def __init__(self,
geometric_model='affine',
output_size=(240, 240),
crop_factor=1.0,
padding_factor=0.5,
use_cuda=True,
normalize=None):
assert isinstance(output_size, (tuple))
assert isinstance(crop_factor, (float))
assert isinstance(padding_factor, (float))
assert isinstance(use_cuda, (bool))
self.out_h, self.out_w = output_size
self.crop_factor = crop_factor
self.padding_factor = padding_factor
# self.crop_layer = crop_layer
self.use_cuda = use_cuda
self.normalize = normalize
# Initialize an affine transformation to resize the cropped object to (240, 240)
# if self.crop_layer == 'object':
# self.rescalingTnf = GeometricTnf(geometric_model='affine', out_h=self.out_h, out_w=self.out_w,
# use_cuda=self.use_cuda)
# Initialize geometric transformation (tps or affine) to warp the image to form the training pair
self.geometricTnf = GeometricTnf(geometric_model=geometric_model,
out_h=self.out_h,
out_w=self.out_w,
use_cuda=self.use_cuda)
def __init__(self, crop_layer='img'):
super().__init__()
self.crop_layer = crop_layer
if self.crop_layer == 'img':
self.affineTnf = GeometricTnf(geometric_model='affine')
elif self.crop_layer == 'pool4':
self.RoIPool = ROIPool((15, 15), 1.0 / 16.0)
def __init__(self, csv_file, dataset_path, output_size=(240, 240), normalize=None, random_crop=False):
self.dataframe = pd.read_csv(csv_file) # Read images data
self.img_A_names = self.dataframe.iloc[:, 0] # Get source image & target image name
self.img_B_names = self.dataframe.iloc[:, 1]
self.flips = self.dataframe.iloc[:, 3].values.astype('int')
self.dataset_path = dataset_path # Path for reading images
self.out_h, self.out_w = output_size
self.normalize = normalize
self.random_crop = random_crop
# Initialize an affine transformation to resize the image to (240, 240)
self.affineTnf = GeometricTnf(geometric_model='affine', out_h=self.out_h, out_w=self.out_w, use_cuda=False)
def __init__(self,
csv_file,
dataset_path,
output_size=(240, 240),
normalize=None):
self.dataframe = pd.read_csv(csv_file) # Read images data
self.dataset_path = dataset_path # Path for reading images
self.out_h, self.out_w = output_size
self.normalize = normalize
# Initialize an affine transformation to resize the image to (240, 240)
self.affineTnf = GeometricTnf(geometric_model='affine',
out_h=self.out_h,
out_w=self.out_w,
use_cuda=False)
def __init__(self,
geometric_model='affine',
output_size=(240, 240),
crop_factor=1.0,
padding_factor=0.5,
use_cuda=True,
normalize=None):
assert isinstance(output_size, (tuple))
assert isinstance(crop_factor, (float))
assert isinstance(padding_factor, (float))
assert isinstance(use_cuda, (bool))
self.out_h, self.out_w = output_size
self.crop_factor = crop_factor
self.padding_factor = padding_factor
# self.crop_layer = crop_layer
self.use_cuda = use_cuda
self.normalize = normalize
self.geometricTnf = GeometricTnf(geometric_model=geometric_model,
out_h=self.out_h,
out_w=self.out_w,
use_cuda=self.use_cuda)
def __init__(self,
csv_file,
dataset_path,
output_size=(240, 240),
geometric_model='affine',
dataset_size=0,
normalize=None,
random_sample=False,
random_t=0.5,
random_s=0.5,
random_alpha=1 / 6,
random_t_tps=0.4,
random_crop=False):
self.dataframe = pd.read_csv(csv_file) # Read images data
if dataset_size != 0:
dataset_size = min((dataset_size, len(self.dataframe)))
self.dataframe = self.dataframe.iloc[0:dataset_size, :]
self.img_A_names = self.dataframe.iloc[:,
0] # Get source image & target image name
self.img_B_names = self.dataframe.iloc[:, 1]
self.categories = self.dataframe.iloc[:, 2].values
self.flips = self.dataframe.iloc[:, 3].values.astype('int')
self.random_sample = random_sample
if not self.random_sample:
self.theta_array = self.dataframe.iloc[:, 4:].values.astype(
'float') # Get ground-truth tps parameters
self.dataset_path = dataset_path # Path for reading images
self.out_h, self.out_w = output_size
self.geometric_model = geometric_model
self.normalize = normalize
self.random_t = random_t
self.random_s = random_s
self.random_alpha = random_alpha
self.random_t_tps = random_t_tps
self.random_crop = random_crop
# Initialize an affine transformation to resize the image to (240, 240)
self.affineTnf = GeometricTnf(geometric_model='affine',
out_h=self.out_h,
out_w=self.out_w,
use_cuda=False)
def affTpsTnf(self, source_image, theta_aff, theta_aff_tps, use_cuda=True):
tpstnf = GeometricTnf(geometric_model='tps',
out_h=240,
out_w=240,
use_cuda=use_cuda)
sampling_grid_tps = tpstnf(image_batch=source_image,
theta_batch=theta_aff_tps,
padding_factor=0.5,
crop_factor=1.0,
return_sampling_grid=True)[1]
X = sampling_grid_tps[:, :, :, 0].unsqueeze(3)
Y = sampling_grid_tps[:, :, :, 1].unsqueeze(3)
Xp = X * theta_aff[:, 0].unsqueeze(1).unsqueeze(
2) + Y * theta_aff[:, 1].unsqueeze(1).unsqueeze(
2) + theta_aff[:, 2].unsqueeze(1).unsqueeze(2)
Yp = X * theta_aff[:, 3].unsqueeze(1).unsqueeze(
2) + Y * theta_aff[:, 4].unsqueeze(1).unsqueeze(
2) + theta_aff[:, 5].unsqueeze(1).unsqueeze(2)
sampling_grid = torch.cat((Xp, Yp), 3)
warped_image_batch = F.grid_sample(source_image, sampling_grid)
return warped_image_batch
def train_fn_detect(epoch,
model,
faster_rcnn,
aff_theta,
loss_fn,
optimizer,
dataloader,
triple_generation,
use_cuda=True,
log_interval=100,
vis=None,
show=False):
"""
Train the model with synthetically training triple:
{source image, target image, refer image (warped source image), theta_GT} from PF-PASCAL.
1. Train the transformation parameters theta_st from source image to target image;
2. Train the transformation parameters theta_tr from target image to refer image;
3. Combine theta_st and theta_st to obtain theta from source image to refer image, and compute loss between
theta and theta_GT.
"""
tpsTnf = GeometricTnf(geometric_model='tps', use_cuda=use_cuda)
epoch_loss = 0
if (epoch % 5 == 0 or epoch == 1) and vis is not None:
stride_images = len(dataloader) / 3
watch_images = torch.Tensor(24, 3, 240, 240)
# means for normalize of caffe resnet and vgg
# pixel_means = torch.Tensor(np.array([[[[102.9801, 115.9465, 122.7717]]]]).astype(np.float32)).cuda()
stride_loss = len(dataloader) / 35
iter_loss = np.zeros(36)
begin = time.time()
for batch_idx, batch in enumerate(dataloader):
''' Move input batch to gpu '''
# batch['source_image'].shape & batch['target_image'].shape: (batch_size, 3, 240, 240)
# batch['theta'].shape-tps: (batch_size, 18)-random or (batch_size, 18, 1, 1)-(pre-set from csv)
if use_cuda:
batch = batch_cuda(batch)
''' Get the training triple {source image, target image, refer image (warped source image), theta_GT}'''
batch_triple = triple_generation(batch)
''' Train the model '''
optimizer.zero_grad()
# Predict tps parameters between images
box_info_s = faster_rcnn(
im_data=batch_triple['source_im'],
im_info=batch_triple['source_im_info'][:, 3:],
gt_boxes=batch_triple['source_gt_boxes'],
num_boxes=batch_triple['source_num_boxes'])[0:3]
box_info_t = faster_rcnn(
im_data=batch_triple['target_im'],
im_info=batch_triple['target_im_info'][:, 3:],
gt_boxes=batch_triple['target_gt_boxes'],
num_boxes=batch_triple['target_num_boxes'])[0:3]
box_info_r = faster_rcnn(
im_data=batch_triple['refer_im'],
im_info=batch_triple['refer_im_info'][:, 3:],
gt_boxes=batch_triple['refer_gt_boxes'],
num_boxes=batch_triple['refer_num_boxes'])[0:3]
all_box_s = select_boxes(rois=box_info_s[0],
cls_prob=box_info_s[1],
bbox_pred=box_info_s[2],
im_infos=batch_triple['source_im_info'][:,
3:])
all_box_t = select_boxes(rois=box_info_t[0],
cls_prob=box_info_t[1],
bbox_pred=box_info_t[2],
im_infos=batch_triple['target_im_info'][:,
3:])
all_box_r = select_boxes(rois=box_info_r[0],
cls_prob=box_info_r[1],
bbox_pred=box_info_r[2],
im_infos=batch_triple['source_im_info'][:,
3:])
box_s, box_t, box_r = select_box(all_box_s, all_box_t, all_box_r)
theta_st = aff_theta(boxes_s=box_s, boxes_t=box_t)
theta_tr = aff_theta(boxes_s=box_t, boxes_t=box_r)
# theta.shape: (batch_size, 18) for tps
batch_st = {
'source_image': batch_triple['source_image'],
'target_image': batch_triple['target_image']
}
batch_tr = {
'source_image': batch_triple['target_image'],
'target_image': batch_triple['refer_image']
}
theta_aff_tps_st, theta_aff_st = model(
batch_st, theta_st) # from source image to target image
theta_aff_tps_tr, theta_aff_tr = model(
batch_tr, theta_tr) # from target image to refer image
