glenn-jocher commited on
Commit
a21bd06
·
1 Parent(s): 455f7b8

Update train.py forward simplification

Browse files
Files changed (1) hide show
  1. train.py +3 -9
train.py CHANGED
@@ -265,18 +265,12 @@ def train(hyp, opt, device, tb_writer=None):
265
  ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] # new shape (stretched to gs-multiple)
266
  imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
267
 
268
- # Autocast
269
  with amp.autocast(enabled=cuda):
270
- # Forward
271
- pred = model(imgs)
272
-
273
- # Loss
274
- loss, loss_items = compute_loss(pred, targets.to(device), model) # scaled by batch_size
275
  if rank != -1:
276
  loss *= opt.world_size # gradient averaged between devices in DDP mode
277
- # if not torch.isfinite(loss):
278
- # logger.info('WARNING: non-finite loss, ending training ', loss_items)
279
- # return results
280
 
281
  # Backward
282
  scaler.scale(loss).backward()
 
265
  ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] # new shape (stretched to gs-multiple)
266
  imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
267
 
268
+ # Forward
269
  with amp.autocast(enabled=cuda):
270
+ pred = model(imgs) # forward
271
+ loss, loss_items = compute_loss(pred, targets.to(device), model) # loss scaled by batch_size
 
 
 
272
  if rank != -1:
273
  loss *= opt.world_size # gradient averaged between devices in DDP mode
 
 
 
274
 
275
  # Backward
276
  scaler.scale(loss).backward()