• Automation framework: CK
  
• Development repository: ck-object-detection
  
• Source: GitHub
  
• Available command lines:
  
  • ck run program:retrain-tfzoo-cocotokitti --cmd_key=a:create_dataset (META)
    
  • ck run program:retrain-tfzoo-cocotokitti --cmd_key=b:retrain (META)
    
  • ck run program:retrain-tfzoo-cocotokitti --cmd_key=c:export_model (META)
    
• Support for host OS: any
  
• Support for target OS: any
  
• Tags: object-detection,tf,lang-python,retrain,cocotokitti
  
• How to get the stable version via the client:
  
  pip install cbench
  cb download program:retrain-tfzoo-cocotokitti --version=2.5.0 --all
  ck run program:retrain-tfzoo-cocotokitti
• How to get the development version:
  
  pip install ck
  ck pull repo:ck-object-detection
  ck run program:retrain-tfzoo-cocotokitti
  

CK Object Detection Retraining Utilities

Warning: Not all the operations in this program can be automatized, some jobs in the pipeline MUST be manually handled by the user. Moreover, this will work only with models coming from the tensorflow zoo.

Prerequisites

Install CK

Please follow the CK installation instructions.

Pull CK repositories

$ ck pull repo:ck-tensorflow

Install TensorFlow

Install TensorFlow v1.14 from source (built with Bazel):

$ ck install package --tags=lib,tensorflow,v1.14,vsrc,vcpu

or from a binary x86_64 package (installed via pip):

$ ck install package --tags=lib,tensorflow,v1.14,vprebuilt,vcpu

Replace vcpu with vcuda to install TensorFlow with GPU support.

Or you can choose interactively from any available version of TensorFlow:

$ ck install package --tags=lib,tensorflow

Install TensorFlow models

$ ck install package --tags=tensorflow,model,api

Install one or more object detection model package:

$ ck install package --tags=tf,model,object-detection

Datasets

Install the kitti full dataset

$ ck install package --tags=dataset,object-detection,kitti-full

Other Dependencies

Install the pycocotool

$ ck install package --tags=tool,coco

Programs

There are three different programs in this pipeline:

In order to obtain a completely working model, you will need to run them in this order and add some manual work between the dataset creation and the retrain.

a:create_dataset

This program will prepare the required data for the retraining. All the parameters are provided by default ( I am not sure about the CK_OUTPATH, which is an empty string, if it means current working directory or if i just forgot it, need to test it). They can be overridden by setting the appropriate CK env variables.

After the create dataset, the pipeline.config is copied in the /tmp folder, which is the working directory. Now you have to manually edit the file. There are some more mechanical tasks, and other less intuitive and it may not work. In that case, it is possible that they changed some parameters in the tensorflow object detection repository, and you may have to search on google how to edit the pipeline in order to retrain/finetune.

The mandatory (and easy) tasks that must be performed are: - Change the dataset-related info in the first block. In particular, we target the kitti dataset so this have 7 classes (car,van,truck,pedestrian,person_sitting,cyclist,tram). Moreover, the images in the dataset have a different shape, so the resizer must be set at 375(height)*1242(width). - Modify the path wherever the PATH_TO_BE_CONFIGURED string appears. You must use ABSOLUTE path. - If you want to retrain, you will have to provide a starting checkpoint (again, you will need an absolute path). And if not set you will need to set the from_detection_checkpoint: true and load_all_detection_checkpoint_vars: true

Other possible issues (from my experience): - if a schedule with step 0 appears in the pipeline.config, in the manual_step_learning_rate, has to be removed. It is an old version of the API and has been removed and it causes the training to fail. - with lowproposals models, the fiels "second_stage_batch_size" has to be added, immediately after "first_stage_objectness_loss_weight", with a value lower than "first_stage_max_proposals".

b:retrain

This program will perform the actual retraining. All the parameters are set by default beside one, the number of steps for the retraining. This must be set when calling the program, using the CK env variable 'CK_NUM_STEPS'.

c:export_model

This program will create the frozen graph, with the checkpoints and everything that is needed to "restart" in the future, and also creating the graph that can be used for inference. It will require a mandatory parameter, 'CK_CHECKPOINT_PREFIX', that will select the checkpoint to use for the export. it must be one of the checkpoints in the 'model.ckpt' folder, contained in the working directory (called tmp). The exported graph will be crated in the 'output_dir' inside the tmp folder. This can be overridden by changing the 'CK_OUT_DIR' env variable. All the other parameters are set by default, and can be overridden by using the appropriate CK variables

• 2.5.0 (2020-09-04)
• 2.0.3 (2020-09-03)
• 1.0.1 (2020-09-02)
• 1.0.0 (2019-10-08)

• .cm/desc.json (3 bytes)
• .cm/info.json (321 bytes)
• .cm/meta.json (1934 bytes)
• README.md (4906 bytes)
• create_ds.py (1536 bytes)
• export.py (2166 bytes)
• kitti_label_map.pbtxt (248 bytes)
• retrain.py (4575 bytes)