# show_images(batch_st=batch_st, batch_tr=batch_tr, box_s=box_s, box_t=box_t, box_r=box_r)
loss = loss_fn(theta_st=theta_aff_tps_st,
theta_tr=theta_aff_tps_tr,
theta_GT=batch_triple['theta_GT'])
loss.backward()
optimizer.step()
epoch_loss += loss.item()
if (batch_idx + 1) % log_interval == 0:
end = time.time()
print(
'Train epoch: {} [{}/{} ({:.0%})]\t\tCurrent batch loss: {:.6f}\t\tTime cost ({} batches): {:.4f} s'
.format(epoch, batch_idx + 1,
len(dataloader), (batch_idx + 1) / len(dataloader),
loss.item(), batch_idx + 1, end - begin))
if (epoch % 5 == 0 or epoch == 1) and vis is not None:
if (batch_idx + 1) % stride_images == 0 or batch_idx == 0:
watch_images = add_watch(
watch_images, batch_st, batch_tr, tpsTnf, theta_aff_tps_st,
theta_aff_tps_tr,
int((batch_idx + 1) / stride_images) * 6)
if (batch_idx + 1) % stride_loss == 0 or batch_idx == 0:
iter_loss[int((batch_idx + 1) /
stride_loss)] = epoch_loss / (batch_idx + 1)
# tmp_images = batch_triple['target_im'].permute(0, 2, 3, 1) + pixel_means
# tmp_images = tmp_images[:, :, :, [2, 1, 0]].permute(0, 3, 1, 2)
# vis.image(torchvision.utils.make_grid(tmp_images, nrow=1, padding=3))
if show:
# if dual:
# warped_image_aff = affTnf(batch_st['source_image'], theta_aff_st_1)
# warped_image_aff_2 = affTnf(batch_tr['source_image'], theta_aff_tr_1)
# warped_image_aff_3 = affTnf(warped_image_aff, theta_aff_tr_1)
# show_images(batch_st, batch_tr, warped_image_aff.detach(), warped_image_aff_2.detach(), warped_image_aff_3.detach())
#
# warped_image_aff = affTnf(batch_st['source_image'], theta_aff_st)
# warped_image_aff_2 = affTnf(batch_tr['source_image'], theta_aff_tr)
# warped_image_aff_3 = affTnf(warped_image_aff, theta_aff_tr)
# show_images(batch_st, batch_tr, warped_image_aff.detach(), warped_image_aff_2.detach(), warped_image_aff_3.detach())
#
# warped_image_aff_tps = tpsTnf(batch_st['source_image'], theta_aff_tps_st)
# warped_image_aff_tps_2 = tpsTnf(batch_tr['source_image'], theta_aff_tps_tr)
# warped_image_aff_tps_3 = tpsTnf(warped_image_aff_tps, theta_aff_tps_tr)
# show_images(batch_st, batch_tr, warped_image_aff_tps.detach(), warped_image_aff_tps_2.detach(), warped_image_aff_tps_3.detach())
#
# warped_image = affTnf(batch_st['source_image'], theta_aff_st_1)
# warped_image = affTnf(warped_image, theta_aff_st)
# warped_image = tpsTnf(warped_image, theta_aff_tps_st)
# warped_image_2 = affTnf(batch_tr['source_image'], theta_aff_tr_1)
# warped_image_2 = affTnf(warped_image_2, theta_aff_tr)
# warped_image_2 = tpsTnf(warped_image_2, theta_aff_tps_tr)
# warped_image_3 = affTnf(warped_image, theta_aff_tr_1)
# warped_image_3 = affTnf(warped_image_3, theta_aff_tr)
# warped_image_3 = tpsTnf(warped_image_3, theta_aff_tps_tr)
# show_images(batch_st, batch_tr, warped_image.detach(), warped_image_2.detach(), warped_image_3.detach())
# else:
# warped_image_aff = affTnf(batch_st['source_image'], theta_st)
# warped_image_aff_2 = affTnf(batch_tr['source_image'], theta_tr)
# warped_image_aff_3 = affTnf(warped_image_aff, theta_tr)
# show_images(batch_st, batch_tr, warped_image_aff.detach(), warped_image_aff_2.detach(), warped_image_aff_3.detach())
warped_image_tps = tpsTnf(batch_st['source_image'],
theta_aff_tps_st)
warped_image_tps_2 = tpsTnf(batch_tr['source_image'],
theta_aff_tps_tr)
warped_image_tps_3 = tpsTnf(warped_image_tps, theta_aff_tps_tr)
show_images(batch_triple, warped_image_tps.detach(),
warped_image_tps_2.detach(),
warped_image_tps_3.detach(), box_s, box_t, box_r)
end = time.time()
# Visualize watch images & train loss
if (epoch % 5 == 0 or epoch == 1) and vis is not None:
opts = dict(jpgquality=100,
title='Epoch ' + str(epoch) +
' source warped_sr target warped_tr refer warped_sr')
# Un-normalize for caffe resnet and vgg
# watch_images = watch_images.permute(0, 2, 3, 1) + pixel_means
# watch_images = watch_images[:, :, :, [2, 1, 0]].permute(0, 3, 1, 2)
watch_images = normalize_image(watch_images, forward=False) * 255.0
vis.image(torchvision.utils.make_grid(watch_images, nrow=6, padding=3),
opts=opts)
# vis.images(watch_images, nrow=6, padding=3, opts=opts)
opts_loss = dict(
xlabel='Iterations (' + str(stride_loss) + ')',
ylabel='Loss',
title='GM ResNet101 Detect&Affine&TPS Training Loss in Epoch ' +
str(epoch),
legend=['Loss'],
width=2000)
vis.line(iter_loss, np.arange(36), opts=opts_loss)
epoch_loss /= len(dataloader)
print('Train set -- Average loss: {:.6f}\t\tTime cost: {:.4f}'.format(
epoch_loss, end - begin))
return epoch_loss, end - begin
def train_fn_dual(epoch, model, loss_fn, optimizer, dataloader, triple_generation, use_cuda=True, log_interval=100, vis=None, show=False):
"""
Train the model with synthetically training triple:
{source image, target image, refer image (warped source image), theta_GT} from PF-PASCAL.
1. Train the transformation parameters theta_st from source image to target image;
2. Train the transformation parameters theta_tr from target image to refer image;
3. Combine theta_st and theta_st to obtain theta from source image to refer image, and compute loss between
theta and theta_GT.
"""
geoTnf = ComposedGeometricTnf(use_cuda=use_cuda)
affTnf = GeometricTnf(geometric_model='affine', use_cuda=use_cuda)
tpsTnf = GeometricTnf(geometric_model='tps', use_cuda=use_cuda)
epoch_loss = 0
if (epoch % 5 == 0 or epoch == 1) and vis is not None:
stride_images = len(dataloader) / 3
group_size = 6
watch_images = torch.ones(group_size * 4, 3, 260, 240).cuda()
# watch_images = torch.ones(group_size * 20, 3, 260, 240).cuda()
image_names = list()
fnt = cv2.FONT_HERSHEY_COMPLEX
# means for normalize of caffe resnet and vgg
# pixel_means = torch.Tensor(np.array([[[[102.9801, 115.9465, 122.7717]]]]).astype(np.float32)).cuda()
stride_loss = len(dataloader) / 105
iter_loss = np.zeros(106)
begin = time.time()
for batch_idx, batch in enumerate(dataloader):
''' Move input batch to gpu '''
# batch['source_image'].shape & batch['target_image'].shape: (batch_size, 3, 240, 240)
# batch['theta'].shape-tps: (batch_size, 18)-random or (batch_size, 18, 1, 1)-(pre-set from csv)
if use_cuda:
batch = batch_cuda(batch)
''' Get the training triple {source image, target image, refer image (warped source image), theta_GT}'''
batch_triple = triple_generation(batch)
''' Train the model '''
optimizer.zero_grad()
# Predict tps parameters between images
# theta.shape: (batch_size, 18) for tps
batch_st = {'source_image': batch_triple['source_image'], 'target_image': batch_triple['target_image']}
batch_tr = {'source_image': batch_triple['target_image'], 'target_image': batch_triple['refer_image']}
theta_aff_tps_st, theta_aff_st = model(batch_st) # from source image to target image
theta_aff_tps_tr, theta_aff_tr = model(batch_tr) # from target image to refer image
# show_images(batch_st=batch_st, batch_tr=batch_tr, box_s=box_s, box_t=box_t, box_r=box_r)
# loss = loss_fn(theta_st=theta_aff_tps_st, theta_tr=theta_aff_tps_tr, theta_GT=batch_triple['theta_GT'])
loss = loss_fn(theta_aff_tps_st=theta_aff_tps_st, theta_aff_st=theta_aff_st, theta_aff_tps_tr=theta_aff_tps_tr,
theta_aff_tr=theta_aff_tr, theta_aff_GT=batch_triple['theta_aff_GT'], theta_tps_GT=batch_triple['theta_tps_GT'])
loss.backward()
optimizer.step()
epoch_loss += loss.item()
if (batch_idx+1) % log_interval == 0:
end = time.time()
print('Train epoch: {} [{}/{} ({:.0%})]\t\tCurrent batch loss: {:.6f}\t\tTime cost ({} batches): {:.4f} s'
.format(epoch, batch_idx+1, len(dataloader), (batch_idx+1) / len(dataloader), loss.item(), batch_idx + 1, end - begin))
if (epoch % 5 == 0 or epoch == 1) and vis is not None:
if (batch_idx + 1) % stride_images == 0 or batch_idx == 0:
watch_images, image_names = add_watch(watch_images, image_names, batch_st, batch_tr, affTnf, tpsTnf, geoTnf, theta_aff_tps_st, theta_aff_st, theta_aff_tps_tr, theta_aff_tr, int((batch_idx + 1) / stride_images) * group_size)
# if batch_idx <= 19:
# watch_images, image_names = add_watch(watch_images, image_names, batch_st, batch_tr, affTnf, tpsTnf, geoTnf, theta_aff_tps_st, theta_aff_st, theta_aff_tps_tr, theta_aff_tr, batch_idx * group_size, batch_triple)
# if batch_idx == 19:
# opts = dict(jpgquality=100, title='Epoch ' + str(epoch) + ' source warped_st target warped_tr refer warped_sr')
# watch_images[:, :, 0:240, :] = normalize_image(watch_images[:, :, 0:240, :], forward=False)
# watch_images *= 255.0
# watch_images = watch_images.permute(0, 2, 3, 1).cpu().numpy().astype(np.uint8)
# for i in range(len(image_names)):
# cv2.putText(watch_images[i], image_names[i], (80, 255), fnt, 0.5, (0, 0, 0), 1)
# watch_images = torch.Tensor(watch_images.astype(np.float32))
# watch_images = watch_images.permute(0, 3, 1, 2)
# vis.image(torchvision.utils.make_grid(watch_images, nrow=group_size, padding=3), opts=opts)
if (batch_idx + 1) % stride_loss == 0 or batch_idx == 0:
