def val(model, criterion, loader, meta, amp):
model.eval()
accuracy = []
mean_loss = 0
for batch_idx, data in enumerate(loader):
x = data['image']
y = data['classes']
x = x.to('cuda')
if amp != '01':
x = x.half()
y = y.to('cuda')
y_hat = model(x)
probs = torch.sigmoid(y_hat)
loss = criterion(y_hat, y)
prediction = probs.flatten() > 0.5
gt = y.flatten() > 0.5
accuracy.extend(t2np(prediction) == t2np(gt))
mean_loss += t2np(loss).mean()
accuracy = np.mean(accuracy)
mean_loss = np.mean(mean_loss / len(loader))
meta['val']['loss'].append(mean_loss)
meta['val']['accuracy'].append(accuracy)
meta['val']['iteration'].append(meta['iteration'][-1])
print('\nval loss {:^.3f} | val accuracy {:^.3f}'.format(
meta['val']['loss'][-1], meta['val']['accuracy'][-1]))
def val(epoch):
net.eval()
val_loss = 0
for batch_idx, (inputs, loc_targets, cls_targets) in enumerate(valloader):
inputs = inputs.to('cuda')
loc_targets = loc_targets.to('cuda')
cls_targets = cls_targets.to('cuda')
loc_preds, cls_preds = net(inputs)
# pos = (cls_targets > 0)
loss = criterion(loc_preds, loc_targets, cls_preds, cls_targets,
batch_idx % 1 == 0)
loss, inputs, loc_targets, cls_targets = t2np(loss), t2np(
inputs), t2np(loc_targets), t2np(cls_targets)
val_loss += loss.mean()
if batch_idx % 1 == 0:
print('val_loss: %.4f | avg_loss: %.4f' % (loss.mean(), val_loss /
(batch_idx + 1)))
import sys
sys.stdout.flush()
print('val_loss: %.4f | avg_loss: %.4f' % (loss.mean(), val_loss /
(batch_idx + 1)))
save_checkpoint(val_loss, epoch, len(valloader))
def visualize2DGrad():
plt.figure(1)
plt.subplot(121)
plt.imshow(utils.t2np(grad[0, :, :, 0])) #, clim=(-1, 1))
plt.title('gradX')
plt.colorbar()
plt.subplot(122)
plt.imshow(utils.t2np(grad[0, :, :, 1])) #, clim=(-1, 1))
plt.title('gradY')
plt.colorbar()
plt.show()
def visualize1D():
plt.figure(1)
plt.subplot(121)
plt.plot(I0[0, :, 0])
plt.subplot(122)
plt.plot(utils.t2np(out[0, :, 0]))
plt.show()
def train(epoch):
from time import time
print('\nTrain Epoch: %d' % epoch)
net.train()
train_loss = 0
tik = time()
tok = time()
for batch_idx, (inputs, loc_targets,
cls_targets) in enumerate(trainloader):
# if True:
# print(cls_targets.unique())
# print(inputs.shape)
# print(inputs)
inputs = inputs.to('cuda')
loc_targets = loc_targets.to('cuda')
cls_targets = cls_targets.to('cuda')
# print(cls_targets[:10])
optimizer.zero_grad()
loc_preds, cls_preds = net(inputs)
# print(cls_preds)
# pos = (cls_targets > 0).detach()
# num_pos = t2np(pos).astype(np.int).sum()
# if num_pos==0:
# print('zero num positive')
# continue
# print('loc_preds', loc_preds.shape)
# print('cls_preds', cls_preds.shape)
# print('loc_targets', loc_targets.shape)
# print('cls_targets', cls_targets.shape)
# print('loc_preds', loc_preds.shape)
loss = criterion(loc_preds, loc_targets, cls_preds, cls_targets,
batch_idx % 20 == 0)
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
# nn.utils.clip_grad_norm(net.parameters(), max_norm=1.0)
optimizer.step()
lr_s.step()
train_loss += t2np(loss).mean()
if batch_idx % 20 == 0:
tok = time()
print(
'batch: %d/%d | time: %.2f min | train_loss: %.3f | avg_loss: %.4f | lr: %.6f '
% (batch_idx, len(trainloader),
(tok - tik) / 60, loss.detach().to('cpu').numpy().mean(),
