MSELoss()
for bi, (img, patches, path) in enumerate(dataloader):
if "frame_40" not in path[0]: continue
out = torch.zeros(6, 10, 3, 128, 128)
# enc = torch.zeros(6, 10, 16, 8, 8)
avg_loss = 0
for i in range(6):
for j in range(10):
x = Variable(patches[:, :, i, j, :, :]).cuda()
y = model(x)
# e = model.enc_x.data
# p = torch.tensor(np.random.permutation(e.reshape(-1, 1)).reshape(1, 16, 8, 8)).cuda()
# out[i, j] = model.decode(p).data
# enc[i, j] = model.enc_x.data
out[i, j] = y.data
loss = mse_loss(y, x)
avg_loss += 0.6 * loss.item()
print('[%5d/%5d] loss: %f' % (bi, len(dataloader), avg_loss))
# save output
out = np.transpose(out, (0, 3, 1, 4, 2))
out = np.reshape(out, (768, 1280, 3))
out = np.transpose(out, (2, 0, 1))
y = torch.cat((img[0], out), dim=2)
save_imgs(imgs=y.unsqueeze(0),
to_size=(3, 768, 2 * 1280),
name=f"./test/{args.out_dir}/test_{bi}.png")
break
def img_sample(self, iter_time, save_dir, batch_size=4):
imgs, clses, segs, irises, coordinates = self.data.train_random_batch_include_iris(
batch_size=batch_size)
feed = {
self.model.mask_tfph: segs,
self.model.cls_tfph: clses,
self.model.rate_tfph: 0.5, # rate: 1 - keep_prob
self.model.iden_model.img_tfph: irises,
self.model.iden_model.train_mode: False
}
samples = self.sess.run(self.model.g_sample, feed_dict=feed)
utils.save_imgs(img_stores=[segs, samples, imgs],
iter_time=iter_time,
save_dir=save_dir,
is_vertical=True)
def save(self):
torm = {'anime': ['imdb', 'artist', 'album', 'company'],
'books': ['artist', 'album', 'year', 'company', 'tube', 'imdb', 'scrs'],
'games': ['imdb', 'artist', 'album'],
'movies': ['artist', 'album', 'company'],
'music': ['scrs', 'imdb', 'company', 'tube', 'lngs'],
'pics': ['imdb', 'artist', 'album', 'year', 'company', 'tube'],
'tv': ['artist', 'album', 'company'],
'misc': ['scrs', 'imdb', 'artist', 'album', 'year', 'company', 'tube', 'lngs']
}
toadd = {'anime': ['year', 'tube', 'lngs'],
'books': ['lngs'],
'games': ['year', 'company', 'lngs', 'tube'],
'movies': ['imdb', 'year', 'tube', 'lngs'],
'music': ['album', 'artist', 'year'],
'pics': ['lngs'],
'tv': ['imdb', 'year', 'tube', 'lngs'],
'misc': [],
}
for k in torm.get(self.cleaned_data['section'], []):
del self.instance.attrs[k]
for k in toadd.get(self.cleaned_data['section'], []):
self.instance.attrs[k] = self.cleaned_data[k]
self.instance.section = self.cleaned_data['section']
if self.cleaned_data['section'] == 'misc':
self.instance.subcat = 'misc'
else:
self.instance.subcat = self.cleaned_data[self.cleaned_data['section']]
if self.cleaned_data['notify'] == 'yes':
if self.instance.author.attrs.has_key('notify'):
l = self.instance.author.attrs['notify']
l.append(self.instance.id)
self.instance.author.attrs['notify'] = l
else:
self.instance.author.attrs['notify'] = [self.instance.id]
if self.cleaned_data['featured'] == 'yes':
if self.instance.author.attrs.has_key('featured'):
l = self.instance.author.attrs['featured']
l.append(self.instance.id)
self.instance.author.attrs['featured'] = l
else:
self.instance.author.attrs['featured'] = [self.instance.id]
if self.cleaned_data['poster']:
#removing old image if new uploaded
drop_pics(self.instance)
if self.cleaned_data['poster']:
self.instance.poster = save_imgs(self.cleaned_data['poster'], self.instance.author.id, 'poster')
self.instance.title = self.cleaned_data['title']
self.instance.text = self.cleaned_data['text']
_save_torrent(self.instance, self.cleaned_data['torrent'])
if self.cleaned_data['tube']:
self.instance.attrs['tube'] = self.cleaned_data['tube']
_update_notification_delivery_queue(self.cleaned_data['tags'], self.instance.author)
self.instance.save()
self.instance.tags = self.cleaned_data['tags']
def test(cfg):
os.makedirs(f"./test/{cfg['exp_name']}", exist_ok=True)
model = CAE().cuda()
model.load_state_dict(torch.load(cfg['chkpt']))
model.eval()
logger.info("Loaded model from", cfg['chkpt'])
dataset = ImageFolder720p(cfg['dataset_path'])
dataloader = DataLoader(dataset, batch_size=1, shuffle=cfg['shuffle'])
logger.info(f"Done setup dataloader: {len(dataloader)}")
mse_loss = nn.MSELoss()
for bi, (img, patches, path) in enumerate(dataloader):
out = torch.zeros(6, 10, 3, 128, 128)
# enc = torch.zeros(6, 10, 16, 8, 8)
avg_loss = 0
for i in range(6):
for j in range(10):
x = Variable(patches[:, :, i, j, :, :]).cuda()
y = model(x)
# e = model.enc_x.data
# p = torch.tensor(np.random.permutation(e.reshape(-1, 1)).reshape(1, 16, 8, 8)).cuda()
# out[i, j] = model.decode(p).data
# enc[i, j] = model.enc_x.data
out[i, j] = y.data
loss = mse_loss(y, x)
avg_loss += (1 / 60) * loss.item()
logger.debug('[%5d/%5d] avg_loss: %f' % (bi, len(dataloader), avg_loss))
# save output
out = np.transpose(out, (0, 3, 1, 4, 2))
out = np.reshape(out, (768, 1280, 3))
out = np.transpose(out, (2, 0, 1))
y = torch.cat((img[0], out), dim=2)
save_imgs(imgs=y.unsqueeze(0), to_size=(3, 768, 2 * 1280), name=f"./test/{cfg['exp_name']}/test_{bi}.png")
def test(solver, model_dir, log_dir, test_dir):
if FLAGS.load_model is not None:
flag, iter_time = solver.load_model(logger=None, model_dir=model_dir, is_train=False)
if flag is True:
print(' [!] Load Success! Iter: {}'.format(iter_time))
else:
exit(' [!] Failed to restore model {}'.format(FLAGS.load_gan_model))
segs, outputs, clses, imgs = solver.generate_test_imgs()
solver.test_top_k(segs, outputs, clses, log_dir)
print('Saving imgs...')