iter_loss[int((batch_idx + 1) / stride_loss)] = epoch_loss / (batch_idx + 1)
# tmp_images = normalize_image(batch_tr['target_image'], forward=False) * 255.0
# vis.images(tmp_images, nrow=8, padding=3)
# if show:
# if dual:
# warped_image_aff = affTnf(batch_st['source_image'], theta_aff_st_1)
# warped_image_aff_2 = affTnf(batch_tr['source_image'], theta_aff_tr_1)
# warped_image_aff_3 = affTnf(warped_image_aff, theta_aff_tr_1)
# show_images(batch_st, batch_tr, warped_image_aff.detach(), warped_image_aff_2.detach(), warped_image_aff_3.detach())
#
# warped_image_aff = affTnf(batch_st['source_image'], theta_aff_st)
# warped_image_aff_2 = affTnf(batch_tr['source_image'], theta_aff_tr)
# warped_image_aff_3 = affTnf(warped_image_aff, theta_aff_tr)
# show_images(batch_st, batch_tr, warped_image_aff.detach(), warped_image_aff_2.detach(), warped_image_aff_3.detach())
#
# warped_image_aff_tps = tpsTnf(batch_st['source_image'], theta_aff_tps_st)
# warped_image_aff_tps_2 = tpsTnf(batch_tr['source_image'], theta_aff_tps_tr)
# warped_image_aff_tps_3 = tpsTnf(warped_image_aff_tps, theta_aff_tps_tr)
# show_images(batch_st, batch_tr, warped_image_aff_tps.detach(), warped_image_aff_tps_2.detach(), warped_image_aff_tps_3.detach())
#
# warped_image = affTnf(batch_st['source_image'], theta_aff_st_1)
# warped_image = affTnf(warped_image, theta_aff_st)
# warped_image = tpsTnf(warped_image, theta_aff_tps_st)
# warped_image_2 = affTnf(batch_tr['source_image'], theta_aff_tr_1)
# warped_image_2 = affTnf(warped_image_2, theta_aff_tr)
# warped_image_2 = tpsTnf(warped_image_2, theta_aff_tps_tr)
# warped_image_3 = affTnf(warped_image, theta_aff_tr_1)
# warped_image_3 = affTnf(warped_image_3, theta_aff_tr)
# warped_image_3 = tpsTnf(warped_image_3, theta_aff_tps_tr)
# show_images(batch_st, batch_tr, warped_image.detach(), warped_image_2.detach(), warped_image_3.detach())
# else:
# warped_image_aff = affTnf(batch_st['source_image'], theta_st)
# warped_image_aff_2 = affTnf(batch_tr['source_image'], theta_tr)
# warped_image_aff_3 = affTnf(warped_image_aff, theta_tr)
# show_images(batch_st, batch_tr, warped_image_aff.detach(), warped_image_aff_2.detach(), warped_image_aff_3.detach())
# warped_image_tps = tpsTnf(batch_st['source_image'], theta_st)
# warped_image_tps_2 = tpsTnf(batch_tr['source_image'], theta_tr)
# warped_image_tps_3 = tpsTnf(warped_image_tps, theta_tr)
# show_images(batch_st, batch_tr, warped_image_tps.detach(), warped_image_tps_2.detach(), warped_image_tps_3.detach())
end = time.time()
# Visualize watch images & train loss
if (epoch % 5 == 0 or epoch == 1) and vis is not None:
opts = dict(jpgquality=100, title='Epoch ' + str(epoch) + ' source warped_st target warped_tr refer warped_sr')
watch_images[:, :, 0:240, :] = normalize_image(watch_images[:, :, 0:240, :], forward=False)
watch_images *= 255.0
watch_images = watch_images.permute(0, 2, 3, 1).cpu().numpy().astype(np.uint8)
for i in range(watch_images.shape[0]):
cv2.putText(watch_images[i], image_names[i], (80, 255), fnt, 0.5, (0, 0, 0), 1)
watch_images = torch.Tensor(watch_images.astype(np.float32))
watch_images = watch_images.permute(0, 3, 1, 2)
vis.image(torchvision.utils.make_grid(watch_images, nrow=group_size, padding=3), opts=opts)
# Un-normalize for caffe resnet and vgg
# watch_images = watch_images.permute(0, 2, 3, 1) + pixel_means
# watch_images = watch_images[:, :, :, [2, 1, 0]].permute(0, 3, 1, 2)
opts_loss = dict(xlabel='Iterations (' + str(stride_loss) + ')',
ylabel='Loss',
title='GM ResNet101 AffTPS Training Loss in Epoch ' + str(epoch),
legend=['Loss'],
width=2000)
vis.line(iter_loss, np.arange(106), opts=opts_loss)
epoch_loss /= len(dataloader)
print('Train set -- Average loss: {:.6f}\t\tTime cost: {:.4f}'.format(epoch_loss, end - begin))
return epoch_loss, end - begin
def vis_pf(vis,
dataloader,
theta,
theta_weak,
theta_inver,
theta_weak_inver,
results,
results_weak,
dataset_name,
use_cuda=True):
# Visualize watch images
affTnf = GeometricTnf(geometric_model='affine', use_cuda=use_cuda)
tpsTnf = GeometricTnf(geometric_model='tps', use_cuda=use_cuda)
pt = PointTnf(use_cuda=use_cuda)
watch_images = torch.ones(len(dataloader) * 6, 3, 280, 240)
watch_keypoints = -torch.ones(len(dataloader) * 6, 2, 20)
if use_cuda:
watch_images = watch_images.cuda()
watch_keypoints = watch_keypoints.cuda()
num_points = np.ones(len(dataloader) * 6).astype(np.int8)
correct_index = list()
image_names = list()
metrics = list()
# Colors for keypoints
cmap = plt.get_cmap('tab20')
colors = list()
for c in range(20):
r = cmap(c)[0] * 255
g = cmap(c)[1] * 255
b = cmap(c)[2] * 255
colors.append((b, g, r))
fnt = cv2.FONT_HERSHEY_COMPLEX
# means for normalize of caffe resnet and vgg
# pixel_means = torch.Tensor(np.array([[[[102.9801, 115.9465, 122.7717]]]]).astype(np.float32))
for batch_idx, batch in enumerate(dataloader):
if use_cuda:
batch = batch_cuda(batch)
# Theta and theta_inver
theta_aff = theta['aff'][batch_idx].unsqueeze(0)
theta_aff_tps = theta['aff_tps'][batch_idx].unsqueeze(0)
theta_weak_aff = theta_weak['aff'][batch_idx].unsqueeze(0)
theta_weak_aff_tps = theta_weak['aff_tps'][batch_idx].unsqueeze(0)
theta_aff_inver = theta_inver['aff'][batch_idx].unsqueeze(0)
theta_aff_tps_inver = theta_inver['aff_tps'][batch_idx].unsqueeze(0)
theta_weak_aff_inver = theta_weak_inver['aff'][batch_idx].unsqueeze(0)
theta_weak_aff_tps_inver = theta_weak_inver['aff_tps'][
batch_idx].unsqueeze(0)
# Warped image
warped_aff = affTnf(batch['source_image'], theta_aff)
warped_aff_tps = tpsTnf(warped_aff, theta_aff_tps)
warped_weak_aff = affTnf(batch['source_image'], theta_weak_aff)
warped_weak_aff_tps = tpsTnf(warped_weak_aff, theta_weak_aff_tps)
watch_images[batch_idx * 6, :, 0:240, :] = batch['source_image']
watch_images[batch_idx * 6 + 1, :, 0:240, :] = warped_aff
watch_images[batch_idx * 6 + 2, :, 0:240, :] = warped_aff_tps
watch_images[batch_idx * 6 + 3, :, 0:240, :] = batch['target_image']
watch_images[batch_idx * 6 + 4, :, 0:240, :] = warped_weak_aff
watch_images[batch_idx * 6 + 5, :, 0:240, :] = warped_weak_aff_tps
# Warped keypoints
source_im_size = batch['source_im_info'][:, 0:3]
target_im_size = batch['target_im_info'][:, 0:3]
source_points = batch['source_points']
target_points = batch['target_points']
source_points_norm = PointsToUnitCoords(P=source_points,
im_size=source_im_size)
target_points_norm = PointsToUnitCoords(P=target_points,
im_size=target_im_size)
warped_points_aff_norm = pt.affPointTnf(theta=theta_aff_inver,
points=source_points_norm)
warped_points_aff = PointsToPixelCoords(P=warped_points_aff_norm,
im_size=target_im_size)
pck_aff, index_aff, N_pts = pck(target_points, warped_points_aff,
dataset_name)
warped_points_aff = relocate(warped_points_aff, target_im_size)
warped_points_aff_tps_norm = pt.tpsPointTnf(theta=theta_aff_tps_inver,
points=source_points_norm)
warped_points_aff_tps_norm = pt.affPointTnf(
theta=theta_aff_inver, points=warped_points_aff_tps_norm)
warped_points_aff_tps = PointsToPixelCoords(
P=warped_points_aff_tps_norm, im_size=target_im_size)
pck_aff_tps, index_aff_tps, _ = pck(target_points,
warped_points_aff_tps,
dataset_name)
warped_points_aff_tps = relocate(warped_points_aff_tps, target_im_size)
warped_points_weak_aff_norm = pt.affPointTnf(
theta=theta_weak_aff_inver, points=source_points_norm)
warped_points_weak_aff = PointsToPixelCoords(
P=warped_points_weak_aff_norm, im_size=target_im_size)
pck_weak_aff, index_weak_aff, _ = pck(target_points,
warped_points_weak_aff,
dataset_name)
warped_points_weak_aff = relocate(warped_points_weak_aff,
target_im_size)
warped_points_weak_aff_tps_norm = pt.tpsPointTnf(
theta=theta_weak_aff_tps_inver, points=source_points_norm)
warped_points_weak_aff_tps_norm = pt.affPointTnf(
theta=theta_weak_aff_inver, points=warped_points_weak_aff_tps_norm)
warped_points_weak_aff_tps = PointsToPixelCoords(
P=warped_points_weak_aff_tps_norm, im_size=target_im_size)
pck_weak_aff_tps, index_weak_aff_tps, _ = pck(
target_points, warped_points_weak_aff_tps, dataset_name)
