thtrieu / darkflow

Translate darknet to tensorflow. Load trained weights, retrain/fine-tune using tensorflow, export constant graph def to mobile devices
GNU General Public License v3.0
6.13k stars 2.08k forks source link

the issue ckpt files to pb files #597

Open cdragon2 opened 6 years ago

cdragon2 commented 6 years ago

i think this is same issue with #297 i trained this command python3 flow --model cfg/tiny-yolo-voc-136c-combo18.cfg --train --dataset ./train/train15/Images --annotation ./train/train15/Annotations --labels ./englishlabels.txt --save 10000 --epoch 300000 --trainer adam --lr 1e-5 --gpu 1.0 --batch 32

and then when i check predictions, it was good

but after making pb file like this command python3 flow --model cfg/tiny-yolo-voc-136c-combo18.cfg --imgdir ./sample/english/ --lr 1e-5 --gpu 0.7 --batch 32 --load 74568 --labels ./englishlabels.txt --savepb

when i check predictions by pb file, that was different with ckpt file. i don't know why. the command that making pb file is incorrect?

Help me plz~~ .

liuhantao9 commented 6 years ago

I am wondering where did you get ckpt file from?

cdragon2 commented 6 years ago

@liuhantao9 ckpt files are in default path(./ckpt/) and --load 74568 means that i use /tiny-yolo-voc-136c-combo18-74568 checkpoints file

ssusie commented 6 years ago

Have the same problem...

alexanderfrey commented 6 years ago

+1

sillybear92 commented 6 years ago

I had the same problem. When training for the first time, you must load the pre-train weights. I solved it this way.