for i in range(segs.shape[0]):
if i % 100 == 0:
print('[{}/{}] saving...'.format(i, segs.shape[0]))
utils.save_imgs(img_stores=[imgs[i:i+1], segs[i:i+1], outputs[i:i+1]], save_dir=test_dir,
img_name=os.path.basename(solver.data.test_paths[i]), is_vertical=False, margin=0)
def save(self): if self.cleaned_data['image']: self.instance.image = save_imgs(self.cleaned_data['image'], self.instance.id, 'avatar') if self.cleaned_data['name']: self.instance.full_name = self.cleaned_data['name'] if not self.cleaned_data['text']: self.instance.text = "" for k in ['title', 'birthday', 'im', 'text']: if self.cleaned_data[k]: setattr(self.instance, k, self.cleaned_data[k]) self.instance.save()
def sample(self, iterTime, saveDir, num_imgs=4):
feed = {
self.model.ratePh: 0.5, # rate: 1 - keep_prob
self.model.trainMode: True
}
img, predCls, segImg = self.sess.run([
self.model.imgTrain, self.model.predClsTrain,
self.model.segImgTrain
],
feed_dict=feed)
# if batch_size is bigger than num_imgs, we just show num_imgs
num_imgs = np.minimum(num_imgs, img.shape[0])
# Save imgs
utils.save_imgs(
img_stores=[img[:num_imgs], predCls[:num_imgs], segImg[:num_imgs]],
iterTime=iterTime,
saveDir=saveDir,
is_vertical=True)
def save(self):
if self.cleaned_data['image']:
self.instance.image = save_imgs(self.cleaned_data['image'],
self.instance.id, 'avatar')
if self.cleaned_data['name']:
self.instance.full_name = self.cleaned_data['name']
if not self.cleaned_data['text']:
self.instance.text = ""
for k in ['title', 'birthday', 'im', 'text']:
if self.cleaned_data[k]:
setattr(self.instance, k, self.cleaned_data[k])
self.instance.save()
def eval(args):
e1 = E1(args.sep, int((args.resize / 64)))
e2 = E2(args.sep, int((args.resize / 64)))
decoder = Decoder(int((args.resize / 64)))
if torch.cuda.is_available():
e1 = e1.cuda()
e2 = e2.cuda()
decoder = decoder.cuda()
if args.load != '':
save_file = os.path.join(args.load, 'checkpoint')
_iter = load_model_for_eval(save_file, e1, e2, decoder)
e1 = e1.eval()
e2 = e2.eval()
decoder = decoder.eval()
if not os.path.exists(args.out) and args.out != "":
os.mkdir(args.out)
save_imgs(args, e1, e2, decoder, _iter)
def save(self): from tagging.models import TaggedItem self.instance.poster = save_imgs(self.cleaned_data['poster'], self.instance.author.id, 'poster') self.instance.title=self.cleaned_data['title'] self.instance.text=self.cleaned_data['text'] if not CENSORSHIP: self.instance.approved = True _save_torrent(self.instance, self.cleaned_data['torrent']) self.instance.save() self.instance.tags = self.cleaned_data['tags'] _update_notification_delivery_queue(self.cleaned_data['tags'], self.instance.author) return self.instance.id
def save(self):
from tagging.models import TaggedItem
self.instance.poster = save_imgs(self.cleaned_data['poster'],
self.instance.author.id, 'poster')
self.instance.title = self.cleaned_data['title']
self.instance.text = self.cleaned_data['text']
if not CENSORSHIP:
self.instance.approved = True
_save_torrent(self.instance, self.cleaned_data['torrent'])
self.instance.save()
self.instance.tags = self.cleaned_data['tags']
_update_notification_delivery_queue(self.cleaned_data['tags'],
self.instance.author)
return self.instance.id
def train(cfg: Namespace) -> None:
print(cfg.device)
assert cfg.device == 'cpu' or (cfg.device == 'cuda'
and T.cuda.is_available())
logger.info('training: experiment %s' % (cfg.exp_name))
# make dir-tree
exp_dir = ROOT_DIR / 'experiments' / cfg.exp_name
for d in ['out', 'checkpoint', 'logs']:
os.makedirs(exp_dir / d, exist_ok=True)
cfg.to_file(exp_dir / 'train_config.txt')
# tb writer
writer = SummaryWriter(exp_dir / 'logs')
model = CAE()
model.train()
if cfg.device == 'cuda':
model.cuda()
logger.info(f'loaded model on {cfg.device}')
dataset = ImageFolder720p(cfg.dataset_path)
dataloader = DataLoader(dataset,
batch_size=cfg.batch_size,
shuffle=cfg.shuffle,
num_workers=cfg.num_workers)
logger.info('loaded dataset')
optimizer = optim.Adam(model.parameters(),
lr=cfg.learning_rate,
weight_decay=1e-5)
loss_criterion = nn.MSELoss()
avg_loss, epoch_avg = 0.0, 0.0
ts = 0
# EPOCHS
for epoch_idx in range(cfg.start_epoch, cfg.num_epochs + 1):
# BATCHES
for batch_idx, data in enumerate(dataloader, start=1):
img, patches, _ = data
if cfg.device == 'cuda':
patches = patches.cuda()
avg_loss_per_image = 0.0
for i in range(6):
for j in range(10):
optimizer.zero_grad()
x = patches[:, :, i, j, :, :]
y = model(x)
loss = loss_criterion(y, x)
avg_loss_per_image += (1 / 60) * loss.item()
loss.backward()
optimizer.step()
avg_loss += avg_loss_per_image
epoch_avg += avg_loss_per_image
if batch_idx % cfg.batch_every == 0:
writer.add_scalar('train/avg_loss', avg_loss / cfg.batch_every,
ts)
for name, param in model.named_parameters():
writer.add_histogram(name, param, ts)
logger.debug('[%3d/%3d][%5d/%5d] avg_loss: %.8f' %
(epoch_idx, cfg.num_epochs, batch_idx,
len(dataloader), avg_loss / cfg.batch_every))
avg_loss = 0.0
ts += 1
# -- end batch every
if batch_idx % cfg.save_every == 0:
out = T.zeros(6, 10, 3, 128, 128)
for i in range(6):
for j in range(10):
x = patches[0, :, i, j, :, :].unsqueeze(0).cuda()
out[i, j] = model(x).cpu().data
out = np.transpose(out, (0, 3, 1, 4, 2))
out = np.reshape(out, (768, 1280, 3))
out = np.transpose(out, (2, 0, 1))
y = T.cat((img[0], out), dim=2).unsqueeze(0)
save_imgs(imgs=y,
to_size=(3, 768, 2 * 1280),
name=exp_dir / f'out/{epoch_idx}_{batch_idx}.png')
# -- end save every
# -- end batches
if epoch_idx % cfg.epoch_every == 0:
epoch_avg /= (len(dataloader) * cfg.epoch_every)
writer.add_scalar('train/epoch_avg_loss',
avg_loss / cfg.batch_every,
epoch_idx // cfg.epoch_every)
logger.info('Epoch avg = %.8f' % epoch_avg)
epoch_avg = 0.0
T.save(model.state_dict(),
exp_dir / f'checkpoint/model_{epoch_idx}.state')
# -- end epoch every
# -- end epoch
# save final model
T.save(model.state_dict(), exp_dir / 'model_final.state')
# cleaning
writer.close()
def evaluate(model,
eval_dataset,
num_workers=0,
print_detail=True,
save_img=True):
"""
Launch evalution.
Args:
model（nn.Layer): A sementic segmentation model.
eval_dataset (paddle.io.Dataset): Used to read and process validation datasets.
num_workers (int, optional): Num workers for data loader. Default: 0.
print_detail (bool, optional): Whether to print detailed information about the evaluation process. Default: True.
Returns:
float: The mIoU of validation datasets.
float: The accuracy of validation datasets.
