Hi I have a model which is already trained, pruned and retrained
and now I have a totally new ground truth labeled dataset that I want to check my model with it.
option A- If I run inference I do not get there an option to get mAP (I think)
option 2 - I can run evaluation BUT, here you say that evaluation must be run on THE SAME training-specification file
You can use option2. To run evaluation.
You can use the new path to the new validation directory.
Oh I already tried…
When I get to the cell of the retrained model, I did like this:
Here is a screen from both specs files:
Please help me solve this issue. Thank you
Can you share below?
here is s label example:
18-26-41_84_RGB.txt (1020 Bytes)
Here are 2 logs:
1st one is when the evaluation DOES work - when I run it using the original specs file:
log_where_evaluation_DOES_work (2).txt (2.4 MB)
2nd log is where the evaluation DID NOTwork:
log_where_evaluation_does NOT_work.txt (128.9 KB)
Also here you can see how I build my docker paths in the notebook cell:
Please change the label files.
Got label marked as difficult(occlusion > 0), please set occlusion field in KITTI label to 0
Refer to Data Annotation Format - NVIDIA Docs
For example, please change to
someclass1 0 0 0 317.2704700000002 182.39025750000013 360.75424000000027 232.31606750000017 0 0 0 0 0 0 0
Wow! First of all thank you, it did help!
Let be ask you something please.
Actually if you noticed I had 4 datasets each included training and validation folders. What I did was taking one of them putting in some new directory and tried to run evaluation on it like if it was new.
Now When I look at my 4 datasets 3 of them have this strange -1 value in occlusion instead of 0. Do you think it can also be a reason to relatively poor training mAP (about 0.4-0.5)?
You can set occlusion to 0 in all the labels. And train again.
I am on it, thank you so much again!
Please answer me in another place about if it is posible to make automatc split to train and val without tfrecords