search

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

Issue after re-train of darknet model #423

Open EnricoBeltramo opened 6 years ago

EnricoBeltramo commented 6 years ago

I created a model with 4 classes (cat,dog,car,human) starting from and original tiny-yolo-voc downloaded from https://pjreddie.com/darknet/yolo/, and after I trained it using 2007 dataset downloaded from: https://pjreddie.com/projects/pascal-voc-dataset-mirror/

I used follow command to train (I started from tiny-yolo-voc model and weigths downloaded from darknet) flow --model cfg\tiny-yolo-voc-4c.cfg --load tiny-yolo-voc.weights --train --dataset "\VOCtrainval_06-Nov-2007\VOCdevkit\VOC2007\JPEGImages" --annotation "\VOCtrainval_06-Nov-2007\VOCdevkit\VOC2007\Annotations"

I stopped @ checkpoint 22125

after saved in pb format: flow --model cfg/tiny-yolo-voc-4c.cfg --load -1 --savepb

and after tested the new trained flow --pbLoad built_graph/tiny-yolo-voc-4c.pb --metaLoad built_graph/tiny-yolo-voc-4c.meta --imgdir sample_img/

but the new network looks don't be able to detect anything in sample dir images. Did I do something wrong?

lpnueg4 commented 6 years ago

@EnricoBeltramo Have you solved this problem? I meet this problem too. python3 ./flow --model cfg/ff.cfg --load 35000 --imgdir test/ff/images/ --json use the ckpt, I can get result, but when I run python3 ./flow --model cfg/ff.cfg --load 3500 --savepb and run python3 ./flow --imgdir test/ff/tt/ --pbLoad built_graph/ff.pb --metaLoad built_graph/ff.meta --gpu 0.8 --json the json file was empty @thtrieu HELP!