Tensorflow API를 이용한 프로젝트
TF_Badminton IN-OUT Lineman Project
A project to read the in-out of badminton shuttlecock using the tensorflow object detection API
Created on a Colab basis.
Use tensorflow object detection API
1. Setup
- drive mount
2. Setting the Work Environment
- import the models for object detection
- tensorflow ssd mobilenet
- add tf libraries to pythonpath
- download pre-trained model
- ssd_mobilenet_v2_quantized_300x300_coco_2019_01_03
3. Update model pipeline configurations
- model
- Line 3: num_classes = number of classes in your dataset.
- train_config
- Line 157: fine_tune_checkpoint = MODEL_DIR/training/model.ckpt
- train_input_reader
- Line 162: label_map_path = MODEL_DIR/data/labelmap.pbtxt
- Line 164: input_path = MODEL_DIR/data/train.record
- eval_config
- Line 168: num_examples = number of test images
- eval_input_reader
- Line 174: label_map_path = MODEL_DIR/data/labelmap.pbtxt
- Line 178: input_path = MODEL_DIR/data/test.record
4. Train
- if ckpt-0: 37:36[click to run] ~ 39:35[activate first step] => about 2 min to start learning (2019-12-14)
- num_train_steps=100000
5. Tflite converting
title Download cocoapi_clone
- helped metrics per category successfully on github
- cocoapi_clone from some user for us
- Without it, 'export_tflite_ssd_graph.py' doesn't work