Commit History
refactor dataloader
22fb2b0
small dataset bug fix #140
256a3e8
model.names multi-GPU bug fix #94
b50fdf1
Fix yaml saving (don't sort keys), reorder --opt keys, bug fix hyp dict accessor
de19165
Fix bug in --help from percent sign in help string
2d396be
add stride to datasets.py
b8557f8
Turn opt into dictionary before sending it to yaml
611aacf
Default optimizer SGD
bc4ef48
Mode all optimizer settings to 'hyp.yaml', integrate proper momentum with Adam optimizer
7abf202
Fix help message for cfg files
7edbf65
Remove --resume functionality and related checks/logic.
d64ad0f
update google_utils import
6c1b87a
remove fast, add merge
1f1917e
Add plot_results save location to log_dir
e572bb0
Fix DDP bug in single process multiple device use cases
899f1d4
Merge pull request #107 from Lornatang/fix-reference-bugs
cdf1eac
unverified
backbone as FP16, save default to FP32
cce95e7
save ckpt in FP16 #119
d9b64c2
AutoAnchor bug fix # 117
9fdb0fb
Syntax fixes
c8152c8
Add save_dir arg to test.test, use arg as location for saving batch jpgs
9b7386f
Add save_dir to plot_lr_scheduler and plot_labels
945307b
Move results.txt from weights/ to log_dir
3b2b330
Fix hyp file read in and dict update.
ade023c
fix refrence bug
2368603
tb_writer bug fix
8db51c7
remove old print statements
5f2eeba
dist.destroy_process_group() bug fix
afe1df3
add update default hyp dict with provided yaml
333f678
add logic for resuming and getting hyp for resume run
a448c3b
add util function to get most recent last.pt file
25e51bc
add save_dir arg to plot_lr_scheduler, default to current dir.
490f1e7
change weights dir (wdir) to be unique to each run, under log_dir
4418809
add save yaml of opt and hyp to tensorboard log_dir in train()
d9f446c
add parser arg for hyp yaml file
a85e6d0
check_anchors bug fix
5a50491
check_anchors() bug fix #102
05b8ee5
check_anchors() bug fix #90
ec81c7b
AutoAnchor bug fix #72
8b26e89
update fast mode
bdd9fee
update fast mode
1c0b623
default check_git_status() to True
915b148
FP16 to FP32 ckpt load
14523bb
glob search bug fix #77
c5966ab
assert best possible recall > 0.9 before training
31f3310
Ensures weights/ dir exists
c3d4d32
Lauritzen Kasper Primdal
commited on