Tensorflow Object Detection
|
V
|-----Detection TF2.ipynb --> For detection
|-----Trainer TF2.ipynb --> For training
|-----addons --> LabelImg ->> Github Link to LabelImg
|-----models ->> github TFOD master
|-----scripts
|-----exporter_main_v2.py
|-----generate_tfrecord.py
|-----model_main_tf2.py
|-----workspace
|-----training
|-----annotations --> Contains labelmap and TF record files
|-----exported-models --> Exported model after training
|-----images --> Folders named train, test, eval and
detected-images for their specific images.
Each folder will have image and its corresponding
PASCAL VOC formated .xml files.
|-----pre-trained-models --> Each folder containing downloaded pre-trained models ->> Download from TFOD Model Zoo
|-----under-training-models --> Each folder containing customized pipeline config
for the pretrained models and its checkpoints
(Copy over the pipeline.config file with necessary
changes from either exported-models or pre-trained-models.)
Note:
Some folders may be missing due to upload limitations. So, create them if they don't exist using the same folder names given above. For instance, under-training-models/faster_rcnn folder is missing. So create them, then copy over the pipeline.config file from exported-models/10200 and make necessary changes. This needs to be done if you are planning to train.
See Configure the Training Pipeline for figuring out what changes need to be done in pipeline.config file.
In the anaconda prompt under the specific conda environment created for this project using requirements.txt:
For creating the train data TF record:
python scripts/generate_tfrecord.py -x ./workspace/training/images/train -l ./workspace/training/annotations/label_map.pbtxt -o ./workspace/training/annotations/train.record
For creating the test data TF record:
python scripts/generate_tfrecord.py -x ./workspace/training/images/test -l ./workspace/training/annotations/label_map.pbtxt -o ./workspace/training/annotations/test.record
For training the model:
python scripts/model_main_tf2.py --model_dir=./workspace/training/under-training-models/faster_rcnn --pipeline_config_path=./workspace/training/under-training-models/faster_rcnn/pipeline.config
For evaluating the model as it is getting trained, run the below command in a duplicated command prompt:
python model_main_tf2.py --model_dir=./workspace/training/under-training-models/faster_rcnn --pipeline_config_path=./workspace/training/under-training-models/faster_rcnn/pipeline.config --checkpoint_dir=./workspace/training/under-training-models/faster_rcnn
For exporting the model:
python scripts\exporter_main_v2.py --input_type image_tensor --pipeline_config_path ./workspace/training/under-training-models/faster_rcnn/pipeline.config --trained_checkpoint_dir ./workspace/training/under-training-models/faster_rcnn/ --output_directory ./workspace/training/exported-models
Models must to exported or frozen for restarting training from where we left off. To use exported model for training, change the fine_tune_checkpoint directory to refer to the ckpt-0 of exported model.
For monitoring the training and/or evaluation job progress using TensorBoard during/after the process, run the below command in a duplicated command prompt:
tensorboard --logdir=./workspace/training/under-training-models/faster_rcnn
Our research journal, Feel free to cite it in your research/project: Democratizing AI by converting sketches to code
For more detailed procedure explanation: Training Custom Object Detector using Tensorflow 2 Object Detection API
Tensorflow 2 Object Detection: Tensorflow Model Garden
Tensorflow 2 Model Zoo : Go there
Faster RCNN pre-trained Model: Download from Tensorflow 2 Model Zoo