deviandice/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ExtensibleTrainer - don't compute backward when there is no loss

Browse Source
This commit is contained in:
James Betker 2020-10-02 20:54:06 -06:00
parent 146a9125f2
commit 567b4d50a4
1 changed files with 8 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
codes/models/steps/steps.py
View File

@ -159,12 +159,14 @@ class ConfigurableStep(Module):
# Scale the loss down by the accumulation factor.
total_loss = total_loss / self.env['mega_batch_factor']
# Get dem grads!
if self.env['amp']:
with amp.scale_loss(total_loss, self.optimizers, amp_loss_id) as scaled_loss:
scaled_loss.backward()
else:
total_loss.backward()
# In some cases, the loss could not be set (e.g. all losses have 'after'
if isinstance(loss, torch.Tensor):
# Get dem grads!
if self.env['amp']:
with amp.scale_loss(total_loss, self.optimizers, amp_loss_id) as scaled_loss:
scaled_loss.backward()
else:
total_loss.backward()
# Detach all state variables. Within the step, gradients can flow. Once these variables leave the step
# we must release the gradients.