Train Classifier on ImageNet (ILSVRC2012)

Train Classifier on ImageNet (ILSVRC2012)

Required files:

1. imagenet.labels.list - labels of objects (classes), filenames of training images should contain these labels, for example img_n02096294.jpg, example: https://github.com/AlexeyAB/darknet/blob/master/data/imagenet.labels.list

2. imagenet.shortnames.list - names of objects (classes), example: https://github.com/AlexeyAB/darknet/blob/master/data/imagenet.shortnames.list

3. imagenet1k.train.list - paths to Training images

4. Annotated Training images - you can get these Images and imagenet1k.train.list file by using this script (138 GB) https://github.com/AlexeyAB/darknet/blob/master/scripts/get_imagenet_train.sh

   • or download ILSVRC2012_img_train.tarfrom Torrent by yourself and then run get_imagenet_train.sh in the same folder with this tar-file: http://academictorrents.com/browse.php?search=imagenet&page=0

5. imagenet1k.val.list - paths to Validation images (for checking accuracy Top1 / Top5)

6. Annotated Validation images - you can get these Images and imagenet1k.val.list file by using these scripts:

   • images - 6.3 GB: http://www.image-net.org/challenges/LSVRC/2012/nnoupb/ILSVRC2012_img_val.tar
   • annotations - 2.2 MB: http://www.image-net.org/challenges/LSVRC/2012/nnoupb/ILSVRC2012_bbox_val_v3.tgz
   • Set validation labels: https://github.com/AlexeyAB/darknet/blob/master/scripts/imagenet_label.sh
   • read: https://pjreddie.com/darknet/imagenet/

7. imagenet1k.data - file with content:

classes = 1000
train  = data/imagenet1k.train.list
valid  = data/inet.val.list
backup = backup
labels = data/imagenet.labels.list
names  = data/imagenet.shortnames.list
top=5

• csdarknet53-omega.cfg - network structure: https://github.com/AlexeyAB/darknet/blob/master/cfg/csdarknet53-omega.cfg

Look at other models for Classifier: https://pjreddie.com/darknet/imagenet/

---

Training

• Training command with accuracy checking (Annotated Validation images are required):

./darknet classifier train cfg/imagenet1k.data cfg/csdarknet53-omega.cfg -topk

• Training command without accuracy checking:

./darknet classifier train cfg/imagenet1k.data cfg/csdarknet53-omega.cfg

• Continue training:

./darknet classifier train cfg/imagenet1k.data cfg/csdarknet53-omega.cfg backup/csdarknet53-omega_last.weights -topk

You will get files

• chart.png with Loss & Top1 charts
• backup/csdarknet53-omega_last.weights - trained weights file

---

Classifictation:

Predict:

./darknet classifier predict cfg/imagenet1k.data cfg/csdarknet53-omega.cfg backup/csdarknet53-omega_last.weights dog.jpg

Check accuracy Top1 / Top5:

./darknet classifier valid cfg/imagenet1k.data cfg/csdarknet53-omega.cfg backup/csdarknet53-omega_last.weights