warped_points_weak_aff_tps = relocate(warped_points_weak_aff_tps,
target_im_size)
watch_keypoints[batch_idx * 6, :, :N_pts] = relocate(
batch['source_points'], source_im_size)[:, :, :N_pts]
watch_keypoints[batch_idx * 6 +
1, :, :N_pts] = warped_points_aff[:, :, :N_pts]
watch_keypoints[batch_idx * 6 +
2, :, :N_pts] = warped_points_aff_tps[:, :, :N_pts]
watch_keypoints[batch_idx * 6 + 3, :, :N_pts] = relocate(
batch['target_points'], target_im_size)[:, :, :N_pts]
watch_keypoints[batch_idx * 6 +
4, :, :N_pts] = warped_points_weak_aff[:, :, :N_pts]
watch_keypoints[
batch_idx * 6 +
5, :, :N_pts] = warped_points_weak_aff_tps[:, :, :N_pts]
num_points[batch_idx * 6:batch_idx * 6 + 6] = N_pts
correct_index.append(np.arange(N_pts))
correct_index.append(index_aff)
correct_index.append(index_aff_tps)
correct_index.append(np.arange(N_pts))
correct_index.append(index_weak_aff)
correct_index.append(index_weak_aff_tps)
image_names.append('Source')
image_names.append('Aff')
image_names.append('Aff_tps')
image_names.append('Target')
image_names.append('Rocco_aff')
image_names.append('Rocco_aff_tps')
metrics.append('')
metrics.append('PCK: {:.2%}'.format(pck_aff))
metrics.append('PCK: {:.2%}'.format(pck_aff_tps))
metrics.append('')
metrics.append('PCK: {:.2%}'.format(pck_weak_aff))
metrics.append('PCK: {:.2%}'.format(pck_weak_aff_tps))
opts = dict(jpgquality=100, title=dataset_name)
# Un-normalize for caffe resnet and vgg
# watch_images = watch_images.permute(0, 2, 3, 1) + pixel_means
# watch_images = watch_images[:, :, :, [2, 1, 0]].permute(0, 3, 1, 2)
# watch_images = normalize_image(watch_images, forward=False) * 255.0
watch_images[:, :, 0:240, :] = normalize_image(watch_images[:, :,
0:240, :],
forward=False)
watch_images *= 255.0
watch_images = watch_images.permute(0, 2, 3,
1).cpu().numpy().astype(np.uint8)
watch_keypoints = watch_keypoints.cpu().numpy()
for i in range(watch_images.shape[0]):
pos_name = (80, 255)
if (i + 1) % 6 == 1 or (i + 1) % 6 == 4:
pos_pck = (0, 0)
else:
pos_pck = (70, 275)
cv2.putText(watch_images[i], image_names[i], pos_name, fnt, 0.5,
(0, 0, 0), 1)
cv2.putText(watch_images[i], metrics[i], pos_pck, fnt, 0.5, (0, 0, 0),
1)
if (i + 1) % 6 == 4:
for j in range(num_points[i]):
cv2.drawMarker(
watch_images[i],
(watch_keypoints[i, 0, j], watch_keypoints[i, 1, j]),
colors[j], cv2.MARKER_DIAMOND, 12, 2, cv2.LINE_AA)
else:
for j in correct_index[i]:
cv2.drawMarker(
watch_images[i],
(watch_keypoints[i, 0, j], watch_keypoints[i, 1, j]),
colors[j], cv2.MARKER_CROSS, 12, 2, cv2.LINE_AA)
cv2.drawMarker(watch_images[i],
(watch_keypoints[i + 3 - (i % 6), 0, j],
watch_keypoints[i + 3 - (i % 6), 1, j]),
colors[j], cv2.MARKER_DIAMOND, 12, 2,
cv2.LINE_AA)
watch_images = torch.Tensor(watch_images.astype(np.float32))
watch_images = watch_images.permute(0, 3, 1, 2)
vis.image(torchvision.utils.make_grid(watch_images, nrow=3, padding=3),
opts=opts)
def vis_tss(vis,
dataloader,
theta,
theta_weak,
csv_file,
title,
use_cuda=True):
# Visualize watch images
dataframe = pd.read_csv(csv_file)
scores_aff = dataframe.iloc[:, 5]
scores_aff_tps = dataframe.iloc[:, 6]
scores_weak_aff = dataframe.iloc[:, 7]
scores_weak_aff_tps = dataframe.iloc[:, 8]
affTnf = GeometricTnf(geometric_model='affine', use_cuda=use_cuda)
tpsTnf = GeometricTnf(geometric_model='tps', use_cuda=use_cuda)
watch_images = torch.ones(int(len(dataloader) / 2 * 6), 3, 280, 240)
watch_images_inver = torch.ones(int(len(dataloader) / 2 * 6), 3, 280, 240)
if use_cuda:
watch_images = watch_images.cuda()
watch_images_inver = watch_images_inver.cuda()
image_names = list()
image_names_inver = list()
flow = list()
flow_inver = list()
fnt = cv2.FONT_HERSHEY_COMPLEX
# means for normalize of caffe resnet and vgg
# pixel_means = torch.Tensor(np.array([[[[102.9801, 115.9465, 122.7717]]]]).astype(np.float32))
for batch_idx, batch in enumerate(dataloader):
if use_cuda:
batch = batch_cuda(batch)
theta_aff = theta['aff'][batch_idx].unsqueeze(0)
theta_aff_tps = theta['aff_tps'][batch_idx].unsqueeze(0)
theta_weak_aff = theta_weak['aff'][batch_idx].unsqueeze(0)
theta_weak_aff_tps = theta_weak['aff_tps'][batch_idx].unsqueeze(0)
# Warped image
warped_aff = affTnf(batch['source_image'], theta_aff)
warped_aff_tps = tpsTnf(warped_aff, theta_aff_tps)
warped_weak_aff = affTnf(batch['source_image'], theta_weak_aff)
warped_weak_aff_tps = tpsTnf(warped_weak_aff, theta_weak_aff_tps)
if (batch_idx + 1) % 2 != 0:
watch_images[int(batch_idx / 2 * 6), :,
0:240, :] = batch['source_image']
watch_images[int(batch_idx / 2 * 6) + 1, :, 0:240, :] = warped_aff
watch_images[int(batch_idx / 2 * 6) + 2, :,
0:240, :] = warped_aff_tps
watch_images[int(batch_idx / 2 * 6) + 3, :,
0:240, :] = batch['target_image']
watch_images[int(batch_idx / 2 * 6) + 4, :,
0:240, :] = warped_weak_aff
watch_images[int(batch_idx / 2 * 6) + 5, :,
0:240, :] = warped_weak_aff_tps
image_names.append('Source')
image_names.append('Aff')
image_names.append('Aff_tps')
image_names.append('Target')
image_names.append('Rocco_aff')
image_names.append('Rocco_aff_tps')
flow.append('')
flow.append('Flow: {:.3f}'.format(scores_aff[batch_idx]))
flow.append('Flow: {:.3f}'.format(scores_aff_tps[batch_idx]))
flow.append('')
flow.append('Flow: {:.3f}'.format(scores_weak_aff[batch_idx]))
flow.append('Flow: {:.3f}'.format(scores_weak_aff_tps[batch_idx]))
else:
watch_images_inver[int((batch_idx - 1) / 2 * 6), :,
0:240, :] = batch['source_image']
watch_images_inver[int((batch_idx - 1) / 2 * 6) + 1, :,
0:240, :] = warped_aff
watch_images_inver[int((batch_idx - 1) / 2 * 6) + 2, :,
0:240, :] = warped_aff_tps
watch_images_inver[int((batch_idx - 1) / 2 * 6) + 3, :,
0:240, :] = batch['target_image']
watch_images_inver[int((batch_idx - 1) / 2 * 6) + 4, :,
0:240, :] = warped_weak_aff
watch_images_inver[int((batch_idx - 1) / 2 * 6) + 5, :,
0:240, :] = warped_weak_aff_tps
image_names_inver.append('Source')
image_names_inver.append('Aff')
image_names_inver.append('Aff_tps')
image_names_inver.append('Target')
image_names_inver.append('Rocco_aff')
image_names_inver.append('Rocco_aff_tps')
flow_inver.append('')
flow_inver.append('Flow: {:.3f}'.format(scores_aff[batch_idx]))
flow_inver.append('Flow: {:.3f}'.format(scores_aff_tps[batch_idx]))
flow_inver.append('')
flow_inver.append('Flow: {:.3f}'.format(
scores_weak_aff[batch_idx]))
flow_inver.append('Flow: {:.3f}'.format(
scores_weak_aff_tps[batch_idx]))
opts = dict(jpgquality=100, title=title)
# Un-normalize for caffe resnet and vgg
# watch_images = watch_images.permute(0, 2, 3, 1) + pixel_means
# watch_images = watch_images[:, :, :, [2, 1, 0]].permute(0, 3, 1, 2)
# watch_images = normalize_image(watch_images, forward=False) * 255.0
def draw_image(images, names, flows):
images[:, :, 0:240, :] = normalize_image(images[:, :, 0:240, :],
forward=False)
images *= 255.0
images = images.permute(0, 2, 3, 1).cpu().numpy().astype(np.uint8)
for i in range(images.shape[0]):
pos_name = (80, 255)
if (i + 1) % 6 == 1 or (i + 1) % 6 == 4:
pos_lt_ac = (0, 0)
pos_flow = (0, 0)
else:
pos_flow = (70, 275)
cv2.putText(images[i], names[i], pos_name, fnt, 0.5, (0, 0, 0), 1)
cv2.putText(images[i], flows[i], pos_flow, fnt, 0.5, (0, 0, 0), 1)
images = torch.Tensor(images.astype(np.float32))
images = images.permute(0, 3, 1, 2)
return images
watch_images = draw_image(images=watch_images,
names=image_names,
flows=flow)
watch_images_inver = draw_image(images=watch_images_inver,
names=image_names_inver,
flows=flow_inver)
vis.image(torchvision.utils.make_grid(watch_images, nrow=3, padding=5),
opts=opts)
def vis_caltech(vis,
dataloader,
theta,
theta_weak,
results,
results_weak,
title,
use_cuda=True):
# Visualize watch images
affTnf = GeometricTnf(geometric_model='affine', use_cuda=use_cuda)
tpsTnf = GeometricTnf(geometric_model='tps', use_cuda=use_cuda)
watch_images = torch.ones(len(dataloader) * 6, 3, 280, 240)
if use_cuda:
watch_images = watch_images.cuda()
image_names = list()
lt_acc = list()
iou = list()
fnt = cv2.FONT_HERSHEY_COMPLEX
# means for normalize of caffe resnet and vgg
# pixel_means = torch.Tensor(np.array([[[[102.9801, 115.9465, 122.7717]]]]).astype(np.float32))
for batch_idx, batch in enumerate(dataloader):
if use_cuda:
batch = batch_cuda(batch)
theta_aff = theta['aff'][batch_idx].unsqueeze(0)
theta_aff_tps = theta['aff_tps'][batch_idx].unsqueeze(0)
theta_weak_aff = theta_weak['aff'][batch_idx].unsqueeze(0)
theta_weak_aff_tps = theta_weak['aff_tps'][batch_idx].unsqueeze(0)
