Skip to content

Bug fix in TAPIR #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ret-1 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Bug fix in TAPIR #4

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 14 additions & 10 deletions model/net.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -323,20 +323,21 @@ def finetune(self, dataloaders, dataset_size, log_dir, ckpt_path, epochs=10):
rgbs = rgbs.float().to(self.device) # B, S, H, W, C
rgbs = (2 * (rgbs / 255.0) - 1.0) # normalizing [-1 1]

trajs_g = trajs_g.permute(0, 2, 1, 3).to(self.device)
valids = visibs_g.permute(0, 2, 1).to(self.device) #B, S, N
trajs_g = trajs_g.permute(0, 2, 1, 3).to(self.device) #B, N, S, 2 format:(x,y)
valids = visibs_g.permute(0, 2, 1).to(self.device) #B, N, S
B, S, H, W, C = rgbs.shape
trajs_g[...,0] *= W - 1
trajs_g[...,1] *= H - 1
_, N, _, _ = trajs_g.shape

points_0 = trajs_g[:,0,:,:] # taking all points from frame 0
# from (x, y) to (t, x, y)
points_0_xy = trajs_g[:,:,0,:] #B, N, 2 taking all points from frame 0
# from (x, y) to (t, y, x)
points_0 = points_0_xy[:,:,[1,0]] #format:(y,x)
# preparing the time dimension to be concatenated
#points_0 = points_0.cpu().numpy()
time_dim = torch.zeros((points_0.shape[0], points_0.shape[1], 1)).to(self.device)
# prepending a column to be -> (B, N, 3)
points_0 = torch.concatenate((time_dim, points_0), axis=-1)
points_0 = torch.concatenate((time_dim, points_0), axis=-1) #format:(t, y, x)

if phase == 'train':
self.model.train() # Set model to training mode
Expand All @@ -350,16 +351,19 @@ def finetune(self, dataloaders, dataset_size, log_dir, ckpt_path, epochs=10):
outs, loss = self.model(video=rgbs, query_points=points_0,
points_gt=trajs_g, visibs_gt=valids)

trajs_e = outs['tracks']
trajs_e = outs['tracks'] #B, N, S, 2 format:(x,y)
occlusions = outs['occlusion']
expected_dist = outs['expected_dist']
visibs_e = (1 - F.sigmoid(occlusions)) * (1 - F.sigmoid(expected_dist)) > 0.5
trajs_e[...,0] /= W - 1
trajs_e[...,1] /= H - 1
trajs_g[...,0] /= W - 1
trajs_g[...,1] /= H - 1
points_0[...,0] /= W - 1
points_0[...,1] /= H - 1
points_0_xy[...,0] /= W - 1
points_0_xy[...,1] /= H - 1
#print(points_0.shape, trajs_g.shape, trajs_e.shape, visibs_g.shape)
outputs = evaluate.compute_metrics(points_0.cpu().numpy(), trajs_g.permute(0, 2, 1, 3).cpu().numpy(),
visibs_g.cpu().numpy(), trajs_e.detach().cpu().numpy(), visibs_g.cpu().numpy())
outputs = evaluate.compute_metrics(points_0_xy.cpu().numpy(), trajs_g.cpu().numpy(),
valids.cpu().numpy(), trajs_e.detach().cpu().numpy(), visibs_e.cpu().numpy())
for key, value in outputs.items():
metrics_b[key] += value

Expand Down
2 changes: 1 addition & 1 deletion utils/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ def sequence_loss(flow_preds, flow_gt, vis, valids, gamma=0.8):
flow_loss = 0.0
for i in range(n_predictions):
i_weight = gamma**(n_predictions - i - 1)
flow_pred = flow_preds[i]
flow_pred = flow_preds[i].permute(0, 2, 1, 3)
i_loss = (flow_pred - flow_gt).abs() # B,S,N,2
i_loss = torch.mean(i_loss, dim=3) # B,S,N
#flow_loss += i_weight * basic.reduce_masked_mean(i_loss, valids)
Expand Down