• Automation framework: CK
  
• Development repository: ck-ml
  
• Source: GitHub
  
• Available command lines:
  
  • ck run program:caffe-ssd-train-kitti --cmd_key=prepare (META)
    
  • ck run program:caffe-ssd-train-kitti --cmd_key=train (META)
    
• Support for host OS: any
  
• Support for target OS: any
  
• Tags: caffe-detection,caffe-detection-ssd,ssd,demo,training,object-detection
  
• How to get the stable version via the client:
  
  pip install cbench
  cb download program:caffe-ssd-train-kitti --version=3.0.0 --all
  ck run program:caffe-ssd-train-kitti
• How to get the development version:
  
  pip install ck
  ck pull repo:ck-ml
  ck run program:caffe-ssd-train-kitti
  

caffe-ssd-train-kitti

Demo program for fine-tuning of SSD caffemodels originaly trained on COCO and VOC datasets to detect objects of KITTI dataset.

Requirements

Caffe library:

ck install package:lib-caffe-ssd-cpu
ck install package:lib-caffe-ssd-cuda

Caffe SSD model and pretrained weights:

ck install package:caffemodel-ssd-coco-300
ck install package:caffemodel-ssd-coco-512
ck install package:caffemodel-ssd-voc-300
ck install package:caffemodel-ssd-voc-512

KITTI dataset:

ck install package:dataset-kitti-full

Run

There are two steps.

Prepare

ck run program:caffe-ssd-train-kitti --cmd_key=prepare

Here training and testing datasets are prepared. Original KITTI dataset contains image files, while SSD training uses LMBD as a data source. KITTI images are scaled to a size supported by the selected caffemodel (300 or 512 px) and CK_TRAIN_IMAGES_PERCENT of them are included into the train database and the rest into the test database.

Train

ck run program:caffe-ssd-train-kitti --cmd_key=train

Run training process using prepared data.

Essential anvironment variables:

CK_BATCH_SIZE

Specify the batch size.

CK_DEVICE_ID

The device that will be used in GPU mode. Run mode, CPU or GPU, is governed by the selected caffe library.

CK_SNAPSHOT_INTERVAL

The snapshot interval.

CK_TEST_INTERVAL

The number of iterations between two testing phases.

Results

Trained weights are saved into tmp/snapshots directory.

TODO

• Models *-512 fail with message Check failed: num_priors_ * num_classes_ == bottom[1]->channels(), it is because of dimensions of final layers depend on the number of detection classes. Though that layers are modified in program but something is missed for *-512 models (they have more layers than *-300s have).

• We could implement continued training by loading not the pretrained weights of the selected caffemodel but latest snapshot instead.

• Implement one more command key test that will detect specified number of KITTI images using latest snapshot then will convert detection results and run original KITTI evaluation program ($CK-TOOLS/demo-squeezedet-patched/squeezeDet/src/dataset/kitti-eval/cpp/evaluate_object.cpp) to calculate metrics in the same way as SqueezeDet does.

• 3.0.0 (2021-04-11)
• 1.0.0 (2019-10-08)

• .cm/desc.json (3 bytes)
• .cm/info.json (317 bytes)
• .cm/meta.json (1348 bytes)
• README.md (2337 bytes)
• prepare.py (6143 bytes)
• train.py (5780 bytes)
• utils.py (1839 bytes)