train_loss / (batch_idx + 1), lr_s.get_lr()[0]))
import sys
sys.stdout.flush()
tik = time()
def visualize2D():
plt.figure(1)
plt.subplot(121)
plt.imshow(I0[0, :, :, 0], clim=(-1, 1))
plt.colorbar()
plt.subplot(122)
plt.imshow(utils.t2np(out[0, :, :, 0]), clim=(-1, 1))
plt.colorbar()
plt.show()
def visualize3DGrad():
plt.figure(1)
plt.subplot(331)
plt.imshow(utils.t2np(grad[0, coorCenter, :, :, 0])) #, clim=(-1, 1))
plt.title('gradX - xslice')
plt.colorbar()
plt.subplot(332)
plt.imshow(utils.t2np(grad[0, :, coorCenter, :, 0])) #, clim=(-1, 1))
plt.title('gradX - yslice')
plt.colorbar()
plt.subplot(333)
plt.imshow(utils.t2np(grad[0, :, :, coorCenter, 0])) #, clim=(-1, 1))
plt.title('gradX - zslice')
plt.colorbar()
plt.subplot(334)
plt.imshow(utils.t2np(grad[0, coorCenter, :, :, 1])) # , clim=(-1, 1))
plt.title('gradY - xslice')
plt.colorbar()
plt.subplot(335)
plt.imshow(utils.t2np(grad[0, :, coorCenter, :, 1])) # , clim=(-1, 1))
plt.title('gradY - yslice')
plt.colorbar()
plt.subplot(336)
plt.imshow(utils.t2np(grad[0, :, :, coorCenter, 1])) # , clim=(-1, 1))
plt.title('gradY - zslice')
plt.colorbar()
plt.subplot(337)
plt.imshow(utils.t2np(grad[0, coorCenter, :, :, 2])) # , clim=(-1, 1))
plt.title('gradZ - xslice')
plt.colorbar()
plt.subplot(338)
plt.imshow(utils.t2np(grad[0, :, coorCenter, :, 2])) # , clim=(-1, 1))
plt.title('gradZ - yslice')
plt.colorbar()
plt.subplot(339)
plt.imshow(utils.t2np(grad[0, :, :, coorCenter, 2])) # , clim=(-1, 1))
plt.title('gradZ - zslice')
plt.colorbar()
plt.show()
def visualize3D():
plt.figure(1)
plt.subplot(231)
plt.imshow(I0[0, coorCenter, :, :, 0], clim=(-1, 1))
plt.colorbar()
plt.subplot(232)
plt.imshow(I0[0, :, coorCenter, :, 0], clim=(-1, 1))
plt.colorbar()
plt.subplot(233)
plt.imshow(I0[0, :, :, coorCenter, 0], clim=(-1, 1))
plt.colorbar()
plt.subplot(234)
plt.imshow(utils.t2np(out[0, coorCenter, :, :, 0]), clim=(-1, 1))
plt.colorbar()
plt.subplot(235)
plt.imshow(utils.t2np(out[0, :, coorCenter, :, 0]), clim=(-1, 1))
plt.colorbar()
plt.subplot(236)
plt.imshow(utils.t2np(out[0, :, :, coorCenter, 0]), clim=(-1, 1))
plt.colorbar()
plt.show()
def train_epoch(model, criterion, optim, lr_scheduler, loader, meta):
model.train()
tik = time()
for batch_idx, data in enumerate(loader):
x = data['image']
y = data['classes']
x, y = x.to('cuda'), y.to('cuda')
optim.zero_grad()
y_hat = model(x)
probs = torch.functional.F.sigmoid(y_hat)
loss = criterion(y_hat, y)
with amp.scale_loss(loss, optim) as scaled_loss:
scaled_loss.backward()
optim.step()
lr_scheduler.step()
prediction = probs.flatten() > 0.5
gt = y.flatten() > 0.5
accuracy = (t2np(prediction) == t2np(gt)).mean()
meta['iteration'].append(meta['iteration'][-1] + 1)
meta['lr'].append(lr_scheduler.get_lr()[0])
meta['train']['loss'].append(t2np(loss).mean())
meta['train']['accuracy'].append(accuracy)
if batch_idx % 100 == 0:
tok = time()
print(
'batch {:#5}/{:^5} | loss {:^.3f} | accuracy {:^.3f} | lr {:^.6f} | time {:^.2f}s'
.format(batch_idx, len(loader), meta['train']['loss'][-1],
meta['train']['accuracy'][-1], meta['lr'][-1],
(tok - tik)))
sys.stdout.flush()
tik = time()
def forward(self, loc_preds, loc_targets, cls_preds, cls_targets ,verbose):
batch_size, n_anchors = cls_targets.size()
