def test_transform_siamfc(self):
base_dataset = VOT(self.vot_dir, return_rect=True, download=True)
transform = TransformSiamFC(stats_path=self.stats_path)
dataset = Pairwise(base_dataset, transform=transform, subset='train')
self.assertGreater(len(dataset), 0)
for crop_z, crop_x, labels, weights in dataset:
self.assertAlmostEqual(weights[labels == 1].sum().item(),
weights[labels == 0].sum().item())
self.assertAlmostEqual(weights.sum().item(),
labels[labels >= 0].numel())
self.assertEqual(
weights[labels == transform.ignore_label].sum().item(), 0)
if self.visualize:
crop_z, crop_x, labels, weights = random.choice(dataset)
crop_z = F.to_pil_image(crop_z / 255.0)
crop_x = F.to_pil_image(crop_x / 255.0)
labels = self._rescale(labels.cpu().squeeze().numpy())
weights = self._rescale(weights.cpu().squeeze().numpy())
bndbox_z = np.array([31, 31, 64, 64])
bndbox_x = np.array([95, 95, 64, 64])
show_frame(crop_z, bndbox_z, fig_n=1, pause=1)
show_frame(crop_x, bndbox_x, fig_n=2, pause=1)
show_frame(labels, fig_n=3, pause=1, cmap='hot')
show_frame(weights, fig_n=4, pause=5, cmap='hot')
def test_pairwise(self):
base_dataset = OTB(self.otb_dir, download=True)
frame_range = random.choice([0, 1, 100])
causal = random.choice([True, False])
subset = random.choice(['train', 'val'])
return_index = random.choice([True, False])
rand_choice = random.choice([True, False])
dataset = Pairwise(
base_dataset, pairs_per_video=1, frame_range=frame_range,
causal=causal, subset=subset, return_index=return_index,
rand_choice=rand_choice)
self.assertGreater(len(dataset), 0)
for i, item in enumerate(dataset):
img_z, img_x, bndbox_z, bndbox_x = \
item[0], item[1], item[2], item[3]
if return_index:
print('rand_z:', item[4], '\trand_x:', item[5])
self.assertEqual(img_z.mode, 'RGB')
self.assertEqual(img_x.mode, 'RGB')
self.assertEqual(bndbox_z.shape, (4,))
self.assertEqual(bndbox_x.shape, (4,))
if self.visualize:
item = random.choice(dataset)
img_z, img_x, bndbox_z, bndbox_x = \
item[0], item[1], item[2], item[3]
if return_index:
print('rand_z:', item[4], '\trand_x:', item[5])
show_frame(img_z, bndbox_z, fig_n=1, pause=1)
show_frame(img_x, bndbox_x, fig_n=2, pause=1)
def test_transform_goturn(self):
base_dataset = VOT(self.vot_dir, return_rect=True, download=True)
transform = TransformGOTURN()
dataset = Pairwise(base_dataset,
transform,
pairs_per_video=1,
frame_range=1,
causal=True)
self.assertGreater(len(dataset), 0)
for crop_z, crop_x, labels in dataset:
self.assertEqual(crop_z.size(), crop_x.size())
if self.visualize:
for t in range(10):
crop_z, crop_x, labels = random.choice(dataset)
mean_color = torch.tensor(transform.mean_color).float().view(
3, 1, 1)
crop_z = F.to_pil_image((crop_z + mean_color) / 255.0)
crop_x = F.to_pil_image((crop_x + mean_color) / 255.0)
labels = labels.cpu().numpy()
labels *= transform.out_size / transform.label_scale_factor
bndbox = np.concatenate([labels[:2], labels[2:] - labels[:2]])
show_frame(crop_x, bndbox, fig_n=1, pause=1)
def test_goturn_train(self):
tracker = TrackerGOTURN(net_path=self.net_path)
transform = TransformGOTURN()
base_dataset = VOT(self.vot_dir, return_rect=True, download=True)
dataset = Pairwise(
base_dataset, transform, frame_range=1, causal=True)
dataloader = DataLoader(dataset, batch_size=2, shuffle=True)
# training loop
for it, batch in enumerate(dataloader):
update_lr = it == 0
loss = tracker.step(batch, backward=True, update_lr=update_lr)
print('Iter: {} Loss: {:.6f}'.format(it + 1, loss))
# validation loop
for it, batch in enumerate(dataloader):
loss = tracker.step(batch, backward=False)
print('Val. Iter: {} Loss: {:.6f}'.format(it + 1, loss))
def test_siamfc_train_v2(self):
tracker = TrackerSiamFC(branch='alexv2')
transform = TransformSiamFC(
stats_path=self.stats_path, score_sz=33,
r_pos=8, total_stride=4)
base_dataset = VOT(self.vot_dir, return_rect=True, download=True)
dataset = Pairwise(base_dataset, transform, pairs_per_video=1)
dataloader = DataLoader(dataset, batch_size=2, shuffle=True)
# training loop
for it, batch in enumerate(dataloader):
update_lr = it == 0
loss = tracker.step(batch, backward=True, update_lr=update_lr)
print('Iter: {} Loss: {:.6f}'.format(it + 1, loss))
# val loop
for it, batch in enumerate(dataloader):
loss = tracker.step(batch, backward=False)
print('Val. Iter: {} Loss: {:.6f}'.format(it + 1, loss))