"""
logger.info('Validating')
evaluator = Eval(eval_dataset.NUM_CLASSES)
evaluator.reset()
model.eval()
nranks = paddle.distributed.ParallelEnv().nranks
local_rank = paddle.distributed.ParallelEnv().local_rank
if nranks > 1:
# Initialize parallel environment if not done.
if not paddle.distributed.parallel.parallel_helper._is_parallel_ctx_initialized(
):
paddle.distributed.init_parallel_env()
batch_sampler = paddle.io.DistributedBatchSampler(
eval_dataset, batch_size=1, shuffle=False, drop_last=True)
loader = paddle.io.DataLoader(
eval_dataset,
batch_sampler=batch_sampler,
num_workers=num_workers,
return_list=True,
)
progbar_val = progbar.Progbar(
target=len(loader), verbose=0 if nranks < 2 else 2)
reader_cost_averager = TimeAverager()
batch_cost_averager = TimeAverager()
batch_start = time.time()
with paddle.no_grad():
for idx, (x, y, _, item) in enumerate(loader):
reader_cost_averager.record(time.time() - batch_start)
# Forward
y = y.astype('int64')
pred = model(x) # 1, c, h, w
if len(pred) > 1:
pred = pred[0]
# Convert to numpy
label = y.squeeze(axis=1).numpy() #
argpred = np.argmax(pred.numpy(), axis=1) # 1, 1, H, W
if save_img:
save_imgs(argpred, item, './output/')
# Add to evaluator
evaluator.add_batch(label, argpred)
batch_cost_averager.record(
time.time() - batch_start, num_samples=len(label))
batch_cost = batch_cost_averager.get_average()
reader_cost = reader_cost_averager.get_average()
if local_rank == 0 and print_detail and idx % 10 == 0:
progbar_val.update(idx + 1, [('batch_cost', batch_cost),
('reader cost', reader_cost)])
reader_cost_averager.reset()
batch_cost_averager.reset()
batch_start = time.time()
PA = evaluator.pixel_accuracy()
MPA = evaluator.mean_pixel_accuracy()
MIoU = evaluator.mean_iou()
FWIoU = evaluator.fwiou()
PC = evaluator.mean_precision()
logger.info(
'PA1:{:.3f}, MPA1:{:.3f}, MIoU1:{:.3f}, FWIoU1:{:.3f}, PC:{:.3f}'.
format(PA, MPA, MIoU, FWIoU, PC))
return PA, MPA, MIoU, FWIoU
def train():
mnist = tf.keras.datasets.mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0
# data, info = tf.load("mnist", with_info=True, data_dir='/data/tensorflow_datasets')
train_data = x_train
if not os.path.exists('./images'):
os.makedirs('./images')
# settting hyperparameter
latent_dim = 100
epochs = 800
batch_size = 32
buffer_size = 6000
save_interval = 50
n_critic = 5
strategy = tf.distribute.MirroredStrategy()
print('Number of devies: {}'.format(strategy.num_replicas_in_sync))
with strategy.scope():
generator = Generator()
discriminator = Critic()
gen_optimizer = tf.keras.optimizers.Adam(0.0001, 0.5, 0.9)
disc_optimizer = tf.keras.optimizers.Adam(0.0001, 0.5, 0.9)
train_dataset = train_data.map(normalize).shuffle(buffer_size).batch(
batch_size)
@tf.function
def train_discriminator(images):
noise = tf.random.normal([batch_size, latent_dim])
with tf.GradientTape() as disc_tape:
generated_imgs = generator(noise, training=True)
generated_output = discriminator(generated_imgs, training=True)
real_output = discriminator(images, training=True)
interpolated_img = random_weighted_average(
[images, generated_imgs])
validity_interpolated = discriminator(interpolated_img,
training=True)
disc_loss = discriminator_loss(real_output, generated_output,
validity_interpolated,
interpolated_img)
grad_disc = disc_tape.gradient(disc_loss,
discriminator.trainable_variables)
disc_optimizer.apply_gradients(
zip(grad_disc, discriminator.trainable_variables))
return disc_loss
@tf.function
def train_generator():
noise = tf.random.normal([batch_size, latent_dim])
with tf.GradientTape() as gen_tape:
generated_imgs = generator(noise, training=True)
generated_output = discriminator(generated_imgs, training=True)
gen_loss = generator_loss(generated_output)
grad_gen = gen_tape.gradient(gen_loss, generator.trainable_variables)
gen_optimizer.apply_gradients(
zip(grad_gen, generator.trainable_variables))
return gen_loss
seed = tf.random.normal([16, latent_dim])
for epoch in range(epochs):
start = time.time()
disc_loss = 0
gen_loss = 0
for images in train_dataset:
disc_loss += train_discriminator(images)
if disc_optimizer.iterations.numpy() % n_critic == 0:
gen_loss += train_generator()
print('Time for epoch {} is {} sec - gen_loss = {}, disc_loss = {}'.
format(epoch + 1,
time.time() - start, gen_loss / batch_size,
disc_loss / (batch_size * n_critic)))
if epoch % save_interval == 0:
save_imgs(epoch, generator, seed)
def save(self):
torm = {
'anime': ['imdb', 'artist', 'album', 'company'],
'books':
['artist', 'album', 'year', 'company', 'tube', 'imdb', 'scrs'],
'games': ['imdb', 'artist', 'album'],
'movies': ['artist', 'album', 'company'],
'music': ['scrs', 'imdb', 'company', 'tube', 'lngs'],
'pics': ['imdb', 'artist', 'album', 'year', 'company', 'tube'],
'tv': ['artist', 'album', 'company'],
'misc': [
'scrs', 'imdb', 'artist', 'album', 'year', 'company', 'tube',
'lngs'
]
}
toadd = {
'anime': ['year', 'tube', 'lngs'],
'books': ['lngs'],
'games': ['year', 'company', 'lngs', 'tube'],
'movies': ['imdb', 'year', 'tube', 'lngs'],
'music': ['album', 'artist', 'year'],
'pics': ['lngs'],
'tv': ['imdb', 'year', 'tube', 'lngs'],
'misc': [],
}
for k in torm.get(self.cleaned_data['section'], []):
del self.instance.attrs[k]
for k in toadd.get(self.cleaned_data['section'], []):
self.instance.attrs[k] = self.cleaned_data[k]
self.instance.section = self.cleaned_data['section']
if self.cleaned_data['section'] == 'misc':
self.instance.subcat = 'misc'
else:
self.instance.subcat = self.cleaned_data[
self.cleaned_data['section']]
if self.cleaned_data['notify'] == 'yes':
if self.instance.author.attrs.has_key('notify'):
l = self.instance.author.attrs['notify']
l.append(self.instance.id)
self.instance.author.attrs['notify'] = l
else:
self.instance.author.attrs['notify'] = [self.instance.id]
if self.cleaned_data['featured'] == 'yes':
if self.instance.author.attrs.has_key('featured'):
l = self.instance.author.attrs['featured']
l.append(self.instance.id)
self.instance.author.attrs['featured'] = l
else:
self.instance.author.attrs['featured'] = [self.instance.id]
if self.cleaned_data['poster']:
#removing old image if new uploaded
drop_pics(self.instance)
if self.cleaned_data['poster']:
self.instance.poster = save_imgs(self.cleaned_data['poster'],
self.instance.author.id, 'poster')
self.instance.title = self.cleaned_data['title']
self.instance.text = self.cleaned_data['text']
_save_torrent(self.instance, self.cleaned_data['torrent'])
