Tensorflow-Lite object detection

Transfer Learning .

This repo is just an simple version of

Tensorflow transfer learning on TPU

Deploy

Running on Intel Compute stick 2

Installation

as mention in documentation,

installation requires as follows:

• Protobuf 3.0.0
• Python-tk
• Pillow 1.0
• lxml -tf Slim (which is included in the "tensorflow/models/research/" checkout)
• Jupyter notebook
• Matplotlib
• Tensorflow (>=1.12.0)
• Cython
• contextlib2
• cocoapi

Run training

add python path export PYTHONPATH=$PYTHONPATH=/home/s3nh/tf-mobilenet/models/research:/home/s3nh/tf-mobilenet/models/research/slim

Create tf records

python utils/create_licence_tf_record.py --csv_input=dataset/benchmarks/endtoend/labels.csv --output_dir=training/train.record --image_dir=dataset/benchmarks/endtoend/eu

train, based on pipeline.config file and training files. python3 models/research/object_detection/model_main.py --logtostderr --train_dir=training/ --pipeline_config_path=training/pipeline.config

Results after 4000 steps

Predictions

Classification sample

stored into openvino/deployment_tools/inference_engine/samples/python_samples/classification_sample

Step by step installation

Darft

1 clone tensorflow object detection api repo 2. prepare dataset for object detection instance 3. Convert dataset to .tfrecords dataset 4. train model using chosen pretrained topology. 5. save tf checkpoints 6. convert checkpoint to frozen_inference_graph

Converting to frozen_inference_graph.py

• Question: how to deal with image_tensor input_type, which is probably not allowed on intel ncs2 architecture

• OK python export_inference_graph.py --input_type image_tensor --pipeline_config_path training/faster_rcnn_inception_v2_pets.config --trained_checkpoint_prefix training/model.ckpt-XXXX --output_directory inference_graph

7. convert .pb model to IR format

Convertion process- .

python3 intel/openvino/deployment_tools/model_optimizer/mo_tf.py --input_model=tf-mobilenet/training_mobilenet/frozen_inference_graph.pb --tensorflow_object_detection_api_pipeline_config tf-mobilenet/training_mobilenet/ssd_mobilenet_v2_licence_plate.config --reverse_input_channels

8. Prepare inference engine script.

#REMINDER

Properly convert INCEPTION_V2 FASTER RCNN MODEL

python3 intel/openvino/deployment_tools/model_optimizer/mo_tf.py --input_model=tf-mobilenet/training/frozen_inference_graph.pb --tensorflow_object_detection_api_pipeline_config=tf-mobilenet/training/pipeline.config --reverse_input_channels --input_shape=[1,600,600,3] --log_level=DEBUG --tensorflow_use_custom_operations_config=intel/openvino/deployment_tools/model_optimizer/extensions/front/tf/faster_rcnn_support_api_v1.14.json

COnvertion on blob_input_TF

python3 intel/openvino/deployment_tools/model_optimizer/mo_tf.py --input_model=tf-mobilenet/blob_input_TF/frozen_inference_graph.pb --tensorflow_object_detection_api_pipeline_config=tf-mobilenet/blob_input_TF/pipeline.config --reverse_input_channels --log_level=DEBUG --tensorflow_use_custom_operations_config=intel/openvino/deployment_tools/model_optimizer/extensions/front/tf/faster_rcnn_support_api_v1.14.jsond

Sample object detection

intel/openvino/inference_engine/samples/python_samples/object_detection_sample_ssd/object_detection_sample_ssd.py