Reference

This repository was based on: repository

Steps

0. Set Environment Variables

Linux

export PYTHONPATH=$PYTHONPATH:$PWD:$PWD/slim:$PWD/object_detection

Windows

cmd

 set PYTHONPATH=%CD%;%CD%\slim;%CD%\object_detection

powershell

 $env:PYTHONPATH = (Get-Item -Path ".\").FullName + ';' + (Get-Item -Path ".\slim").FullName + ';' + (Get-Item -Path ".\object_detection").FullName

1. Choose the model

Tensorflow provides a series of pre-trained models: model zoo.

2. Gather and Label Pictures

Retrieve the images and label them. Possible tools:

3. Generate Training Data

PASCAL VOC

First covert all the labeled data (.xml) into a .csv

python xml_to_csv.py [-f] 'folder1,folder2,...,folderN' -im images

YOLO

First covert all the labeled data (.txt) into a .csv

python yolo_to_csv.py [-f] 'folder1,folder2,...,folderN'

Then, edit the label_map.txt, located in the training folder, with the classes of interest.

item {
  id: 1
  name: 'firstclass'
}

item {
  id: 2
  name: 'secondclass'
}

...

item {
  id: 4
  name: 'fourthclass'
}

After this, edit the generate_tfrecord.py with the same classes and ids of the label_map.txt.

def class_text_to_int(row_label):
    if row_label == 'firstclass':
        return 1
    elif row_label == 'secondclass':
        return 2
        ...
    elif row_label == 'fourthclass':
        return 4
    else:
        raise_excep(f"THIS CLASS {row_label} DOES NOT EXIST")

If your annotation files are in the YOLO Format, edit the yolo_to_csv.py with the same classes and ids of the label_map.txt.

def class_int_to_text(row_label):
    if row_label == 1:
        return 'firstclass'
    elif row_label == 2:
        return 'secondclass'
        ...
    elif row_label == 'fourthclass':
        return 4
    else:
        raise_excep(f"THIS CLASS {row_label} DOES NOT EXIST")

Then, generate the TF_record files:

python generate_tfrecord.py --csv_input=images/train_labels.csv --image_dir=images/ --output_path=train.record
python generate_tfrecord.py --csv_input=images/test_labels.csv --image_dir=images/ --output_path=test.record

5. Configure training

Finally, its time to configure the training step.

First, you will edit the .config file that you choosed:

Change num_classes to the number of different objects you want the classifier to detect. Line 9.
Set the right path to the .ckpt file. Line 110.
- fine_tune_checkpoint : "DOWNLOADED_PRETRAINED_MODEL_PATH/model.ckpt"
Set train input path to the train.record. Line 126.
- input_path : ".../train.record"
Set label map path to the training/labelmap.pbtxt. Line 128.
- label_map_path: ".../training/labelmap.pbtxt"
Change num_examples to the number of images you have in the \images\test directory. Line 132.
Set test input path to the train.record. Line 140.
- input_path : ".../test.record"
Set label map path to the training/labelmap.pbtxt. Line 142.
- label_map_path: ".../training/labelmap.pbtxt"

Save the file after the changes have been made. That’s it! The training job is all configured and ready to go!

6. Run the Training

Start the training:

python .\model_main_tf2.py --pipeline_config_path MODEL_DIR/pipeline.config --model_dir CKPT_OUTPUT_PATH --alsologtostderr

Start the evaluation:

python .\model_main_tf2.py --pipeline_config_path MODEL_DIR/pipeline.config --model_dir CKPT_OUTPUT_PATH  --checkpoint_dir CKPT_OUTPUT_PATH--alsologtostderr

To open the tensorboard:

tensorboard --logdir=results

7. Export Inference Graph

To export the model trained:

python .\exporter_main_v2.py --input_type image_tensor --pipeline_config_path MODEL_DIR\pipeline.config --trained_checkpoint_dir CKPT_OUTPUT_PATH --output_directory OUTPUT_DIR

8. Test Inference Graph