if self.cleaned_data['tube']:
self.instance.attrs['tube'] = self.cleaned_data['tube']
_update_notification_delivery_queue(self.cleaned_data['tags'],
self.instance.author)
self.instance.save()
self.instance.tags = self.cleaned_data['tags']
def test(cfg: Namespace) -> None:
logger.info("=== Testing ===")
if cfg.alg == "32":
from autoencoder2bpp import AutoEncoder
elif cfg.alg == "16":
from autoencoder025bpp import AutoEncoder
elif cfg.alg == "8":
from autoencoder006bpp import AutoEncoder
# initial setup
prologue(cfg)
model = CAE()
model.load_state_dict(torch.load(cfg.chkpt))
model.eval()
if cfg.device == "cuda":
model.cuda()
logger.info("Loaded model")
dataset = ImageFolder720p(cfg.dataset_path)
dataloader = DataLoader(dataset, batch_size=1, shuffle=cfg.shuffle)
logger.info("Loaded data")
loss_criterion = nn.MSELoss()
if cfg.weigh_path is not None:
loss_criterion = weighted_loss_function
for batch_idx, data in enumerate(dataloader, start=1):
img, patches, _, weights = data
if cfg.device == 'cuda':
patches = patches.cuda()
out = torch.zeros(3, 256, 256)
avg_loss = 0
x = Variable(patches[:, :, :, :]).cpu()
y = model(x).cpu()
out = y.data
if cfg.weigh_path:
w = Variable(weights)
w = w[:, None, :, :]
w = torch.cat((w, w, w), dim=1)
if cfg.device == "cuda":
w = w.cuda()
loss = loss_criterion(y, x, w)
else:
loss = loss_criterion(y, x)
avg_loss += loss.item()
logger.debug('[%5d/%5d] avg_loss: %f' %
(batch_idx, len(dataloader), avg_loss))
# save output
out = np.reshape(out, (3, 256, 256))
#print(model.encoded)
concat = torch.cat((img[0], out), dim=2).unsqueeze(0)
save_imgs(
imgs=concat,
to_size=(3, 256, 2 * 256),
name=f"../experiments/{cfg.exp_name}/out_test/test_{batch_idx}.png"
)
# final setup
epilogue(cfg)
def save(self):
l = self.topic.attrs.get('scrs', [])
fn = save_imgs(self.cleaned_data['scr'], self.topic.author.id, 'scr')
l.append(fn)
self.topic.attrs['scrs'] = l
self.topic.save()
def train():
if not os.path.exists('./images'):
os.makedirs('./images')
# settting hyperparameter
epochs = 800
batch_size = 4
save_interval = 2
# Model setting
generator = Generator(n_blocks=5)
discriminator = Discriminator()
# Optimizer setting
gen_optimizer = tf.keras.optimizers.Adam(0.0005, 0.9)
disc_optimizer = tf.keras.optimizers.Adam(0.0005, 0.9)
image_ds = tf.data.Dataset.list_files('./data/data512x512/*', shuffle=True)
train_dataset = image_ds.map(lambda x: preprocess_data(x), num_parallel_calls=AUTOTUNE).batch(batch_size).prefetch(AUTOTUNE).cache()
# Loss setting
content_layer = 'block5_conv4' # SRGAN-VGG54
extractor = ContentModel(content_layer)
extractor.trainable = False
cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=False)
mse = tf.keras.losses.MeanSquaredError()
@tf.function
def train_step(lr_images, hr_images):
with tf.GradientTape(persistent=True) as tape:
sr_images = generator(lr_images) # sr -> super resolution
real_output = discriminator(hr_images)
fake_output = discriminator(sr_images)
# adversarial loss
gen_loss = generator_loss(cross_entropy, fake_output) * 1e-3
disc_loss = discriminator_loss(cross_entropy, real_output, fake_output) * 1e-3
# content loss
hr_feat = extractor(hr_images)
sr_feat = extractor(sr_images)
cont_loss = content_loss(mse, hr_feat, sr_feat) * 0.006
perc_loss = cont_loss + gen_loss
grad_gen = tape.gradient(perc_loss, generator.trainable_variables)
grad_disc = tape.gradient(disc_loss, discriminator.trainable_variables)
gen_optimizer.apply_gradients(zip(grad_gen, generator.trainable_variables))
disc_optimizer.apply_gradients(zip(grad_disc, discriminator.trainable_variables))
return gen_loss, disc_loss, cont_loss
total_iter = 0
for epoch in range(1, epochs + 1):
start = time.time()
total_gen_loss = 0
total_disc_loss = 0
total_cont_loss = 0
for i, (lr_images, hr_images) in enumerate(train_dataset, 1):
total_iter += 1
gen_loss, disc_loss, cont_loss = train_step(lr_images, hr_images)
if i % 100 == 0:
print(f'Batch:{i}({total_iter}) -> gen_loss: {gen_loss}, disc_loss: {disc_loss}, cont_loss: {cont_loss}')
save_imgs(epoch, generator, lr_images, hr_images)
total_gen_loss += gen_loss
total_disc_loss += disc_loss
total_cont_loss += cont_loss
print('Time for epoch {} is {} sec -> gen_loss: {}, disc_loss: {}, cont_loss: {}'.format(epoch,
time.time() - start,
total_gen_loss / i,
total_disc_loss / i,
total_cont_loss / i))
def train():
data, info = tfds.load("mnist",
with_info=True,
data_dir='/data/tensorflow_datasets')
train_data = data['train']
if not os.path.exists('./images'):
os.makedirs('./images')
# settting hyperparameter
latent_dim = 100
epochs = 800
batch_size = 32
buffer_size = 6000
save_interval = 50
n_critic = 5
generator = Generator()
discriminator = Critic()
gen_optimizer = tf.keras.optimizers.RMSprop(0.00005)
disc_optimizer = tf.keras.optimizers.RMSprop(0.00005)
train_dataset = train_data.map(normalize).shuffle(buffer_size).batch(
batch_size)
@tf.function
def train_discriminator(images):
noise = tf.random.normal([batch_size, latent_dim])
with tf.GradientTape() as disc_tape:
generated_imgs = generator(noise, training=True)
generated_output = discriminator(generated_imgs, training=True)
real_output = discriminator(images, training=True)
disc_loss = discriminator_loss(real_output, generated_output)
grad_disc = disc_tape.gradient(disc_loss,
discriminator.trainable_variables)
disc_optimizer.apply_gradients(
zip(grad_disc, discriminator.trainable_variables))
for param in discriminator.trainable_variables:
# Except gamma and beta in Batch Normalization
if param.name.split('/')[-1].find(
'gamma') == -1 and param.name.split('/')[-1].find(
'beta') == -1:
param.assign(tf.clip_by_value(param, -0.01, 0.01))
return disc_loss
@tf.function
def train_generator():
noise = tf.random.normal([batch_size, latent_dim])
with tf.GradientTape() as gen_tape:
generated_imgs = generator(noise, training=True)
generated_output = discriminator(generated_imgs, training=True)
gen_loss = generator_loss(generated_output)
grad_gen = gen_tape.gradient(gen_loss, generator.trainable_variables)
gen_optimizer.apply_gradients(
zip(grad_gen, generator.trainable_variables))
return gen_loss
seed = tf.random.normal([16, latent_dim])
for epoch in range(epochs):
start = time.time()
disc_loss = 0
gen_loss = 0
for images in train_dataset:
disc_loss += train_discriminator(images)
if disc_optimizer.iterations.numpy() % n_critic == 0:
gen_loss += train_generator()
print('Time for epoch {} is {} sec - gen_loss = {}, disc_loss = {}'.
