zjc
/
yolov5
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

backbone as FP16, save default to FP32

This commit is contained in:
Glenn Jocher 2020-06-18 00:13:18 -07:00
parent d9b64c27c2
commit cce95e744d
2 changed files with 2 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
train.py
View File

@ -332,7 +332,7 @@ def train(hyp):
ckpt = {'epoch': epoch,
'best_fitness': best_fitness,
'training_results': f.read(),
'model': ema.ema.module.half() if hasattr(model, 'module') else ema.ema.half(),
'model': ema.ema.module if hasattr(model, 'module') else ema.ema,
'optimizer': None if final_epoch else optimizer.state_dict()}
# Save last, best and delete

3
utils/utils.py
View File

@ -627,13 +627,12 @@ def strip_optimizer(f='weights/best.pt'): # from utils.utils import *; strip_op
def create_backbone(f='weights/best.pt', s='weights/backbone.pt'): # from utils.utils import *; create_backbone()
# create backbone 's' from 'f'
device = torch.device('cpu')
x = torch.load(f, map_location=device)
torch.save(x, s) # update model if SourceChangeWarning
x = torch.load(s, map_location=device)
x['optimizer'] = None
x['training_results'] = None
x['epoch'] = -1
x['model'].half() # to FP16
for p in x['model'].parameters():
p.requires_grad = True
torch.save(x, s)