Welcome

This repository is the final Capstone Project for Data Science & Artificial Intelligence program

Notes: Tensorflow model on custom data (Transfer Learning).The objective is to identify Formula One Racing Team. Please contact me if you are interested in the inference graph and custom dataset.

Instructions: Download and install tensorflow model. The files have been converted to Jupyter Notebook for readability. However, i do recommended to use the python files for implementation.

---

1. Clone the master branch of Tensorflow models repository

git clone https://github.com/tensorflow/models.git

2. Install protobuf

conda install -c anaconda protobuf

3. Compile Protobufs

cd models/research

protoc object_detection/protos/*.proto --python_out=.

4. Install Tensorflow Object Detection Library

cd object_detection/packages/tf2

python setup.py

5. Test if you have correctly install all the library

cd object_detection/builders

python model_builder_tf2_test.py

You are ready to go if you see this code:

Ran 20 tests in 13.823s

OK (skipped=1)

6. Clone my repository on a separate folder

git clone https://github.com/RickFSA/Capstone_Object_Detection.git

7. Go to the directory of the folder to open the jupyter notebook

jupyter notebook Object_detection_image.ipynb

From this notebook you need to specify the path to the trained model inference_graph/saved_model (370MB)

It should take 15s to load the model, then you are ready to use the model to predict any image from the F1 Formula images.

I have included some images & video for testing.

Please send a request to [email protected] for the dataset & inference graph.

---

Train on custom dataset

---

1. Generate csv from xml:

python xml_to_csv.py

2. Adjust class label from generate_tfrecord.py

code 35 from files

3. Generate TFRecords from csv:

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

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

4. Config files

chanage input, label, model_checkpoint

5. Model training:

python model_main_tf2.py
--pipeline_config_path=training/faster_rcnn_resnet101_v1_800x1333_coco17_gpu-8.config --model_dir=training --alsologtostderr

6. Tensorboard:

tensorboard --logdir=training/train

7. Extract inference graph (change the config to your selected model):

python exporter_main_v2.py --pipeline_config_path training/faster_rcnn_resnet101_v1_800x1333_coco17_gpu-8.config --trained_checkpoint_dir training --output_directory inference_graph