Use euclidean distance for stop criterion
This commit is contained in:
parent
4f1514accf
commit
1d39184036
6
train.py
6
train.py
|
@ -43,10 +43,6 @@ batch = 0
|
|||
# Accumulated loss for last batches
|
||||
running_loss = 0
|
||||
|
||||
# True in the first training phase where small batches are used, and false in
|
||||
# the second phase where point positions are refined using large batches
|
||||
is_coarse_optim = True
|
||||
|
||||
# List of training examples (not shuffled)
|
||||
classes_ordered = torch.arange(order).repeat(batch_size)
|
||||
|
||||
|
@ -84,7 +80,7 @@ while total_change >= 1e-4:
|
|||
# Check for convergence
|
||||
model.eval()
|
||||
constellation = model.get_constellation()
|
||||
total_change = (constellation - prev_constellation).abs().sum()
|
||||
total_change = (constellation - prev_constellation).norm(dim=1).sum()
|
||||
prev_constellation = constellation
|
||||
|
||||
# Report loss
|
||||
|
|
Loading…
Reference in New Issue