monk-obj-test2 0.0.12

Monk Object Detection's 2_pytorch_finetune

Project Details

Pipeline based on GluonCV Fintuning project - https://gluon-cv.mxnet.io/build/examples_detection/index.html

Installation

Supports

• Python 3.6
• Python 3.7

cd installation

Check the cuda version using the command

nvcc -V

Select the right requirements file and run

cat <selected requirements file> | xargs -n 1 -L 1 pip install

For example for cuda 9.0

cat requirements_cuda9.0.txt | xargs -n 1 -L 1 pip install

Functional Documentation

Link

Pipeline

• Load Dataset

gtf.Dataset(root_dir, img_dir, anno_file, batch_size=batch_size);

• Load Model

gtf.Model(model_name, use_pretrained=pretrained, use_gpu=gpu);

• Set Hyper-parameter

gtf.Set_Learning_Rate(0.001);

• Train

gtf.Train(epochs, params_file);

TODO

• Add SSD support
• Add YoloV3 support
• Add support for Coco-Type Annotated Datasets
• Add support for VOC-Type Annotated Dataset
• Add Faster-RCNN support
• 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

External Contributors list

• https://github.com/THEFASHIONGEEK: Multi GPU feature