# Warped image
warped_aff = affTnf(batch['source_image'], theta_aff)
warped_aff_tps = tpsTnf(warped_aff, theta_aff_tps)
warped_weak_aff = affTnf(batch['source_image'], theta_weak_aff)
warped_weak_aff_tps = tpsTnf(warped_weak_aff, theta_weak_aff_tps)
watch_images[batch_idx * 6, :, 0:240, :] = batch['source_image']
watch_images[batch_idx * 6 + 1, :, 0:240, :] = warped_aff
watch_images[batch_idx * 6 + 2, :, 0:240, :] = warped_aff_tps
watch_images[batch_idx * 6 + 3, :, 0:240, :] = batch['target_image']
watch_images[batch_idx * 6 + 4, :, 0:240, :] = warped_weak_aff
watch_images[batch_idx * 6 + 5, :, 0:240, :] = warped_weak_aff_tps
image_names.append('Source')
image_names.append('Aff')
image_names.append('Aff_tps')
image_names.append('Target')
image_names.append('Rocco_aff')
image_names.append('Rocco_aff_tps')
lt_acc.append('')
lt_acc.append('LT-ACC: {:.2f}'.format(
float(results['aff']['label_transfer_accuracy'][batch_idx])))
lt_acc.append('LT-ACC: {:.2f}'.format(
float(results['aff_tps']['label_transfer_accuracy'][batch_idx])))
lt_acc.append('')
lt_acc.append('LT-ACC: {:.2f}'.format(
float(results_weak['aff']['label_transfer_accuracy'][batch_idx])))
lt_acc.append('LT-ACC: {:.2f}'.format(
float(results_weak['aff_tps']['label_transfer_accuracy']
[batch_idx])))
iou.append('')
iou.append('IoU: {:.2f}'.format(
float(results['aff']['intersection_over_union'][batch_idx])))
iou.append('IoU: {:.2f}'.format(
float(results['aff_tps']['intersection_over_union'][batch_idx])))
iou.append('')
iou.append('IoU: {:.2f}'.format(
float(results_weak['aff']['intersection_over_union'][batch_idx])))
iou.append('IoU: {:.2f}'.format(
float(results_weak['aff_tps']['intersection_over_union']
[batch_idx])))
opts = dict(jpgquality=100, title=title)
# Un-normalize for caffe resnet and vgg
# watch_images = watch_images.permute(0, 2, 3, 1) + pixel_means
# watch_images = watch_images[:, :, :, [2, 1, 0]].permute(0, 3, 1, 2)
# watch_images = normalize_image(watch_images, forward=False) * 255.0
watch_images[:, :, 0:240, :] = normalize_image(watch_images[:, :,
0:240, :],
forward=False)
watch_images *= 255.0
watch_images = watch_images.permute(0, 2, 3,
1).cpu().numpy().astype(np.uint8)
for i in range(watch_images.shape[0]):
pos_name = (80, 255)
if (i + 1) % 6 == 1 or (i + 1) % 6 == 4:
pos_lt_ac = (0, 0)
pos_iou = (0, 0)
else:
pos_lt_ac = (10, 275)
pos_iou = (140, 275)
cv2.putText(watch_images[i], image_names[i], pos_name, fnt, 0.5,
(0, 0, 0), 1)
cv2.putText(watch_images[i], lt_acc[i], pos_lt_ac, fnt, 0.5, (0, 0, 0),
1)
cv2.putText(watch_images[i], iou[i], pos_iou, fnt, 0.5, (0, 0, 0), 1)
watch_images = torch.Tensor(watch_images.astype(np.float32))
watch_images = watch_images.permute(0, 3, 1, 2)
vis.image(torchvision.utils.make_grid(watch_images, nrow=3, padding=5),
opts=opts)
def vis_fn_detect(vis,
model,
faster_rcnn,
aff_theta,
train_loss,
val_pck,
train_lr,
epoch,
num_epochs,
dataloader,
use_cuda=True):
# Visualize watch images
affTnf = GeometricTnf(geometric_model='affine', use_cuda=use_cuda)
tpsTnf = GeometricTnf(geometric_model='tps', use_cuda=use_cuda)
watch_images = torch.Tensor(len(dataloader) * 5, 3, 240, 240)
# means for normalize of caffe resnet and vgg
# pixel_means = torch.Tensor(np.array([[[[102.9801, 115.9465, 122.7717]]]]).astype(np.float32))
for batch_idx, batch in enumerate(dataloader):
if use_cuda:
batch = batch_cuda(batch)
box_info_s = faster_rcnn(im_data=batch['source_im'],
im_info=batch['source_im_info'][:, 3:],
gt_boxes=batch['source_gt_boxes'],
num_boxes=batch['source_num_boxes'])[0:3]
box_info_t = faster_rcnn(im_data=batch['target_im'],
im_info=batch['target_im_info'][:, 3:],
gt_boxes=batch['target_gt_boxes'],
num_boxes=batch['target_num_boxes'])[0:3]
all_box_s = select_boxes(rois=box_info_s[0],
cls_prob=box_info_s[1],
bbox_pred=box_info_s[2],
im_infos=batch['source_im_info'][:, 3:])
all_box_t = select_boxes(rois=box_info_t[0],
cls_prob=box_info_t[1],
bbox_pred=box_info_t[2],
im_infos=batch['target_im_info'][:, 3:])
box_s, box_t = select_box_st(all_box_s, all_box_t)
theta_det = aff_theta(boxes_s=box_s, boxes_t=box_t)
theta_aff_tps, theta_aff = model(batch, theta_det)
warped_image_1 = affTnf(batch['source_image'], theta_det)
warped_image_2 = affTnf(warped_image_1, theta_aff)
warped_image_3 = tpsTnf(warped_image_2, theta_aff_tps)
watch_images[batch_idx * 5] = batch['source_image'][0]
watch_images[batch_idx * 5 + 1] = warped_image_1[0]
watch_images[batch_idx * 5 + 2] = warped_image_2[0]
watch_images[batch_idx * 5 + 3] = warped_image_3[0]
watch_images[batch_idx * 5 + 4] = batch['target_image'][0]
opts = dict(jpgquality=100,
title='Epoch ' + str(epoch) + ' source warped target')
# Un-normalize for caffe resnet and vgg
# watch_images = watch_images.permute(0, 2, 3, 1) + pixel_means
# watch_images = watch_images[:, :, :, [2, 1, 0]].permute(0, 3, 1, 2)
watch_images = normalize_image(watch_images, forward=False) * 255.0
vis.image(torchvision.utils.make_grid(watch_images, nrow=5, padding=5),
opts=opts)
# vis.images(watch_images, nrow=5, padding=3, opts=opts)
if epoch == num_epochs:
epochs = np.arange(1, num_epochs + 1)
# Visualize train loss
opts_loss = dict(xlabel='Epoch',
ylabel='Loss',
title='GM ResNet101 Detect&Affine&TPS Training Loss',
legend=['Loss'],
width=2000)
vis.line(train_loss, epochs, opts=opts_loss)
# Visualize val pck
opts_pck = dict(xlabel='Epoch',
ylabel='Val PCK',
title='GM ResNet101 Detect&Affine&TPS Val PCK',
legend=['PCK'],
width=2000)
vis.line(val_pck, epochs, opts=opts_pck)
# Visualize train lr
opts_lr = dict(
xlabel='Epoch',
ylabel='Learning Rate',
title='GM ResNet101 Detect&Affine&TPS Training Learning Rate',
legend=['LR'],
width=2000)
vis.line(train_lr, epochs, opts=opts_lr)
def vis_fn_dual(vis, train_loss, val_pck, train_lr, epoch, num_epochs, dataloader, theta, thetai, results, title, use_cuda=True):
# Visualize watch images
affTnf = GeometricTnf(geometric_model='affine', use_cuda=use_cuda)
tpsTnf = GeometricTnf(geometric_model='tps', use_cuda=use_cuda)
pt = PointTnf(use_cuda=use_cuda)
group_size = 4
watch_images = torch.ones(len(dataloader) * group_size, 3, 280, 240)
watch_keypoints = -torch.ones(len(dataloader) * group_size, 2, 20)
if use_cuda:
watch_images = watch_images.cuda()
watch_keypoints = watch_keypoints.cuda()
num_points = np.ones(len(dataloader) * group_size).astype(np.int8)
correct_index = list()
image_names = list()
metrics = list()
# Colors for keypoints
cmap = plt.get_cmap('tab20')
colors = list()
for c in range(20):
r = cmap(c)[0] * 255
g = cmap(c)[1] * 255
b = cmap(c)[2] * 255
colors.append((b, g, r))
fnt = cv2.FONT_HERSHEY_COMPLEX
theta, thetai = swap(theta, thetai)
# means for normalize of caffe resnet and vgg
# pixel_means = torch.Tensor(np.array([[[[102.9801, 115.9465, 122.7717]]]]).astype(np.float32))
for batch_idx, batch in enumerate(dataloader):
if use_cuda:
batch = batch_cuda(batch)
batch['source_image'], batch['target_image'] = swap(batch['source_image'], batch['target_image'])
batch['source_im_info'], batch['target_im_info'] = swap(batch['source_im_info'], batch['target_im_info'])
batch['source_points'], batch['target_points'] = swap(batch['source_points'], batch['target_points'])
# Theta and thetai
theta_aff = theta['aff'][batch_idx].unsqueeze(0)
theta_aff_tps = theta['afftps'][batch_idx].unsqueeze(0)
theta_aff_inver = thetai['aff'][batch_idx].unsqueeze(0)
theta_aff_tps_inver = thetai['afftps'][batch_idx].unsqueeze(0)
# Warped image
warped_aff = affTnf(batch['source_image'], theta_aff)
warped_aff_tps = tpsTnf(warped_aff, theta_aff_tps)
watch_images[batch_idx * group_size, :, 0:240, :] = batch['source_image']
watch_images[batch_idx * group_size + 1, :, 0:240, :] = warped_aff
watch_images[batch_idx * group_size + 2, :, 0:240, :] = warped_aff_tps
watch_images[batch_idx * group_size + 3, :, 0:240, :] = batch['target_image']
# Warped keypoints
source_im_size = batch['source_im_info'][:, 0:3]
target_im_size = batch['target_im_info'][:, 0:3]
source_points = batch['source_points']
target_points = batch['target_points']
source_points_norm = PointsToUnitCoords(P=source_points, im_size=source_im_size)
target_points_norm = PointsToUnitCoords(P=target_points, im_size=target_im_size)
warped_points_aff_norm = pt.affPointTnf(theta=theta_aff_inver, points=source_points_norm)
warped_points_aff = PointsToPixelCoords(P=warped_points_aff_norm, im_size=target_im_size)
