Train and Convert SSD Resnet-50 model to TFLite

Retrain the model on your own dataset and convert it into TFLite to deploy on mobile devices and Coral Dev Board.

Pre-requisites

• Tensorflow-gpu==1.12.3 or Tensorflow-gpu==1.14
• Bazel==0.24.0
• cuda==9.0
• cuDNN==7.1

Step 1:

Download the object_detection API from github repository or clone it by using the following command git clone https://github.com/tensorflow/models.git. Following is the hierarchy of the models folder:

models    
│
└───research
│   │    
│   └───object_detection
│   │    
│   └───slim
│   │    
│   └───...
│  
└─── ...

The objective of this work is to convert the pretrained SSD Resnet-50 object detection model into TFLite, therefore only slim and object_detection directories are required from the models.

Step 2

Download a SSD Resnet-50 model from a collection of pretrained models Tensorflow Model Zoo and move it to the object_detection folder.

• Extract it here
• Rename the folder to ssd-resnet-50.

Step 3

Set environment variable using the following command:

export PYTHONPATH=$PYTHONPATH:/home/models:/home/models/research:/home/models/research/slim

Note: I placed the models in home directory.

Step 4

Open terminal and change your directory using cd /home/models/research. Further, generate protoc scripts using the following command:

python3.6 setup.py install
protoc object_detection/protos/*.proto --python_out=.

After generating the protoc scripts change working directory to object_detection.

cd /home/models/research/object_detection

Step 5

Verify your tensoflow setup by running a sample script jupyter notebook object_detection_tutorial.ipynb, if it works you are good to go.

Step 6

Annotate your data and convert xml file into a csv file using the following script:

import os
import glob
import pandas as pd
import xml.etree.ElementTree as ET

for dir in ['train', 'test']:
    path = os.path.join(os.getcwd(), ('dataset/' + dir))
    xml_list = []
    for xml_file in glob.glob(path + '/*.xml'):
        tree = ET.parse(xml_file)
        root = tree.getroot()
        for member in root.findall('object'):
            value = (root.find('filename').text,
                     int(root.find('size')[0].text),
                     int(root.find('size')[1].text),
                     member[0].text,
                     int(member[4][0].text),
                     int(member[4][1].text),
                     int(member[4][2].text),
                     int(member[4][3].text)
                     )
            xml_list.append(value)
    column_name = ['filename', 'width', 'height', 'class', 'xmin', 'ymin', 'xmax', 'ymax']
    df = pd.DataFrame(xml_list, columns=column_name)
    df.to_csv(('dataset/' + dir + '_labels.csv'), index=None)
    print('Successfully converted xml to csv.')

Edit the generate_tfrecord.py and verify the label mapping and execute it to generate the tfrecord file as required by the tensorflow.

# Change the following function in "generate_tfrecord.py"
def class_text_to_int(row_label):
    if row_label == 'cat':
        return 1
    elif row_label == 'dog':
        return 2
    else:
        None

• python3.6 generate_tfrecord.py --csv_input=dataset/train_labels.csv --image_dir=dataset/train --output_path=train.record
• python3.6 generate_tfrecord.py --csv_input=dataset/test_labels.csv --image_dir=dataset/test --output_path=test.record

Step 7

Generate a labelmap.pbtxt in training folder and define all your classes.

item {
  id: 1
  name: 'cat'
}

item {
  id: 2
  name: 'dog'
}

Copy pipeline.config from ssd-resnet-50 folder to training folder and rename it as ssd_resnet_50_config.config for readability. Edit the config file and change the following properties:

• num_classes (Set the number of classes in your dataset)
• fine_tune_checkpoint (Set the path of model.ckpt from ssd-resnet-50 folder)
• label_map_path (Set the path of labelmap.pbtxt from training folder)
• input_path (Set the path of train.record in train_input_reader)
• input_path (Set the path of test.record in eval_input_reader)
• num_examples (in eval_config)
• from_detection_checkpoint: true (Add after fine_tune_checkpoint in case of missing in config file)

Step 8

Execute the following script to start training.

python3.6 train.py --logtostderr --train_dir=training/ --pipeline_config_path=training/ssd_resnet_50_config.config

Step 9

Generate the frozen inference graph (.pb file) of SSD model.

python3.6 export_inference_graph.py --input_type image_tensor --pipeline_config_path training/ssd_resnet_50_config.config --trained_checkpoint_prefix training/model.ckpt-XXXX --output_directory freezed_model

Export frozen graph for TFLite.

python3.6 export_tflite_ssd_graph.py --pipeline_config_path=training/ssd_resnet_50_config.config --trained_checkpoint_prefix=training/model.ckpt-XXXX --output_directory=tflite_model --add_postprocessing_op=true

Step 10

Download the source code of Tensorflow from github or clone it using git clone https://github.com/tensorflow/tensorflow.git. After downloading the source code, open terminal and change working directory to cd /home/tensorflow and execute the following command to convert the SSD model into TFLite.

tflite_convert  --graph_def_file=/home/models/research/object_detection/tflite_model/tflite_graph.pb --output_file=/home/models/research/object_detection/tflite_model/detect.tflite --output_format=TFLITE --input_shapes=1,640,640,3 --input_arrays=normalized_input_image_tensor --output_arrays='TFLite_Detection_PostProcess','TFLite_Detection_PostProcess:1','TFLite_Detection_PostProcess:2','TFLite_Detection_PostProcess:3' --inference_type=FLOAT --mean_values=128 --std_dev_values=127 --change_concat_input_ranges=false --allow_custom_ops

Step 11

Use the .tflite file to deploy the model on mobile devices and Coral Dev Board.

Inference Code

Use the follwing code to test the converted model.

interpreter = tf.lite.Interpreter(model_path=PATH_TO_MODEL)
interpreter.allocate_tensors()

# Get input and output tensors.
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

# Load image for testing
image = cv2.imread(os.path.join(os.getcwd(),IMAGE_NAME))

input_data_f32 = np.array(image, dtype=np.float32)
input_data = np.expand_dims(input_data_f32, axis=0)
interpreter.set_tensor(input_details[0]['index'], input_data)

interpreter.invoke()

output_data = interpreter.get_tensor(output_details[0]['index'])

References

https://www.tensorflow.org/lite/convert/cmdline_examples