format(epoch + 1,
time.time() - start, gen_loss / batch_size,
disc_loss / (batch_size * n_critic)))
if epoch % save_interval == 0:
save_imgs(epoch, generator, seed)
def train(cfg: Namespace) -> None:
logger.info("=== Training ===")
# initial setup
writer = prologue(cfg)
# train-related code
model = CAE()
model.train()
if cfg.device == "cuda":
model.cuda()
logger.debug(f"Model loaded on {cfg.device}")
dataset = ImageFolder720p(cfg.dataset_path)
dataloader = DataLoader(dataset,
batch_size=cfg.batch_size,
shuffle=cfg.shuffle,
num_workers=cfg.num_workers)
logger.debug("Data loaded")
optimizer = optim.Adam(model.parameters(),
lr=cfg.learning_rate,
weight_decay=1e-5)
loss_criterion = nn.MSELoss()
# scheduler = ...
avg_loss, epoch_avg = 0.0, 0.0
ts = 0
# train-loop
for epoch_idx in range(cfg.start_epoch, cfg.num_epochs + 1):
# scheduler.step()
for batch_idx, data in enumerate(dataloader, start=1):
img, y, patches, _ = data
if cfg.device == "cuda":
patches = patches.cuda()
avg_loss_per_image = 0.0 # 初始化单张图片的损失
for i in range(1):
for j in range(1):
optimizer.zero_grad()
x = Variable(patches[:, :, i, j, :, :])
y = model(x)
loss = loss_criterion(y, x)
avg_loss_per_image += (1 / 60) * loss.item()
avg_loss_per_image += (1 / 1) * loss.item()
loss.backward()
optimizer.step()
avg_loss += avg_loss_per_image
epoch_avg += avg_loss_per_image
if batch_idx % cfg.batch_every == 0:
writer.add_scalar("train/avg_loss", avg_loss / cfg.batch_every,
ts)
for name, param in model.named_parameters():
writer.add_histogram(name, param, ts)
logger.debug('[%3d/%3d][%5d/%5d] avg_loss: %.8f' %
(epoch_idx, cfg.num_epochs, batch_idx,
len(dataloader), avg_loss / cfg.batch_every))
avg_loss = 0.0
ts += 1
if batch_idx % cfg.save_every == 0:
# out = torch.zeros(6, 10, 3, 128, 128)
out = torch.zeros(1, 1, 3, 256, 256)
for i in range(1):
for j in range(1):
x = Variable(patches[0, :, i,
j, :, :].unsqueeze(0)).cuda()
out[i, j] = model(x).cpu().data
out = np.transpose(out, (0, 3, 1, 4, 2))
# out = np.reshape(out, (768, 1280, 3))
out = np.reshape(out, (256, 256, 3))
out = np.transpose(out, (2, 0, 1))
y = torch.cat((img[0], out), dim=2).unsqueeze(0)
# save_imgs(imgs=y, to_size=(3, 768, 2 * 1280), name=f"/data2/TDL/paper_fabric/workdir/4_30_short_eassy/{cfg.exp_name}/out/out_{epoch_idx}_{batch_idx}.png")
save_imgs(
imgs=y,
to_size=(3, 256, 2 * 256),
name=
f"/data2/TDL/paper_fabric/workdir/4_30_short_eassy/{cfg.exp_name}/out/out_{epoch_idx}_{batch_idx}.png"
)
# -- batch-loop
if epoch_idx % cfg.epoch_every == 0:
epoch_avg /= (len(dataloader) * cfg.epoch_every)
writer.add_scalar("train/epoch_avg_loss",
avg_loss / cfg.batch_every,
epoch_idx // cfg.epoch_every)
logger.info("Epoch avg = %.8f" % epoch_avg)
epoch_avg = 0.0
torch.save(
model.state_dict(),
f"/data2/TDL/paper_fabric/workdir/4_30_short_eassy/{cfg.exp_name}/chkpt/model_{epoch_idx}.pth"
)
# -- train-loop
# save final model
torch.save(
model.state_dict(),
f"/data2/TDL/paper_fabric/workdir/4_30_short_eassy/{cfg.exp_name}/model_final.pth"
)
# final setup
epilogue(cfg, writer)
def train(cfg: Namespace) -> None:
assert cfg.device == "cpu" or (cfg.device == "cuda"
and T.cuda.is_available())
root_dir = Path(__file__).resolve().parents[1]
logger.info("training: experiment %s" % (cfg.exp_name))
# make dir-tree
exp_dir = root_dir / "experiments" / cfg.exp_name
for d in ["out", "checkpoint", "logs"]:
os.makedirs(exp_dir / d, exist_ok=True)
cfg.to_file(exp_dir / "train_config.json")
# tb tb_writer
tb_writer = SummaryWriter(exp_dir / "logs")
logger.info("started tensorboard writer")
model = CAE()
model.train()
if cfg.device == "cuda":
model.cuda()
logger.info(f"loaded model on {cfg.device}")
dataloader = DataLoader(
dataset=ImageFolder720p(cfg.dataset_path),
batch_size=cfg.batch_size,
shuffle=cfg.shuffle,
num_workers=cfg.num_workers,
)
logger.info(f"loaded dataset from {cfg.dataset_path}")
optimizer = optim.Adam(model.parameters(),
lr=cfg.learning_rate,
weight_decay=1e-5)
loss_criterion = nn.MSELoss()
avg_loss, epoch_avg = 0.0, 0.0
ts = 0
# EPOCHS
for epoch_idx in range(cfg.start_epoch, cfg.num_epochs + 1):
# BATCHES
for batch_idx, data in enumerate(dataloader, start=1):
img, patches, _ = data
if cfg.device == "cuda":
patches = patches.cuda()
avg_loss_per_image = 0.0
for i in range(6):
for j in range(10):
optimizer.zero_grad()
x = patches[:, :, i, j, :, :]
y = model(x)
loss = loss_criterion(y, x)
avg_loss_per_image += (1 / 60) * loss.item()
loss.backward()
optimizer.step()
avg_loss += avg_loss_per_image
epoch_avg += avg_loss_per_image
if batch_idx % cfg.batch_every == 0:
tb_writer.add_scalar("train/avg_loss",
avg_loss / cfg.batch_every, ts)
for name, param in model.named_parameters():
tb_writer.add_histogram(name, param, ts)
logger.debug("[%3d/%3d][%5d/%5d] avg_loss: %.8f" % (
epoch_idx,
cfg.num_epochs,
batch_idx,
len(dataloader),
avg_loss / cfg.batch_every,
))
avg_loss = 0.0
ts += 1
# -- end batch every
if batch_idx % cfg.save_every == 0:
out = T.zeros(6, 10, 3, 128, 128)
for i in range(6):
for j in range(10):
x = patches[0, :, i, j, :, :].unsqueeze(0).cuda()
out[i, j] = model(x).cpu().data
out = np.transpose(out, (0, 3, 1, 4, 2))
out = np.reshape(out, (768, 1280, 3))
out = np.transpose(out, (2, 0, 1))
y = T.cat((img[0], out), dim=2).unsqueeze(0)
save_imgs(
imgs=y,
to_size=(3, 768, 2 * 1280),
name=exp_dir / f"out/{epoch_idx}_{batch_idx}.png",
)
# -- end save every
# -- end batches
if epoch_idx % cfg.epoch_every == 0:
epoch_avg /= len(dataloader) * cfg.epoch_every
tb_writer.add_scalar(
"train/epoch_avg_loss",
avg_loss / cfg.batch_every,
epoch_idx // cfg.epoch_every,
)
logger.info("Epoch avg = %.8f" % epoch_avg)
epoch_avg = 0.0
T.save(model.state_dict(),
exp_dir / f"checkpoint/model_{epoch_idx}.pth")
# -- end epoch every
# -- end epoch
# save final model
T.save(model.state_dict(), exp_dir / "model_final.pth")
# cleaning
tb_writer.close()
def main():
args = parse_args()
images = args.input
vol = load_volume(images, ext='.bmp')
with h5py.File('downsampled_histogram.h5', 'r') as f:
hist = f['hist'][:]
hist = np.cumsum(hist)
