How to fine-tune?

[9thDimension]

Firstly, let me say thanks and congratulations on this ambitious and impressive project.

I've been perusing the code-base and I don't understand it very well. I would like to fine-tune YOLO to my own dataset, using a checkpoint that was learnt on VOC/COCO data.

Just as a test-run, to see what would happen, I ran

./flow --train --model cfg/tiny-yolo.cfg --load bin/tiny-yolo.weights

which came back with the error:

...
Load  |  Yep!  | conv 3x3p1_1  +bnorm  leaky      | (?, 13, 13, 1024)
Load  |  Yep!  | conv 3x3p1_1  +bnorm  leaky      | (?, 13, 13, 1024)
Load  |  Yep!  | conv 1x1p0_1    linear           | (?, 13, 13, 425)
-------+--------+----------------------------------+---------------
Running entirely on CPU
Traceback (most recent call last):
 File "./flow", line 42, in <module>
   tfnet = TFNet(FLAGS)
 File "/home/hal9000/Sources/darkflow/net/build.py", line 51, in __init__
   self.setup_meta_ops()
 File "/home/hal9000/Sources/darkflow/net/build.py", line 94, in setup_meta_ops
   if self.FLAGS.train: self.build_train_op()
 File "/home/hal9000/Sources/darkflow/net/help.py", line 15, in build_train_op
   self.framework.loss(self.out)
 File "/home/hal9000/Sources/darkflow/net/vanilla/train.py", line 9, in loss
   'Loss type {} not implemented'.format(loss_type)
AssertionError: Loss type [region] not implemented

So I ask:

(1) Which options on the flow shell script do I need to run fine-tuning?
(2) How do I point to my dataset, and how should it be formatted?

Thanks again.

[thtrieu]

(1) You did it correctly, the only reason you cannot fine-tune this config is because it is YOLO9000, for which I did not build the training part (as stated in README). If you want to fine-tune older version's configs, point to ./cfg/v1/ or ./cfg/v1.1/

(2) use option --dataset and --annotation, you can see the complete set of options by ./flow --h (again, as stated in README). The format should be identical to that of PASCAL VOC dataset.