_, index_aff, N_pts = pck(target_points, warped_points_aff)
warped_points_aff = relocate(warped_points_aff, target_im_size)
warped_points_aff_tps_norm = pt.tpsPointTnf(theta=theta_aff_tps_inver, points=source_points_norm)
warped_points_aff_tps_norm = pt.affPointTnf(theta=theta_aff_inver, points=warped_points_aff_tps_norm)
warped_points_aff_tps = PointsToPixelCoords(P=warped_points_aff_tps_norm, im_size=target_im_size)
_, index_aff_tps, _ = pck(target_points, warped_points_aff_tps)
warped_points_aff_tps = relocate(warped_points_aff_tps, target_im_size)
watch_keypoints[batch_idx * group_size, :, :N_pts] = relocate(batch['source_points'], source_im_size)[:, :, :N_pts]
watch_keypoints[batch_idx * group_size + 1, :, :N_pts] = warped_points_aff[:, :, :N_pts]
watch_keypoints[batch_idx * group_size + 2, :, :N_pts] = warped_points_aff_tps[:, :, :N_pts]
watch_keypoints[batch_idx * group_size + 3, :, :N_pts] = relocate(batch['target_points'], target_im_size)[:, :, :N_pts]
num_points[batch_idx * group_size:batch_idx * group_size + group_size] = N_pts
correct_index.append(np.arange(N_pts))
correct_index.append(index_aff)
correct_index.append(index_aff_tps)
correct_index.append(np.arange(N_pts))
image_names.append('Source')
image_names.append('Aff')
image_names.append('AffTPS')
image_names.append('Target')
metrics.append('')
metrics.append('PCK: {:.2%}'.format(float(results['aff']['pck'][batch_idx])))
metrics.append('PCK: {:.2%}'.format(float(results['afftps']['pck'][batch_idx])))
metrics.append('')
opts = dict(jpgquality=100, title='Epoch ' + str(epoch) + ' source warped target')
# Un-normalize for caffe resnet and vgg
# watch_images = watch_images.permute(0, 2, 3, 1) + pixel_means
# watch_images = watch_images[:, :, :, [2, 1, 0]].permute(0, 3, 1, 2)
# watch_images = normalize_image(watch_images, forward=False) * 255.0
watch_images[:, :, 0:240, :] = normalize_image(watch_images[:, :, 0:240, :], forward=False)
watch_images *= 255.0
watch_images = watch_images.permute(0, 2, 3, 1).cpu().numpy().astype(np.uint8)
watch_keypoints = watch_keypoints.cpu().numpy()
for i in range(watch_images.shape[0]):
pos_name = (80, 255)
if (i + 1) % group_size == 1 or (i + 1) % group_size == 0:
pos_pck = (0, 0)
else:
pos_pck = (70, 275)
cv2.putText(watch_images[i], image_names[i], pos_name, fnt, 0.5, (0, 0, 0), 1)
cv2.putText(watch_images[i], metrics[i], pos_pck, fnt, 0.5, (0, 0, 0), 1)
if (i + 1) % group_size == 0:
for j in range(num_points[i]):
cv2.drawMarker(watch_images[i], (watch_keypoints[i, 0, j], watch_keypoints[i, 1, j]), colors[j],
cv2.MARKER_CROSS, 12, 2, cv2.LINE_AA)
else:
for j in correct_index[i]:
cv2.drawMarker(watch_images[i], (watch_keypoints[i, 0, j], watch_keypoints[i, 1, j]), colors[j],
cv2.MARKER_DIAMOND, 12, 2, cv2.LINE_AA)
cv2.drawMarker(watch_images[i],
(watch_keypoints[i + (group_size - 1) - (i % group_size), 0, j], watch_keypoints[i + (group_size - 1) - (i % group_size), 1, j]),
colors[j], cv2.MARKER_CROSS, 12, 2, cv2.LINE_AA)
watch_images = torch.Tensor(watch_images.astype(np.float32))
watch_images = watch_images.permute(0, 3, 1, 2)
vis.image(torchvision.utils.make_grid(watch_images, nrow=4, padding=3), opts=opts)
if epoch == num_epochs:
epochs = np.arange(1, num_epochs+1)
# Visualize train loss
opts_loss = dict(xlabel='Epoch',
ylabel='Loss',
title='GM ResNet101 ' + title + ' Training Loss',
legend=['Loss'],
width=2000)
vis.line(train_loss, epochs, opts=opts_loss)
# Visualize val pck
opts_pck = dict(xlabel='Epoch',
ylabel='Val PCK',
title='GM ResNet101 ' + title + ' Val PCK',
legend=['PCK'],
width=2000)
vis.line(val_pck, epochs, opts=opts_pck)
# Visualize train lr
opts_lr = dict(xlabel='Epoch',
ylabel='Learning Rate',
title='GM ResNet101 ' + title + ' Training Learning Rate',
legend=['LR'],
width=2000)
vis.line(train_lr, epochs, opts=opts_lr)
def compute_metric(metric, model, fasterRCNN, dataset, dataloader, args=None):
# Initialize stats
N = len(dataset)
stats = {}
# decide which results should be computed aff/tps/aff+tps
stats['aff'] = {}
stats['tps'] = {}
stats['aff_tps'] = {}
# if two_stage or do_aff:
# stats['aff']={}
# if not two_stage and do_tps:
# stats['tps']={}
# if two_stage:
# stats['aff_tps']={}
# choose metric function and metrics to compute
if metric == 'pck':
metrics = ['pck']
metric_fun = pck_metric
if metric == 'dist':
metrics = ['dist']
# metric_fun = point_dist_metric
elif metric == 'area':
metrics = [
'intersection_over_union', 'label_transfer_accuracy',
'localization_error'
]
# metric_fun = area_metrics
elif metric == 'pascal_parts':
metrics = ['intersection_over_union', 'pck']
# metric_fun = pascal_parts_metrics
elif metric == 'flow':
metrics = ['flow']
# metric_fun = flow_metrics
elif metric == 'inlier_count':
metrics = ['inlier_count']
# metric_fun = inlier_count
# model.return_correlation = True
# initialize vector for storing results for each metric
for key in stats.keys():
for metric in metrics:
stats[key][metric] = np.zeros((N, 1))
affine_theta = AffineTheta(use_cuda=True, original=False, image_size=240)
affTnf = GeometricTnf(geometric_model='affine', use_cuda=True)
# Compute
for i, batch in enumerate(dataloader):
# batch = batch_tnf(batch)
batch_start_idx = args.batch_size * i
batch_end_idx = np.minimum(batch_start_idx + args.batch_size, N)
for k, v in batch.items():
batch[k] = batch[k].cuda()
model.eval()
fasterRCNN.eval()
theta_aff = None
theta_tps = None
theta_aff_tps = None
tnf_batch = {
'source_image': batch['source_image'],
'target_image': batch['target_image']
}
theta_tps = model(tnf_batch)
thresh = 0.05
max_per_image = 50
rois_s, cls_prob_s, bbox_pred_s, _, _, _, _, _ = fasterRCNN(
batch['source_im'], batch['source_im_info'],
batch['source_gt_boxes'], batch['source_num_boxes'])
all_boxes_s = select_boxes(rois_s, cls_prob_s, bbox_pred_s,
batch['source_im_info'], thresh,
max_per_image)
rois_t, cls_prob_t, bbox_pred_t, _, _, _, _, _ = fasterRCNN(
batch['target_im'], batch['target_im_info'],
batch['target_gt_boxes'], batch['target_num_boxes'])
all_boxes_t = select_boxes(rois_t, cls_prob_t, bbox_pred_t,
batch['target_im_info'], thresh,
max_per_image)
boxes_s, boxes_t = select_box(all_boxes_s, all_boxes_t)
boxes_s = boxes_s.cuda()
boxes_t = boxes_t.cuda()
theta_aff = affine_theta(boxes_s=boxes_s,
boxes_t=boxes_t,
source_im_size=None,
target_im_size=None)
batch['source_image'] = affTnf(batch['source_image'], theta_aff)
tnf_batch = {
'source_image': batch['source_image'],
'target_image': batch['target_image']
}
theta_aff_tps = model(tnf_batch)
stats = metric_fun(batch, batch_start_idx, theta_aff, theta_tps,
theta_aff_tps, stats, args)
# if two_stage:
# if model.return_correlation==False:
# theta_aff,theta_aff_tps=model(batch)
# else:
# theta_aff,theta_aff_tps,corr_aff,corr_aff_tps=model(batch)
# elif do_aff:
# theta_aff=model(batch)
# if isinstance(theta_aff,tuple):
# theta_aff=theta_aff[0]
# elif do_tps:
# theta_tps=model(batch)
# if isinstance(theta_tps,tuple):
# theta_tps=theta_tps[0]
#
# if metric=='inlier_count':
# stats = inlier_count(batch,batch_start_idx,theta_aff,theta_tps,theta_aff_tps,corr_aff,corr_aff_tps,stats,args)
# elif metric_fun is not None:
# stats = metric_fun(batch,batch_start_idx,theta_aff,theta_tps,theta_aff_tps,stats,args)
print('Batch: [{}/{} ({:.0f}%)]'.format(i, len(dataloader),
100. * i / len(dataloader)))
# Print results
if metric == 'flow':
print('Flow files have been saved to ' + args.flow_output_dir)
return stats
for key in stats.keys():
print('=== Results ' + key + ' ===')
for metric in metrics:
# print per-class brakedown for PFPascal, or caltech
if isinstance(dataset, PFPASCALDataset):
N_cat = int(np.max(dataset.category)
) # Number of categories in dataset (PF-PASCAL)
for c in range(N_cat):
cat_idx = np.nonzero(
dataset.category == c +
1)[0] # Compute indices of current category
print(
dataset.category_names[c].ljust(15) + ': ',
'{:.2%}'.format(np.mean(stats[key][metric][cat_idx])))
# print mean value
results = stats[key][metric]
good_idx = np.flatnonzero((results != -1) * ~np.isnan(results))
print('Total: ' + str(results.size))
print('Valid: ' + str(good_idx.size))
filtered_results = results[good_idx]
print(metric + ':', '{:.2%}'.format(np.mean(filtered_results)))
print('\n')
return stats
def vis_control(vis,
dataloader,
theta_1,
theta_2,
dataset_name,
use_cuda=True):
# Visualize watch images
tpsTnf_1 = GeometricTnf(geometric_model='tps', use_cuda=use_cuda)