vol = np.floor(255 * hist[vol])
vol[vol > 255] = 255
vol[vol < 0] = 0
vol = vol.astype(np.uint8)
sel = ndi.generate_binary_structure(3, 1)
sel[0] = 0
sel[2] = 0
subject = os.path.basename(args.input)
print(subject)
params = {}
if args.recover_parameters:
# try:
with open(args.parameter_file, 'r') as f:
params = json.load(f)
t_iris = params[subject]['t_iris']
t_pupil = params[subject]['t_pupil']
window_iris = tuple(params[subject]['window_iris'])
window_pupil = tuple(params[subject]['window_pupil'])
# except KeyError:
# pass
if args.t_iris is not None:
t_iris = args.t_iris
if args.t_pupil is not None:
t_pupil = args.t_pupil
if args.window_iris is not None:
window_iris = tuple(args.window_iris)
if args.window_pupil is not None:
window_pupil = tuple(args.window_pupil)
depth = args.depth
radius = int(min(vol.shape[1], vol.shape[2]) / 4)
if args.save_iris:
iris_seg = np.zeros(vol.shape, dtype=np.uint8)
if args.save_pupil:
pupil_seg = np.zeros(vol.shape, dtype=np.uint8)
seg = np.zeros(vol.shape, dtype=np.uint8)
sums, counts = None, None
for i in range(0, vol.shape[0], depth):
subvol = np.copy(vol[i:i + depth])
orig_shape = subvol.shape
if subvol.shape[0] < depth:
subvol = np.concatenate([
subvol,
np.repeat(subvol[-1].reshape(1, subvol.shape[1],
subvol.shape[2]),
depth - subvol.shape[0],
axis=0)
],
axis=0)
if all([
window_iris[j] == window_pupil[j]
for j in range(len(window_iris))
]):
sums, counts = summate(integral_image(subvol),
np.asarray(window_iris))
# Iris Segmentation
iris = 1.0 - threshold_bradley_nd(
subvol, t=t_iris, s=window_iris, sums=sums, counts=counts)
iris = ndi.binary_fill_holes(iris, structure=sel)
# Pupil Segmentation
pupil = 1.0 - threshold_bradley_nd(
subvol, t=t_pupil, s=window_pupil, sums=sums, counts=counts)
pupil = ndi.binary_fill_holes(pupil, structure=sel)
pupil = ndi.binary_erosion(pupil, structure=sel)
pupil = ndi.binary_dilation(pupil, structure=sel)
pupil = ndi.binary_dilation(pupil, structure=sel).astype(np.uint8)
pupil_collapsed = (np.sum(pupil, axis=0) > 1).astype(np.uint8)
pupil_collapsed = remove_small_objects(label(pupil_collapsed),
min_size=200).astype(np.uint8)
circle_mask = np.zeros(pupil_collapsed.shape, dtype=np.uint8)
try:
objs = regionprops(label(pupil_collapsed),
intensity_image=np.mean(subvol, axis=0).astype(
np.uint8))
for obj in objs:
if obj.convex_area > 1000 and obj.solidity < 0.5 or np.sum(
obj.inertia_tensor_eigvals) == 0:
pupil_collapsed[obj.coords[:, 0], obj.coords[:, 1]] = 0
pupil_idx = np.argmax([
o.area * np.abs(o.orientation) * o.solidity /
(o.eccentricity + 1e-7) /
(o.inertia_tensor_eigvals[0] - o.inertia_tensor_eigvals[1])
for o in objs
])
pupil_obj = objs[pupil_idx]
circle_coords = circle(pupil_obj.centroid[0],
pupil_obj.centroid[1],
radius,
shape=pupil_collapsed.shape)
circle_mask[circle_coords] = 1
except (ValueError, IndexError):
pass
pupil = np.logical_and(
pupil,
np.repeat(pupil_collapsed.reshape((1, ) + pupil_collapsed.shape),
pupil.shape[0],
axis=0))
# Final Segmentation
final = np.logical_xor(iris, pupil).astype(np.uint8)
final = ndi.binary_dilation(final, structure=sel)
final[:, circle_mask == 0] = 0
# Save it
seg[i:i + depth] = final[:orig_shape[0]]
if args.save_iris:
iris_seg[i:i + depth] += iris[:orig_shape[0]]
if args.save_pupil:
pupil_seg[i:i + depth] += pupil[:orig_shape[0]]
seg[:, np.sum(seg, axis=0) < 20] = 0
seg = ndi.binary_erosion(seg, structure=sel)
seg = ndi.binary_erosion(seg, structure=sel).astype(np.uint8)
outdir = os.path.join(args.output, subject)
if not os.path.isdir(outdir):
os.makedirs(outdir)
seg[seg.nonzero()] = 255
save_imgs(seg, outdir, prefix=subject)
if args.save_iris:
outdir = os.path.join(args.output, 'iris', subject)
if not os.path.isdir(outdir):
os.makedirs(outdir)
iris_seg[iris_seg.nonzero()] = 255
save_imgs(iris_seg, outdir)
if args.save_pupil:
outdir = os.path.join(args.output, 'pupil', subject)
if not os.path.isdir(outdir):
os.makedirs(outdir)
pupil_seg[pupil_seg.nonzero()] = 255
save_imgs(pupil_seg, outdir)
shutil.copy(args.parameter_file, args.parameter_file + '.bak')
with open(args.parameter_file, 'w') as f:
params[subject] = {
't_iris': t_iris,
't_pupil': t_pupil,
'window_iris': window_iris,
'window_pupil': window_pupil,
}
json.dump(params, f)
def save(self):
l = self.topic.attrs.get('scrs', [])
fn = save_imgs(self.cleaned_data['scr'], self.topic.author.id, 'scr')
l.append(fn)
self.topic.attrs['scrs'] = l
self.topic.save()
def eval(self,
tb_writer=None,
iter_time=None,
save_dir=None,
is_debug=False):
if self.multi_test:
run_ops = [
self.model.mIoU_metric_update,
self.model.accuracy_metric_update,
self.model.precision_metric_update,
self.model.recall_metric_update,
self.model.per_class_accuracy_metric_update, self.model.imgVal,
self.model.predClsVal, self.model.segImgVal,
self.model.img_name_val, self.model.user_id_val,
self.model.imgVal_s1, self.model.imgVal_s2,
self.model.predVal_s1, self.model.predVal_s2,
self.model.segImgVal_s1, self.model.segImgVal_s2
]
else:
run_ops = [
self.model.mIoU_metric_update,
self.model.accuracy_metric_update,
self.model.precision_metric_update,
self.model.recall_metric_update,
self.model.per_class_accuracy_metric_update, self.model.imgVal,
self.model.predClsVal, self.model.segImgVal,
self.model.img_name_val, self.model.user_id_val
]
feed = {
self.model.ratePh: 0., # rate: 1 - keep_prob
self.model.trainMode: False
}
# Initialize/reset the running variables
self.sess.run(self.model.running_vars_initializer)
per_cla_acc_mat = None
for iterTime in range(self.data.numValImgs):
img_s1, img_s2, pred_s1, pred_s2, segImg_s1, segImg_s2 = None, None, None, None, None, None
if self.multi_test:
_, _, _, _, per_cla_acc_mat, img, predCls, segImg, img_name, user_id, \
img_s1, img_s2, pred_s1, pred_s2, segImg_s1, segImg_s2 = self.sess.run(run_ops, feed_dict=feed)
else:
_, _, _, _, per_cla_acc_mat, img, predCls, segImg, img_name, user_id = \
self.sess.run(run_ops, feed_dict=feed)
if iterTime % 100 == 0:
msg = "\r - Evaluating progress: {:.2f}%".format(
(iterTime / self.data.numValImgs) * 100.)
