def show_pairs(images, features, pairs):
dists = np.sqrt(np.sum((features[0] - features[1])**2, axis=1))
ds_utils.denormalize(images[0])
ds_utils.denormalize(images[1])
images[1] = vis.add_error_to_images(images[1],
dists,
size=2.0,
thickness=2,
vmin=0,
vmax=1)
images[1] = vis.add_id_to_images(images[1],
pairs.numpy(),
size=1.2,
thickness=2,
color=(1, 0, 1))
thresh = 0.4
corrects = (dists < thresh) == pairs.cpu().numpy()
colors = [(0, 1, 0) if c else (1, 0, 0) for c in corrects]
images[1] = vis.add_cirle_to_images(images[1], colors)
images[0] = vis._to_disp_images(images[0])
img_rows = [
vis.make_grid(imgs,
fx=0.75,
fy=0.75,
nCols=len(dists),
normalize=False) for imgs in images
]
vis.vis_square(img_rows, nCols=1, normalize=False)
if __name__ == "__main__":
# extract_main()
# exit()
from utils.nn import Batch
from utils import random
random.init_random()
ds = VggFace2(
train=True,
deterministic=True,
use_cache=False,
align_face_orientation=False,
return_modified_images=False,
image_size=256,
)
micro_batch_loader = td.DataLoader(ds,
batch_size=10,
shuffle=True,
num_workers=0)
f = 1.0
t = time.perf_counter()
for iter, data in enumerate(micro_batch_loader):
print("t load:", time.perf_counter() - t)
t = time.perf_counter()
batch = Batch(data, gpu=False)
print("t Batch:", time.perf_counter() - t)
images = nn.denormalized(batch.images)
vis.vis_square(images, fx=f, fy=f, normalize=False, nCols=10, wait=0)
result['landmarks'], self.landmark_sigma, self.landmark_ids)
return result
if __name__ == '__main__':
from utils.nn import Batch, to_numpy
import utils.common
from landmarks import lmconfig as lmcfg
utils.common.init_random(3)
lmcfg.config_landmarks('wflw')
ds = WFLW(train=True, deterministic=True, use_cache=True, daug=0)
# ds.filter_labels({'pose':0, 'blur':0, 'occlusion':1})
dl = td.DataLoader(ds, batch_size=1, shuffle=False, num_workers=0)
cfg.WITH_LANDMARK_LOSS = False
for data in dl:
batch = Batch(data, gpu=False)
images = vis._to_disp_images(batch.images, denorm=True)
# lms = lmutils.convert_landmarks(to_numpy(batch.landmarks), lmutils.LM98_TO_LM68)
lms = batch.landmarks
images = vis.add_landmarks_to_images(images,
lms,
draw_wireframe=False,
color=(0, 255, 0),
radius=3)
vis.vis_square(images, nCols=1, fx=1., fy=1., normalize=False)
return np.concatenate(sizes)
if __name__ == '__main__':
import torch
from utils import vis
from utils.nn import Batch
from datasets import ds_utils
from datasets import affectnet
torch.manual_seed(0)
torch.cuda.manual_seed_all(0)
train = True
datasets = [
affectnet.AffectNet(train=train, max_samples=1000),
# vggface2.VggFace2(train=train, max_samples=1000),
]
multi_ds = MultiFaceDataset(datasets, train=True, max_samples=5000)
print(multi_ds)
dl = td.DataLoader(multi_ds, batch_size=40, shuffle=False, num_workers=0)
for data in dl:
batch = Batch(data, gpu=False)
inputs = batch.images.clone()
ds_utils.denormalize(inputs)
imgs = vis.add_landmarks_to_images(inputs.numpy(),
batch.landmarks.numpy())
# imgs = vis.add_pose_to_images(inputs.numpy(), batch.poses.numpy())
# imgs = vis.add_emotion_to_images(imgs, batch.emotions.numpy())
vis.vis_square(imgs, nCols=20, fx=0.6, fy=0.6, normalize=False)
parser.add_argument('--nd', default=1, type=int)
args = parser.parse_args()
extract_features(st=args.st, nd=args.nd)
if __name__ == '__main__':
# extract_main()
# exit()
import torch
from utils import vis, face_processing
from utils.nn import Batch
torch.manual_seed(0)
torch.cuda.manual_seed_all(0)
ds = LFWImages(train=True, deterministic=True, max_samples=40, use_cache=False)
dl = torch.utils.data.DataLoader(ds, batch_size=10, shuffle=True, num_workers=1)
data = next(iter(dl))
inputs = Batch(data).images
ds_utils.denormalize(inputs)
# imgs = vis.add_id_to_images(inputs.numpy(), data[1].numpy())
# imgs = vis.add_pose_to_images(inputs.numpy(), pose.numpy())
# imgs = vis.add_landmarks_to_images(inputs.numpy(), landmarks.numpy())
imgs = inputs.detach().cpu().numpy()
vis.vis_square(imgs, fx=1.0, fy=1.0, normalize=False)
args = parser.parse_args()
extract_features(st=args.st, nd=args.nd)
if __name__ == '__main__':
# extract_main()
# create_annotations(split='dev', num_ids=500)
# extract_crops()
# read_openface_csvs()
# exit()
from utils import vis, face_processing