tpsTnf_2 = GeometricTnf2(geometric_model='tps', use_cuda=use_cuda)
group_size = 5
watch_images = torch.ones(len(dataloader) * group_size, 3, 340, 340)
if use_cuda:
watch_images = watch_images.cuda()
# Colors for keypoints
cmap = plt.get_cmap('tab20')
colors = list()
for c in range(20):
r = cmap(c)[0] * 255
g = cmap(c)[1] * 255
b = cmap(c)[2] * 255
colors.append((b, g, r))
fnt = cv2.FONT_HERSHEY_COMPLEX
# means for normalize of caffe resnet and vgg
# pixel_means = torch.Tensor(np.array([[[[102.9801, 115.9465, 122.7717]]]]).astype(np.float32))
for batch_idx, batch in enumerate(dataloader):
if use_cuda:
batch = batch_cuda(batch)
# Theta and theta_inver
theta_tps_1 = theta_1['tps'][batch_idx].unsqueeze(0)
theta_tps_2 = theta_2['tps'][batch_idx].unsqueeze(0)
# Warped image
warped_tps_1 = tpsTnf_1(batch['source_image'], theta_tps_1)
warped_tps_2 = tpsTnf_2(batch['source_image'], theta_tps_2)
watch_images[batch_idx * group_size, :, 50:290,
50:290] = batch['source_image']
watch_images[batch_idx * group_size + 1, :, 50:290,
50:290] = warped_tps_1
watch_images[batch_idx * group_size + 2, :, 50:290,
50:290] = batch['source_image']
watch_images[batch_idx * group_size + 3, :, 50:290,
50:290] = warped_tps_2
watch_images[batch_idx * group_size + 4, :, 50:290,
50:290] = batch['target_image']
opts = dict(jpgquality=100, title=dataset_name)
watch_images[:, :, 50:290,
50:290] = normalize_image(watch_images[:, :, 50:290, 50:290],
forward=False)
watch_images *= 255.0
watch_images = watch_images.permute(0, 2, 3,
1).cpu().numpy().astype(np.uint8)
im_size = torch.Tensor([[240, 240]]).cuda()
for i in range(watch_images.shape[0]):
if i % group_size == 0:
cp_norm = theta_1['tps'][int(i / group_size)].view(1, 2, -1)
cp = PointsToPixelCoords(P=cp_norm, im_size=im_size)
cp = cp.squeeze().cpu().numpy() + 50
for j in range(9):
cv2.drawMarker(watch_images[i], (cp[0, j], cp[1, j]),
(0, 0, 255), cv2.MARKER_TILTED_CROSS, 12, 2,
cv2.LINE_AA)
for j in range(2):
for k in range(3):
# vertical grid
cv2.line(watch_images[i],
(cp[0, j + k * 3], cp[1, j + k * 3]),
(cp[0, j + k * 3 + 1], cp[1, j + k * 3 + 1]),
(0, 0, 255), 2, cv2.LINE_AA)
# horizontal grid
cv2.line(watch_images[i],
(cp[0, j * 3 + k], cp[1, j * 3 + k]),
(cp[0, j * 3 + k + 3], cp[1, j * 3 + k + 3]),
(0, 0, 255), 2, cv2.LINE_AA)
if i % group_size == 1:
cp_norm = torch.Tensor(
[-1, -1, -1, 0, 0, 0, 1, 1, 1, -1, 0, 1, -1, 0, 1, -1, 0,
1]).cuda().view(1, 2, -1)
cp = PointsToPixelCoords(P=cp_norm, im_size=im_size)
cp = cp.squeeze().cpu().numpy() + 50
for j in range(9):
cv2.drawMarker(watch_images[i], (cp[0, j], cp[1, j]),
(0, 0, 255), cv2.MARKER_TILTED_CROSS, 12, 2,
cv2.LINE_AA)
for j in range(1):
for k in range(3):
# vertical grid
cv2.line(watch_images[i],
(cp[0, j + k * 3], cp[1, j + k * 3]),
(cp[0, j + k * 3 + 1], cp[1, j + k * 3 + 1]),
(0, 0, 255), 2, cv2.LINE_AA)
# horizontal grid
cv2.line(watch_images[i],
(cp[0, j * 3 + k], cp[1, j * 3 + k]),
(cp[0, j * 3 + k + 3], cp[1, j * 3 + k + 3]),
(0, 0, 255), 2, cv2.LINE_AA)
if i % group_size == 2:
cp_norm = theta_2['tps'][int(i / group_size)][:18].view(1, 2, -1)
cp = PointsToPixelCoords(P=cp_norm, im_size=im_size)
cp = cp.squeeze().cpu().numpy() + 50
for j in range(9):
cv2.drawMarker(watch_images[i], (cp[0, j], cp[1, j]),
(0, 0, 255), cv2.MARKER_TILTED_CROSS, 12, 2,
cv2.LINE_AA)
for j in range(2):
for k in range(3):
# vertical grid
cv2.line(watch_images[i],
(cp[0, j + k * 3], cp[1, j + k * 3]),
(cp[0, j + k * 3 + 1], cp[1, j + k * 3 + 1]),
(0, 0, 255), 2, cv2.LINE_AA)
# horizontal grid
cv2.line(watch_images[i],
(cp[0, j * 3 + k], cp[1, j * 3 + k]),
(cp[0, j * 3 + k + 3], cp[1, j * 3 + k + 3]),
(0, 0, 255), 2, cv2.LINE_AA)
if i % group_size == 3:
cp_norm = theta_2['tps'][int(i / group_size)][18:].view(1, 2, -1)
cp = PointsToPixelCoords(P=cp_norm, im_size=im_size)
cp = cp.squeeze().cpu().numpy() + 50
for j in range(9):
cv2.drawMarker(watch_images[i], (cp[0, j], cp[1, j]),
(0, 0, 255), cv2.MARKER_TILTED_CROSS, 12, 2,
cv2.LINE_AA)
for j in range(2):
for k in range(3):
# vertical grid
cv2.line(watch_images[i],
(cp[0, j + k * 3], cp[1, j + k * 3]),
(cp[0, j + k * 3 + 1], cp[1, j + k * 3 + 1]),
(0, 0, 255), 2, cv2.LINE_AA)
# horizontal grid
cv2.line(watch_images[i],
(cp[0, j * 3 + k], cp[1, j * 3 + k]),
(cp[0, j * 3 + k + 3], cp[1, j * 3 + k + 3]),
(0, 0, 255), 2, cv2.LINE_AA)
watch_images = torch.Tensor(watch_images.astype(np.float32))
watch_images = watch_images.permute(0, 3, 1, 2)
vis.image(torchvision.utils.make_grid(watch_images, nrow=5, padding=5),
opts=opts)
def vis_pf(vis,
dataloader,
theta_1,
theta_2,
theta_inver_1,
theta_inver_2,
results_1,
results_2,
dataset_name,
use_cuda=True):
# Visualize watch images
tpsTnf_1 = GeometricTnf(geometric_model='tps', use_cuda=use_cuda)
tpsTnf_2 = GeometricTnf2(geometric_model='tps', use_cuda=use_cuda)
pt_1 = PointTnf(use_cuda=use_cuda)
pt_2 = PointTPS(use_cuda=use_cuda)
group_size = 4
watch_images = torch.ones(len(dataloader) * group_size, 3, 280, 240)
watch_keypoints = -torch.ones(len(dataloader) * group_size, 2, 20)
if use_cuda:
watch_images = watch_images.cuda()
watch_keypoints = watch_keypoints.cuda()
num_points = np.ones(len(dataloader) * 6).astype(np.int8)
correct_index = list()
image_names = list()
metrics = list()
# Colors for keypoints
cmap = plt.get_cmap('tab20')
colors = list()
for c in range(20):
r = cmap(c)[0] * 255
g = cmap(c)[1] * 255
b = cmap(c)[2] * 255
colors.append((b, g, r))
fnt = cv2.FONT_HERSHEY_COMPLEX
# means for normalize of caffe resnet and vgg
# pixel_means = torch.Tensor(np.array([[[[102.9801, 115.9465, 122.7717]]]]).astype(np.float32))
for batch_idx, batch in enumerate(dataloader):
if use_cuda:
batch = batch_cuda(batch)
# Theta and theta_inver
theta_tps_1 = theta_1['tps'][batch_idx].unsqueeze(0)
theta_tps_2 = theta_2['tps'][batch_idx].unsqueeze(0)
thetai_tps_1 = theta_inver_1['tps'][batch_idx].unsqueeze(0)
thetai_tps_2 = theta_inver_2['tps'][batch_idx].unsqueeze(0)
# Warped image
warped_tps_1 = tpsTnf_1(batch['source_image'], theta_tps_1)
warped_tps_2 = tpsTnf_2(batch['source_image'], theta_tps_2)
watch_images[batch_idx * group_size, :,
0:240, :] = batch['source_image']
watch_images[batch_idx * group_size + 1, :, 0:240, :] = warped_tps_1
watch_images[batch_idx * group_size + 2, :, 0:240, :] = warped_tps_2
watch_images[batch_idx * group_size + 3, :,
0:240, :] = batch['target_image']
# Warped keypoints
source_im_size = batch['source_im_info'][:, 0:3]
target_im_size = batch['target_im_info'][:, 0:3]
source_points = batch['source_points']
target_points = batch['target_points']
source_points_norm = PointsToUnitCoords(P=source_points,
im_size=source_im_size)
target_points_norm = PointsToUnitCoords(P=target_points,
im_size=target_im_size)
warped_points_tps_norm_1 = pt_1.tpsPointTnf(theta=thetai_tps_1,
points=source_points_norm)
warped_points_tps_1 = PointsToPixelCoords(P=warped_points_tps_norm_1,
im_size=target_im_size)
pck_tps_1, index_tps_1, N_pts = pck(target_points, warped_points_tps_1,
dataset_name)
warped_points_tps_1 = relocate(warped_points_tps_1, target_im_size)
warped_points_tps_norm_2 = pt_2.tpsPointTnf(theta=thetai_tps_2,
points=source_points_norm)
warped_points_tps_2 = PointsToPixelCoords(P=warped_points_tps_norm_2,
im_size=target_im_size)
pck_tps_2, index_tps_2, _ = pck(target_points, warped_points_tps_2,
dataset_name)
warped_points_tps_2 = relocate(warped_points_tps_2, target_im_size)
watch_keypoints[batch_idx * group_size, :, :N_pts] = relocate(
batch['source_points'], source_im_size)[:, :, :N_pts]
watch_keypoints[batch_idx * group_size +
1, :, :N_pts] = warped_points_tps_1[:, :, :N_pts]
watch_keypoints[batch_idx * group_size +
2, :, :N_pts] = warped_points_tps_2[:, :, :N_pts]
watch_keypoints[batch_idx * group_size + 3, :, :N_pts] = relocate(
batch['target_points'], target_im_size)[:, :, :N_pts]
num_points[batch_idx * group_size:batch_idx * group_size +
group_size] = N_pts
correct_index.append(np.arange(N_pts))
correct_index.append(index_tps_1)
correct_index.append(index_tps_2)
correct_index.append(np.arange(N_pts))
image_names.append('Source')
image_names.append('TPS')
image_names.append('TPS_Jitter')
image_names.append('Target')
metrics.append('')
metrics.append('PCK: {:.2%}'.format(pck_tps_1))
metrics.append('PCK: {:.2%}'.format(pck_tps_2))
metrics.append('')
opts = dict(jpgquality=100, title=dataset_name)