# Print it.
sys.stdout.write(msg)
sys.stdout.flush()
############################################################################################################
if not self.is_train:
# Save images
utils.save_imgs(img_stores=[img, predCls, segImg],
saveDir=save_dir,
img_name=img_name.astype('U26'),
is_vertical=False)
if not self.is_train and is_debug and self.multi_test:
# # Step 1: save rotated images
predCls_s1 = np.argmax(
pred_s1, axis=-1) # predict class using argmax function
utils.save_imgs(img_stores=[img_s1, predCls_s1, segImg_s1],
saveDir=os.path.join(save_dir, 'debug'),
name_append='step1_',
img_name=img_name.astype('U26'),
is_vertical=True)
# Step 2: save inverse-roated images
predCls_s2 = np.argmax(
pred_s2, axis=-1) # predict class using argmax function
utils.save_imgs(img_stores=[img_s2, predCls_s2, segImg_s2],
saveDir=os.path.join(save_dir, 'debug'),
name_append='step2_',
img_name=img_name.astype('U26'),
is_vertical=True)
# Step 3: Save comparison image that includes img, single_pred, multi-test_pred, gt
utils.save_imgs(img_stores=[
img,
np.expand_dims(predCls_s1[5], axis=0), predCls, segImg
],
saveDir=os.path.join(save_dir, 'debug'),
name_append='step3_',
img_name=img_name.astype('U26'),
is_vertical=False)
############################################################################################################
# Calculate the mIoU
mIoU, accuracy, precision, recall, f1_score, metric_summary_op = self.sess.run(
[
self.model.mIoU_metric, self.model.accuracy_metric,
self.model.precision_metric, self.model.recall_metric,
self.model.f1_score_metric, self.model.metric_summary_op
])
if self.is_train:
# Write to tensorboard
tb_writer.add_summary(metric_summary_op, iter_time)
tb_writer.flush()
mIoU *= 100.
accuracy *= 100.
precision *= 100.
recall *= 100.
f1_score *= 100.
per_cla_acc_mat *= 100.
return mIoU, accuracy, per_cla_acc_mat, precision, recall, f1_score
def train(args):
if not os.path.exists(args.out):
os.makedirs(args.out)
_iter = 0
comp_transformA = transforms.Compose([
transforms.CenterCrop(args.cropA),
transforms.Resize(args.resize),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
comp_transformB = transforms.Compose([
transforms.CenterCrop(args.cropB),
transforms.Resize(args.resize),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
domA_train = CustomDataset(os.path.join(args.root, 'trainA.txt'), transform=comp_transformA)
domB_train = CustomDataset(os.path.join(args.root, 'trainB.txt'), transform=comp_transformB)
A_label = torch.full((args.bs,), 1)
B_label = torch.full((args.bs,), 0)
B_separate = torch.full((args.bs, args.sep * (args.resize // 64) * (args.resize // 64)), 0)
e1 = E1(args.sep, args.resize // 64)
e2 = E2(args.sep, args.resize // 64)
decoder = Decoder(args.resize // 64)
disc = Disc(args.sep, args.resize // 64)
mse = nn.MSELoss()
bce = nn.BCELoss()
if torch.cuda.is_available():
e1 = e1.cuda()
e2 = e2.cuda()
decoder = decoder.cuda()
disc = disc.cuda()
A_label = A_label.cuda()
B_label = B_label.cuda()
B_separate = B_separate.cuda()
mse = mse.cuda()
bce = bce.cuda()
ae_params = list(e1.parameters()) + list(e2.parameters()) + list(decoder.parameters())
ae_optimizer = optim.Adam(ae_params, lr=args.lr, betas=(0.5, 0.999))
disc_params = disc.parameters()
disc_optimizer = optim.Adam(disc_params, lr=args.disclr, betas=(0.5, 0.999))
if args.load != '':
save_file = os.path.join(args.load, 'checkpoint')
_iter = load_model(save_file, e1, e2, decoder, ae_optimizer, disc, disc_optimizer)
e1 = e1.train()
e2 = e2.train()
decoder = decoder.train()
disc = disc.train()
print('Started training...')
while True:
domA_loader = torch.utils.data.DataLoader(domA_train, batch_size=args.bs,
shuffle=True, num_workers=6)
domB_loader = torch.utils.data.DataLoader(domB_train, batch_size=args.bs,
shuffle=True, num_workers=6)
if _iter >= args.iters:
break
for domA_img, domB_img in zip(domA_loader, domB_loader):
if domA_img.size(0) != args.bs or domB_img.size(0) != args.bs:
break
domA_img = Variable(domA_img)
domB_img = Variable(domB_img)
if torch.cuda.is_available():
domA_img = domA_img.cuda()
domB_img = domB_img.cuda()
domA_img = domA_img.view((-1, 3, args.resize, args.resize))
domB_img = domB_img.view((-1, 3, args.resize, args.resize))
ae_optimizer.zero_grad()
A_common = e1(domA_img)
A_separate = e2(domA_img)
A_encoding = torch.cat([A_common, A_separate], dim=1)
B_common = e1(domB_img)
B_encoding = torch.cat([B_common, B_separate], dim=1)
A_decoding = decoder(A_encoding)
B_decoding = decoder(B_encoding)
loss = mse(A_decoding, domA_img) + mse(B_decoding, domB_img)
if args.discweight > 0:
preds_A = disc(A_common)
preds_B = disc(B_common)
loss += args.discweight * (bce(preds_A, B_label) + bce(preds_B, B_label))
loss.backward()
torch.nn.utils.clip_grad_norm_(ae_params, 5)
ae_optimizer.step()
if args.discweight > 0:
disc_optimizer.zero_grad()
A_common = e1(domA_img)
B_common = e1(domB_img)
disc_A = disc(A_common)
disc_B = disc(B_common)
loss2 = bce(disc_A, A_label) + bce(disc_B, B_label)
loss2.backward()
torch.nn.utils.clip_grad_norm_(disc_params, 5)
disc_optimizer.step()
if _iter % args.progress_iter == 0:
print('Outfile: %s | Iteration %d | loss %.6f | loss1: %.6f | loss2: %.6f' % (args.out, _iter, loss+loss2, loss, loss2))
if _iter % args.display_iter == 0:
e1 = e1.eval()
e2 = e2.eval()
decoder = decoder.eval()
save_imgs(args, e1, e2, decoder, _iter)
e1 = e1.train()
e2 = e2.train()
decoder = decoder.train()
if _iter % args.save_iter == 0:
save_file = os.path.join(args.out, 'checkpoint_%d' % _iter)
save_model(save_file, e1, e2, decoder, ae_optimizer, disc, disc_optimizer, _iter)
_iter += 1
def train(cfg):
os.makedirs(f"out/{cfg['exp_name']}", exist_ok=True)
os.makedirs(f"checkpoints/{cfg['exp_name']}", exist_ok=True)
# dump config for current experiment
with open(f"checkpoints/{cfg['exp_name']}/setup.cfg", "wt") as f:
for k, v in cfg.items():
f.write("%15s: %s\n" % (k, v))
model = CAE().cuda()
if cfg['load']:
model.load_state_dict(torch.load(cfg['chkpt']))
logger.info("Loaded model from", cfg['chkpt'])