fg_idx = cls_targets > 0
n_fg = fg_idx.float().sum()
fg_mask = fg_idx.unsqueeze(2).expand_as(loc_preds)
fg_loc_preds = loc_preds[fg_mask]
fg_loc_targets = loc_targets[fg_mask]
loc_loss = 0.4 * F.smooth_l1_loss(fg_loc_preds, fg_loc_targets, reduction='none').sum()
fbg_idx = cls_targets != -1
# print('cls_targets shape', cls_targets.shape)
fbg_mask = fbg_idx.expand_as(cls_targets)
# print('fgb_mask shape', fbg_mask.shape)
fgb_cls_preds = cls_preds[fbg_mask].squeeze()
fgb_cls_targets = cls_targets[fbg_mask]
# print('fgb_cls_preds', fgb_cls_preds.shape)
# print('fgb_cls_targets', fgb_cls_targets.shape)
ohe_targets = self._ohe(fgb_cls_targets)
# print(ohe_targets[:10])
# print(ohe_targets.shape)
# print(fgb_cls_preds.shape)
cls_loss = self.focal_loss(fgb_cls_preds, ohe_targets)
if verbose:
np_loc_loss = t2np(loc_loss)
np_cls_loss = t2np(cls_loss)
np_n_fg = t2np(n_fg)
# np_n_fg = batch_size
print('loc_loss: %.5f | cls_loss: %.4f' % (np_loc_loss / np_n_fg, np_cls_loss / np_n_fg), end=' | ')
return (loc_loss + cls_loss) / n_fg
import utils
dim = 1
sz = np.tile(30, dim) # size of the desired images: (sz)^dim
params = dict()
params['len_s'] = sz.min() // 6
params['len_l'] = sz.min() // 3
# create a default image size with two sample squares
cs = eg.CreateSquares(sz)
I0, I1 = cs.create_image_pair(params)
# create the source and target image as pyTorch variables
ISource = torch.from_numpy(I0)
ITarget = torch.from_numpy(I1)
# spacing so that everything is in [0,1]^2 for now
spacing = 1. / (sz - 1)
print('Spacing = ' + str(spacing))
s = sf.SmootherFactory(sz, spacing).create_smoother('diffusion', {'iter': 10})
r = s.smooth_scalar_field(ISource)
plt.figure(1)
plt.plot(utils.t2np(ISource))
plt.plot(utils.t2np(r))
plt.plot(utils.t2np(ISource))
plt.show()
start = time.time()
for iter in range(100):
def closure():
optimizer.zero_grad()
# Forward pass: Compute predicted y by passing x to the model
IWarped = model(ISource)
# Compute loss
loss = criterion(IWarped, ITarget)
loss.backward()
return loss
optimizer.step(closure)
cIWarped = model(ISource)
if iter % 1 == 0:
energy, similarityEnergy, regEnergy = criterion.get_energy(
cIWarped, ITarget)
print(
'Iter {iter}: E={energy}, similarityE={similarityE}, regE={regE}'.
format(iter=iter,
energy=utils.t2np(energy),
similarityE=utils.t2np(similarityEnergy),
regE=utils.t2np(regEnergy)))
if iter % 10 == 0:
vizReg.showCurrentImages(iter, ISource, ITarget, cIWarped)
print('time:', time.time() - start)
# now create the grids xvals = np.array(np.linspace(-1, 1, sz)) yvals = np.array(np.linspace(-1, 1, sz)) YY, XX = np.meshgrid(xvals, yvals) grid = np.zeros([1, sz, sz, 2], dtype='float32') grid[0, :, :, 0] = XX + 0.2 grid[0, :, :, 1] = YY ISource = torch.from_numpy(I0) gridV = torch.from_numpy(grid) #print input2.data #s = STN(layout = 'BCHW') s = STN() start = time.time() out = s(ISource, gridV) plt.figure(1) plt.subplot(121) plt.imshow(I0[0, :, :, 0]) plt.subplot(122) plt.imshow(utils.t2np(out[0, :, :, 0])) print(out.size(), 'time:', time.time() - start) start = time.time() out.backward(gridV.data) print(gridV.grad.size(), 'time:', time.time() - start)
def visualize1DGrad():
plt.figure(1)
plt.plot(utils.t2np(grad[0, :]))
plt.title('grad')
plt.show()