# Un-normalize for caffe resnet and vgg
# watch_images = watch_images.permute(0, 2, 3, 1) + pixel_means
# watch_images = watch_images[:, :, :, [2, 1, 0]].permute(0, 3, 1, 2)
# watch_images = normalize_image(watch_images, forward=False) * 255.0
watch_images[:, :, 0:240, :] = normalize_image(watch_images[:, :,
0:240, :],
forward=False)
watch_images *= 255.0
watch_images = watch_images.permute(0, 2, 3,
1).cpu().numpy().astype(np.uint8)
watch_keypoints = watch_keypoints.cpu().numpy()
for i in range(watch_images.shape[0]):
pos_name = (80, 255)
if (i + 1) % group_size == 1 or (i + 1) % group_size == 0:
pos_pck = (0, 0)
else:
pos_pck = (70, 275)
cv2.putText(watch_images[i], image_names[i], pos_name, fnt, 0.5,
(0, 0, 0), 1)
cv2.putText(watch_images[i], metrics[i], pos_pck, fnt, 0.5, (0, 0, 0),
1)
if (i + 1) % group_size == 0:
for j in range(num_points[i]):
cv2.drawMarker(
watch_images[i],
(watch_keypoints[i, 0, j], watch_keypoints[i, 1, j]),
colors[j], cv2.MARKER_DIAMOND, 12, 2, cv2.LINE_AA)
else:
for j in correct_index[i]:
cv2.drawMarker(
watch_images[i],
(watch_keypoints[i, 0, j], watch_keypoints[i, 1, j]),
colors[j], cv2.MARKER_CROSS, 12, 2, cv2.LINE_AA)
cv2.drawMarker(watch_images[i],
(watch_keypoints[i + (group_size - 1) -
(i % group_size), 0, j],
watch_keypoints[i + (group_size - 1) -
(i % group_size), 1, j]),
colors[j], cv2.MARKER_DIAMOND, 12, 2,
cv2.LINE_AA)
watch_images = torch.Tensor(watch_images.astype(np.float32))
watch_images = watch_images.permute(0, 3, 1, 2)
vis.image(torchvision.utils.make_grid(watch_images, nrow=4, padding=5),
opts=opts)
def __init__(self,
aff_output_dim=6,
tps_output_dim=18,
use_cuda=True,
feature_extraction_cnn='resnet101',
feature_extraction_last_layer='',
fr_feature_size=15,
fr_kernel_sizes=[7, 5],
fr_channels=[128, 64],
fixed_blocks=3,
normalize_features=True,
normalize_matches=True,
batch_normalization=True,
crop_layer=None,
pytorch=False,
caffe=False,
return_correlation=False):
super(DualGeometricMatching, self).__init__()
self.normalize_features = normalize_features
self.normalize_matches = normalize_matches
self.return_correlation = return_correlation
# self.return_correlation = True
self.AffTnf = GeometricTnf(geometric_model='affine', use_cuda=use_cuda)
self.crop_layer = crop_layer
# Feature extraction networks for two images
self.FeatureExtraction = FeatureExtraction(
feature_extraction_cnn=feature_extraction_cnn,
last_layer=feature_extraction_last_layer,
normalization=normalize_features,
fixed_blocks=fixed_blocks,
pytorch=pytorch,
caffe=caffe)
# Matching layer based on normalized feature maps of two images
self.FeatureCorrelation = FeatureCorrelation(
shape='3D', normalization=normalize_matches)
for param in self.parameters():
param.requires_grad = False
# Regression layer based on correlated feature map for predicting parameters of geometric transformation
self.ThetaRegression = ThetaRegression(
output_dim=aff_output_dim,
feature_size=fr_feature_size,
kernel_sizes=fr_kernel_sizes,
channels=fr_channels,
batch_normalization=batch_normalization)
# Feature cropping on specific layer
# self.FeatureCrop = FeatureCrop(crop_layer=crop_layer)
# if self.crop_layer == 'conv1':
# self.FeatureExtraction_2 = FeatureExtraction_2(feature_extraction_cnn=feature_extraction_cnn,
# first_layer='pool1', last_layer='pool4',
# pretrained=pretrained)
self.ThetaRegression2 = ThetaRegression(
output_dim=tps_output_dim,
feature_size=fr_feature_size,
kernel_sizes=fr_kernel_sizes,
channels=fr_channels,
batch_normalization=batch_normalization)
def train_fn(epoch,
model,
loss_fn,
loss_cycle_fn,
lambda_c,
optimizer,
dataloader,
triple_generation,
geometric_model='tps',
use_cuda=True,
log_interval=100,
vis=None):
"""
Train the model with synthetically training triple:
{source image, target image, refer image (warped source image), theta_GT} from PF-PASCAL.
1. Train the transformation parameters theta_st from source image to target image;
2. Train the transformation parameters theta_tr from target image to refer image;
3. Combine theta_st and theta_st to obtain theta from source image to refer image, and compute loss between
theta and theta_GT.
"""
geoTnf = GeometricTnf(geometric_model=geometric_model, use_cuda=use_cuda)
epoch_loss = 0
if (epoch % 5 == 0 or epoch == 1) and vis is not None:
stride_images = len(dataloader) / 3
group_size = 6
watch_images = torch.ones(group_size * 4, 3, 260, 240).cuda()
image_names = list()
fnt = cv2.FONT_HERSHEY_COMPLEX
stride_loss = len(dataloader) / 105
iter_loss = np.zeros(106)
begin = time.time()
for batch_idx, batch in enumerate(dataloader):
''' Move input batch to gpu '''
# batch['source_image'].shape & batch['target_image'].shape: (batch_size, 3, 240, 240)
# batch['theta'].shape-tps: (batch_size, 18)-random or (batch_size, 18, 1, 1)-(pre-set from csv)
if use_cuda:
batch = batch_cuda(batch)
''' Get the training triple {source image, target image, refer image (warped source image), theta_GT}'''
batch_triple = triple_generation(batch)
''' Train the model '''
optimizer.zero_grad()
loss = 0
# Predict tps parameters between images
# theta.shape: (batch_size, 18) for tps, theta.shape: (batch_size, 6) for affine
theta_st, theta_ts, theta_tr, theta_rt = model(batch_triple)
loss_match = loss_fn(theta_st=theta_st,
theta_tr=theta_tr,
theta_GT=batch_triple['theta_GT'])
loss_cycle_st = loss_cycle_fn(theta_AB=theta_st, theta_BA=theta_ts)
loss_cycle_ts = loss_cycle_fn(theta_AB=theta_ts, theta_BA=theta_st)
loss_cycle_tr = loss_cycle_fn(theta_AB=theta_tr, theta_BA=theta_rt)
loss_cycle_rt = loss_cycle_fn(theta_AB=theta_rt, theta_BA=theta_tr)
loss = loss_match + lambda_c * (loss_cycle_st + loss_cycle_ts +
loss_cycle_tr + loss_cycle_rt) / 4
loss.backward()
optimizer.step()
epoch_loss += loss.item()
if (batch_idx + 1) % log_interval == 0:
end = time.time()
print(
'Train epoch: {} [{}/{} ({:.0%})]\t\tCurrent batch loss: {:.6f}\t\tTime cost ({} batches): {:.4f} s'
.format(epoch, batch_idx + 1,
len(dataloader), (batch_idx + 1) / len(dataloader),
loss.item(), batch_idx + 1, end - begin))
if (epoch % 5 == 0 or epoch == 1) and vis is not None:
if (batch_idx + 1) % stride_images == 0 or batch_idx == 0:
watch_images, image_names = add_watch(
watch_images, image_names, batch_triple, geoTnf, theta_st,
theta_tr,
int((batch_idx + 1) / stride_images) * group_size)
# if batch_idx <= 3:
# watch_images, image_names = add_watch(watch_images, image_names, batch_triple, geoTnf, theta_st,
# theta_tr, batch_idx * group_size)
# if batch_idx == 3:
# opts = dict(jpgquality=100, title='Epoch ' + str(epoch) + ' source warped_sr target warped_tr refer warped_sr')
# watch_images[:, :, 0:240, :] = normalize_image(watch_images[:, :, 0:240, :], forward=False)
# watch_images *= 255.0
# watch_images = watch_images.permute(0, 2, 3, 1).cpu().numpy().astype(np.uint8)
# for i in range(watch_images.shape[0]):
# cv2.putText(watch_images[i], image_names[i], (80, 255), fnt, 0.5, (0, 0, 0), 1)
# watch_images = torch.Tensor(watch_images.astype(np.float32))
# watch_images = watch_images.permute(0, 3, 1, 2)
# vis.image(torchvision.utils.make_grid(watch_images, nrow=group_size, padding=3), opts=opts)
if (batch_idx + 1) % stride_loss == 0 or batch_idx == 0:
iter_loss[int((batch_idx + 1) /
stride_loss)] = epoch_loss / (batch_idx + 1)
end = time.time()
# Visualize watch images & train loss
if (epoch % 5 == 0 or epoch == 1) and vis is not None:
opts = dict(jpgquality=100,
title='Epoch ' + str(epoch) +
' source warped_sr target warped_tr refer warped_sr')
watch_images[:, :, 0:240, :] = normalize_image(watch_images[:, :,
0:240, :],
forward=False)
watch_images *= 255.0
watch_images = watch_images.permute(0, 2, 3,
1).cpu().numpy().astype(np.uint8)
for i in range(watch_images.shape[0]):
cv2.putText(watch_images[i], image_names[i], (80, 255), fnt, 0.5,
(0, 0, 0), 1)
watch_images = torch.Tensor(watch_images.astype(np.float32))
watch_images = watch_images.permute(0, 3, 1, 2)
vis.image(torchvision.utils.make_grid(watch_images,
nrow=group_size,
padding=3),
opts=opts)
opts_loss = dict(xlabel='Iterations (' + str(stride_loss) + ')',
ylabel='Loss',
title='GM ResNet101 ' + geometric_model +
' Training Loss in Epoch ' + str(epoch),
legend=['Loss'],
width=2000)
vis.line(iter_loss, np.arange(106), opts=opts_loss)
epoch_loss /= len(dataloader)
print('Train set -- Average loss: {:.6f}\t\tTime cost: {:.4f}'.format(
epoch_loss, end - begin))
return epoch_loss, end - begin