model.train()
logger.info("Done setup model")
dataset = ImageFolder720p(cfg['dataset_path'])
dataloader = DataLoader(dataset,
batch_size=cfg['batch_size'],
shuffle=cfg['shuffle'],
num_workers=cfg['num_workers'])
logger.info(
f"Done setup dataloader: {len(dataloader)} batches of size {cfg['batch_size']}"
)
mse_loss = nn.MSELoss()
adam = torch.optim.Adam(model.parameters(),
lr=cfg['learning_rate'],
weight_decay=1e-5)
sgd = torch.optim.SGD(model.parameters(), lr=cfg['learning_rate'])
optimizer = adam
ra = 0
for ei in range(cfg['resume_epoch'], cfg['num_epochs']):
for bi, (img, patches, _) in enumerate(dataloader):
avg_loss = 0
for i in range(6):
for j in range(10):
x = Variable(patches[:, :, i, j, :, :]).cuda()
y = model(x)
loss = mse_loss(y, x)
avg_loss += (1 / 60) * loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
ra = avg_loss if bi == 0 else ra * bi / (bi + 1) + avg_loss / (bi +
1)
logger.debug('[%3d/%3d][%5d/%5d] avg_loss: %f, ra: %f' %
(ei + 1, cfg['num_epochs'], bi + 1, len(dataloader),
avg_loss, ra))
# save img
if bi % cfg['out_every'] == 0:
out = torch.zeros(6, 10, 3, 128, 128)
for i in range(6):
for j in range(10):
x = Variable(patches[0, :, i,
j, :, :].unsqueeze(0)).cuda()
out[i, j] = model(x).cpu().data
out = np.transpose(out, (0, 3, 1, 4, 2))
out = np.reshape(out, (768, 1280, 3))
out = np.transpose(out, (2, 0, 1))
y = torch.cat((img[0], out), dim=2).unsqueeze(0)
save_imgs(imgs=y,
to_size=(3, 768, 2 * 1280),
name=f"out/{cfg['exp_name']}/out_{ei}_{bi}.png")
# save model
if bi % cfg['save_every'] == cfg['save_every'] - 1:
torch.save(
model.state_dict(),
f"checkpoints/{cfg['exp_name']}/model_{ei}_{bi}.state")
# save final model
torch.save(model.state_dict(),
f"checkpoints/{cfg['exp_name']}/model_final.state")
def train():
(train_data, _), (_, _) = tf.keras.datasets.mnist.load_data()
if not os.path.exists('./images'):
os.makedirs('./images')
# settting hyperparameter
latent_dim = 100
epochs = 800
batch_size = 32
buffer_size = 6000
save_interval = 50
n_critic = 5
generator = Generator()
discriminator = Critic()
gen_optimizer = tf.keras.optimizers.Adam(0.0001, 0.5, 0.9)
disc_optimizer = tf.keras.optimizers.Adam(0.0001, 0.5, 0.9)
# Rescale -1 to 1
train_data = train_data / 127.5 - 1.
train_data = np.expand_dims(train_data, axis=3).astype('float32')
train_dataset = tf.data.Dataset.from_tensor_slices(train_data).shuffle(
buffer_size).batch(batch_size)
@tf.function
def train_discriminator(images):
noise = tf.random.normal([batch_size, latent_dim])
with tf.GradientTape() as disc_tape:
generated_imgs = generator(noise, training=True)
generated_output = discriminator(generated_imgs, training=True)
real_output = discriminator(images, training=True)
interpolated_img = random_weighted_average(
[images, generated_imgs])
validity_interpolated = discriminator(interpolated_img,
training=True)
disc_loss = discriminator_loss(real_output, generated_output,
validity_interpolated,
interpolated_img)
grad_disc = disc_tape.gradient(disc_loss,
discriminator.trainable_variables)
disc_optimizer.apply_gradients(
zip(grad_disc, discriminator.trainable_variables))
return disc_loss
@tf.function
def train_generator():
noise = tf.random.normal([batch_size, latent_dim])
with tf.GradientTape() as gen_tape:
generated_imgs = generator(noise, training=True)
generated_output = discriminator(generated_imgs, training=True)
gen_loss = generator_loss(generated_output)
grad_gen = gen_tape.gradient(gen_loss, generator.trainable_variables)
gen_optimizer.apply_gradients(
zip(grad_gen, generator.trainable_variables))
return gen_loss
seed = tf.random.normal([16, latent_dim])
for epoch in range(epochs):
start = time.time()
disc_loss = 0
gen_loss = 0
for images in train_dataset:
disc_loss += train_discriminator(images)
if disc_optimizer.iterations.numpy() % n_critic == 0:
gen_loss += train_generator()
print('Time for epoch {} is {} sec - gen_loss = {}, disc_loss = {}'.
format(epoch + 1,
time.time() - start, gen_loss / batch_size,
disc_loss / (batch_size * n_critic)))
if epoch % save_interval == 0:
save_imgs(epoch, generator, seed)
def train():
(train_data, _), (_, _) = tf.keras.datasets.mnist.load_data()
print(type(train_data))
if not os.path.exists('./images'):
os.makedirs('./images')
# settting hyperparameter
latent_dim = 100
epochs = 800
batch_size = 200
buffer_size = 6000
save_interval = 50
generator = Generator()
discriminator = Discriminator()
gen_optimizer = tf.keras.optimizers.Adam(0.0002, 0.5)
disc_optimizer = tf.keras.optimizers.Adam(0.0002, 0.5)
# Rescale -1 to 1
train_data = train_data / 127.5 - 1.
train_data = np.expand_dims(train_data, axis=3).astype('float32')
train_dataset = tf.data.Dataset.from_tensor_slices(train_data).shuffle(
buffer_size).batch(batch_size)
cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)
@tf.function
def train_step(images):
noise = tf.random.normal([batch_size, latent_dim])
with tf.GradientTape(persistent=True) as tape:
generated_images = generator(noise)
real_output = discriminator(images)
generated_output = discriminator(generated_images)
gen_loss = generator_loss(cross_entropy, generated_output)
disc_loss = discriminator_loss(cross_entropy, real_output,
generated_output)
grad_gen = tape.gradient(gen_loss, generator.trainable_variables)
grad_disc = tape.gradient(disc_loss, discriminator.trainable_variables)
gen_optimizer.apply_gradients(
zip(grad_gen, generator.trainable_variables))
disc_optimizer.apply_gradients(
zip(grad_disc, discriminator.trainable_variables))
return gen_loss, disc_loss
seed = tf.random.normal([16, latent_dim])
for epoch in range(epochs):
start = time.time()
total_gen_loss = 0
total_disc_loss = 0
for images in train_dataset:
gen_loss, disc_loss = train_step(images)
total_gen_loss += gen_loss
total_disc_loss += disc_loss
print('Time for epoch {} is {} sec - gen_loss = {}, disc_loss = {}'.
format(epoch + 1,
time.time() - start, total_gen_loss / batch_size,
total_disc_loss / batch_size))
if epoch % save_interval == 0:
save_imgs(epoch, generator, seed)