Open 1726400669 opened 4 years ago
have you tried training using eager mode does it work?
I made a tutorial on training from scratch. https://github.com/zzh8829/yolov3-tf2/blob/master/docs/training_voc.md let me know if this helps you, in my experience training without transfer learning from darknet is almost impossible
Hi everyone, I used the instructions on the tutorial in this repo to train it on my custom pictures. I converted the darknet weight using convert.py and then used the eager_fit mode and transfer learning using this command: python train.py --dataset ./data/Ortho_Dataset/ortho_train.tfrecord --val_dataset ./data/Ortho_Dataset/ortho_valid.tfrecord --classes ./data/obj.names --num_classes 1 --mode eager_fit --transfer darknet --batch_size 16 --epochs 10 --weights ./checkpoints/yolov3.tf --weights_num_classes 80
The training results are shown in this picture.
Although I donot encounter any problem when doing my taining, the model cannot detect anything from yolov3 and it returns zero and no detections. Refer to the following pic:
Does anyone have any idea what could be the problem?
Thanks,
After I made my dataset and trained for 50 epochs with 1 class, I couldn't detect anything. I have changed the classes of
detect.py
to mine, weights to./checkpoints/yolov3_train_50.tf
, image to mine, andnum_classes
to 1, but the detection results are all 0 and lots of Warning like https://github.com/zzh8829/yolov3-tf2/issues/108 .This is where I modify train.py
and This is where I modify model.py:
Thanks.