Skip to content

Latest commit

 

History

History

13_tf_obj_2

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 

Project Details

Pipeline based on - https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2.md




Supported Models

  • ssd_mobilenet_v2_320
  • ssd_mobilenet_v1_fpn_640
  • ssd_mobilenet_v2_fpnlite_320
  • ssd_mobilenet_v2_fpnlite_640
  • ssd_resnet50_v1_fpn_320
  • ssd_resnet50_v1_fpn_640
  • ssd_resnet101_v1_fpn_320
  • ssd_resnet101_v1_fpn_640
  • ssd_resnet152_v1_fpn_320
  • ssd_resnet152_v1_fpn_640
  • faster_rcnn_resnet50_v1_640
  • faster_rcnn_resnet50_v1_1024
  • faster_rcnn_resnet101_v1_640
  • faster_rcnn_resnet101_v1_1024
  • faster_rcnn_resnet152_v1_640
  • faster_rcnn_resnet152_v1_1024
  • faster_rcnn_inception_resnet_v2_640
  • faster_rcnn_inception_resnet_v2_1024
  • efficientdet_d0
  • efficientdet_d1
  • efficientdet_d2
  • efficientdet_d3
  • efficientdet_d4
  • efficientdet_d5
  • efficientdet_d6
  • efficientdet_d7


Installation

Supports

  • Python 3.6
  • Cuda 10.0 (Other cuda version support is experimental)
  • Tensorflow 1.15

cd installation

chmod +x install_cuda10.sh && ./install_cuda10.sh




Pipeline

  • Load Dataset

gtf.set_train_dataset(train_img_dir, train_anno_dir, class_list_file, batch_size=2, trainval_split = 0.8)

gtf.create_tfrecord(data_output_dir="data_tfrecord")

  • Set Model

gtf.set_model_params(model_name="ssd_mobilenet_v2_320")

  • Set Hyper Params

gtf.set_hyper_params(num_train_steps=1000, lr=0.1)

  • Train

%run Monk_Object_Detection/12_tf_obj_1/lib/train.py

  • Export Model

%run Monk_Object_Detection/12_tf_obj_1/lib/export.py




TODO

  • Add support for Coco-Type Annotated Datasets
  • Add support for VOC-Type Annotated Dataset
  • Test on Kaggle and Colab
  • Add validation feature & data pipeline
  • Add Optimizer selection feature
  • Enable Learning-Rate Scheduler Support
  • Enable Layer Freezing
  • Set Verbosity Levels
  • Add Project management and version control support (Similar to Monk Classification)
  • Add Graph Visualization Support
  • Enable batch proessing at inference
  • Add feature for top-k output visualization
  • Add Multi-GPU training
  • Auto correct missing or corrupt images - Currently skips them
  • Add Experimental Data Analysis Feature