akashAtmana / Tensorflow_model_learning

Tensorflow_model_learning

Common Setup and First Model

1. Pre-requisite

Python (Version 3.9) 
Protobuf
Anaconda

2. Create Virtual Environment

git clone https://github.com/tensorflow/models.git
conda create -n {environment_name} pip python=3.9
conda activate {environment_name}

3. Install Protobuf in project

conda install protobuf
Add use_protobuf.py into models/research/ directory
cd models/research/
python use_protobuf.py object_detection/protos protoc
protoc object_detection/protos/*.proto --python_out=.

4. Install Dependencies

Copy /models/research/object_detection/packages/tf2/setup.py to models/research/ directory
python -m pip install .

5. Test all dependencies are installed or not

python object_detection\builders\model_builder_tf2_test.py
If any dependecy is pending: conda install {name}

6. Tensorflow Model Zoo

pip install wget
python model_downloader.py

7. Run Detection Script

Create a folder "outputs"
python .\detect_from_image.py -m ssd_mobilenet_v2_320x320_coco17_tpu-8\saved_model -l .\models\research\object_detection\data\mscoco_complete_label_map.pbtxt -i .\models\research\object_detection\test_images

8. Check Results in Outputs folder

9. Detection through Webcam

python .\detect_from_webcam.py -m ssd_mobilenet_v2_320x320_coco17_tpu-8\saved_model -l .\models\research\object_detection\data\mscoco_complete_label_map.pbtxt

Training New Model from Scratch

1. Pre- Setup

Follow the previous steps 1 to 5 as it is

2. Preparing Images

Collect images for each category
Keep 80% of the Images from each category into one folder /images/train
Keep the rest 20% in another folder /images/test
pip install labelImg
labelImg.exe
Open the train directory in LabelImg and annotate each image- XML file for each image will be created
Do the same for the test folder

3. Convert XML to CSV

python xml_to_csv.py

4. Generate TF records

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

5. Tensorflow Model Zoo

pip install wget
python model_downloader.py

6. Create LabelMap

7. Edit Config file

8. Start Training

9. Export Model

10. Run Detection Script