ds = VoxCeleb(train=True, max_samples=50000, use_cache=True)
print(ds)
dl = td.DataLoader(ds, batch_size=40, shuffle=False, num_workers=0)
from utils.nn import Batch
for data in dl:
batch = Batch(data)
print(batch.ids)
ds_utils.denormalize(batch.images)
# vis.show_images_in_batch(batch.images.detach().cpu())
vis.vis_square(batch.images.detach().cpu(),
fx=0.7,
fy=0.7,
normalize=False)
# print(item)
lm_x, lm_y = lm[0], lm[1]
cv2.circle(img, (int(lm_x), int(lm_y)), 3, (0, 0, 255), -1)
cv2.imshow('landmarks', cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
cv2.waitKey(0)
if __name__ == '__main__':
from utils import vis
import torch
torch.manual_seed(0)
torch.cuda.manual_seed_all(0)
ds = W300(train=True, deterministic=True, use_cache=True,
test_split='challenging', daug=0, align_face_orientation=False,
return_modified_images=False)
dl = td.DataLoader(ds, batch_size=50, shuffle=False, num_workers=0)
print(ds)
cfg.WITH_LANDMARK_LOSS = False
for data in dl:
batch = Batch(data, gpu=False)
inputs = batch.images.clone()
imgs = vis._to_disp_images(inputs, denorm=True)
imgs = vis.add_landmarks_to_images(imgs, batch.landmarks, radius=3, color=(0,255,0))
# imgs = vis.add_landmarks_to_images(imgs, data['landmarks_of'].numpy(), color=(1,0,0))
vis.vis_square(imgs, nCols=5, fx=1, fy=1, normalize=False)
def calc_triplet_loss(outputs,
c,
return_acc=False,
images=None,
feature_name=None,
wnd_title=None):
margin = 0.2
eps = 1e-8
debug = False
is_expressions = (not isinstance(
c, list)) and len(c.shape) > 1 and c.shape[1] == 3
pos_id, neg_id = make_triplets(outputs, c, debug=debug)
X, P, N = outputs[:, :], outputs[pos_id, :], outputs[neg_id, :]
dpos = torch.sqrt(torch.sum((X - P)**2, dim=1) + eps)
dneg = torch.sqrt(torch.sum((X - N)**2, dim=1) + eps)
loss = torch.mean(
torch.clamp(dpos - dneg + margin, min=0.0, max=margin * 2.0))
# show triplets
if images is not None:
from utils import vis
from datasets import ds_utils
if debug and is_expressions:
for i in range(10):
print(c[:, 0][i].item(), c[pos_id, 0][i].item(),
c[neg_id, 0][i].item())
# ids, vids = c[0], c[1]
# print(vids[:5])
# print(vids[pos_id][:5])
# print(vids[neg_id][:5])
nimgs = 20
losses = to_numpy(
torch.clamp(dpos - dneg + margin, min=0.0, max=margin * 2.0))
# print("Acc: ", 1 - sum(dpos[:nimgs] >= dneg[:nimgs]).item()/float(len(dpos[:nimgs])))
# print("L : ", losses.mean())
images_ref = ds_utils.denormalized(images[:nimgs].clone())
images_pos = ds_utils.denormalized(images[pos_id][:nimgs].clone())
images_neg = ds_utils.denormalized(images[neg_id][:nimgs].clone())
colors = [(0, 1, 0) if c else (1, 0, 0) for c in dpos < dneg]
f = 0.75
images_ref = vis.add_error_to_images(vis.add_cirle_to_images(
images_ref, colors),
losses,
size=1.0,
vmin=0,
vmax=0.5,
thickness=2,
format_string='{:.2f}')
images_pos = vis.add_error_to_images(images_pos,
to_numpy(dpos),
size=1.0,
vmin=0.5,
vmax=1.0,
thickness=2,
format_string='{:.2f}')
images_neg = vis.add_error_to_images(images_neg,
to_numpy(dneg),
size=1.0,
vmin=0.5,
vmax=1.0,
thickness=2,
format_string='{:.2f}')
if is_expressions:
emotions = to_numpy(c[:, 0]).astype(int)
images_ref = vis.add_emotion_to_images(images_ref, emotions)
images_pos = vis.add_emotion_to_images(images_pos,
emotions[pos_id])
images_neg = vis.add_emotion_to_images(images_neg,
emotions[neg_id])
elif feature_name == 'id':
ids = to_numpy(c).astype(int)
images_ref = vis.add_id_to_images(images_ref, ids, loc='tr')
images_pos = vis.add_id_to_images(images_pos,
ids[pos_id],
loc='tr')
images_neg = vis.add_id_to_images(images_neg,
ids[neg_id],
loc='tr')
img_ref = vis.make_grid(images_ref, nCols=nimgs, padsize=1, fx=f, fy=f)
img_pos = vis.make_grid(images_pos, nCols=nimgs, padsize=1, fx=f, fy=f)
img_neg = vis.make_grid(images_neg, nCols=nimgs, padsize=1, fx=f, fy=f)
title = 'triplets'
if feature_name is not None:
title += " " + feature_name
if wnd_title is not None:
title += " " + wnd_title
vis.vis_square([img_ref, img_pos, img_neg],
nCols=1,
padsize=1,
normalize=False,
wait=10,
title=title)
# plt.plot(to_numpy((X[:nimgs]-P[:nimgs]).abs()), 'b')
# plt.plot(to_numpy((X[:nimgs]-N[:nimgs]).abs()), 'r')
# plt.show()
if return_acc:
return loss, sum(dpos >= dneg).item() / float(len(dpos))
else:
return loss