xview3-kohei-solution

Usage

This repository assumes that the given data set is stored in the following locations:

$ ls data/input/xview3/*.csv
data/input/xview3/train.csv  data/input/xview3/validation.csv

$ ls data/input/xview3/downloaded
00a035722196ee86t.tar.gz  4455faa0cb4824f4t.tar.gz  85fe34d1aee53a7ft.tar.gz  c07f6ec980c2c149t.tar.gz
014261f774287442t.tar.gz  4518c556b38a5fa4t.tar.gz  864390795b0439b1t.tar.gz  c0831dbd6d7f3c56t.tar.gz
(snip)

In the case of training, create a virtual env as follows:

# Setup: Create the virtual env
$ poetry config virtualenvs.in-project true
$ poetry install
$ poetry run pip install albumentations timm segmentation_models_pytorch torch==1.10.0+cu113 torchvision==0.11.1+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

Training

To save storage space, my preprocessing code assumes that the input image file is the original file in .tar.gz format. It does NOT assume pre-extracted files. The batch size and other settings are based on the assumption that two RTX3080 GPU cards are used.

$ bash train.sh

Inference

My containerized inference code follows the xView3 evaluation protocol. The detailed usage of the xview3 evaluation protocol is described in https://iuu.xview.us/verify.

To build the Docker Image, you need the model weights generated by train.sh. The pre-computed files can be downloaded from my Google Drive.

You will also need to download and extract gshhg-shp-2.3.7.zip as external data. GSHHG data can be found at https://www.ngdc.noaa.gov/mgg/shorelines/data/gshhg/latest/

The Docker image containing these model weights and external data has already been uploaded to Docker Hub (smly/kohei-xview3:latest).

## Setup: Build the docker image -----
$ docker build --no-cache -t kohei-xview3 .

## Inference -----
$ docker run \
    --shm-size 16G \
    --gpus=1 \
    --mount type=bind,source=/home/xv3data,target=/on-docker/xv3data \
    kohei-xview3 \
    /on-docker/xv3data/ \
    0157baf3866b2cf9v \
    /on-docker/xv3data/prediction/prediction.csv

Kohei's 5th place solution for xview3 challenge

Related tags

Overview

xview3-kohei-solution

Usage

Training

Inference

Owner

Kohei Ozaki

BarcodeRattler - A Raspberry Pi Powered Barcode Reader to load a game on the Mister FPGA using MBC

Request execution of Galaxy SARS-CoV-2 variation analysis workflows on input data you provide.

Implementation of the state of the art beat-detection, downbeat-detection and tempo-estimation model

TART - A PyTorch implementation for Transition Matrix Representation of Trees with Transposed Convolutions

Deep Learning Algorithms for Hedging with Frictions

FACIAL: Synthesizing Dynamic Talking Face With Implicit Attribute Learning. ICCV, 2021.

PyTorch implementation of the Quasi-Recurrent Neural Network - up to 16 times faster than NVIDIA's cuDNN LSTM

Official codebase for "B-Pref: Benchmarking Preference-BasedReinforcement Learning" contains scripts to reproduce experiments.

Code for "Localization with Sampling-Argmax", NeurIPS 2021

NFT-Price-Prediction-CNN - Using visual feature extraction, prices of NFTs are predicted via CNN (Alexnet and Resnet) architectures.

Near-Duplicate Video Retrieval with Deep Metric Learning

Spiking Neural Network for Computer Vision using SpikingJelly framework and Pytorch-Lightning

MetaBalance: Improving Multi-Task Recommendations via Adapting Gradient Magnitudes of Auxiliary Tasks

IsoGCN code for ICLR2021

Simple machine learning library / 簡單易用的機器學習套件

Multivariate Time Series Transformer, public version

用强化学习DQN算法，训练AI模型来玩合成大西瓜游戏，提供Keras版本和PARL（paddle）版本

Instant-Teaching: An End-to-End Semi-Supervised Object Detection Framework

Pyramid Pooling Transformer for Scene Understanding

Unsupervised Foreground Extraction via Deep Region Competition