Giters

mseok / stylizing_textimage

2019 Fall, CS470 term project

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

stylizing_textimage

KAIST 2019 Fall, CS470 final project

Table of Contents

Introduction

This is pytorch implementation of Character Image Synthesis Based on Selected Content and Referenced Style Embedding, 2019.

Installation

  • All the code required to get started

Clone

  • Clone this repo to your local machine using below command
$ git clone https://github.com/mseok/stylizing_textimage

Setup

update and install package below

  • pytorch 1.3.1 with torchvision
  • CUDA 10.1
  • Seaborn
  • Matplotlib
  • OpenCV2
  • Numpy
### Datasets
> we use MCGAN datasets, you can get these with below command you should in main directory

~~~shell
$ mkdir datasets && cd datasets
$ wget https://people.eecs.berkeley.edu/\~sazadi/MCGAN/datasets/Capitals64.tar.gz
$ wget https://people.eecs.berkeley.edu/\~sazadi/MCGAN/datasets/Capitals_colorGrad64.tar.gz

Pretrained models

Due to large size of our pretrained save file, we uploaded it on google drive. You should execute following command in main directory. Note that this only works for LINUX system!!

wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=FILEID' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=FILEID" -O FILENAME && rm -rf /tmp/cookies.txt
  • Following command represent above instructions.
$ mkdir results/ && mkdir results/selector/ && cd results/selector
$ wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1fTX1wt_daS1xWJbRIdq6GkJiWY1EEzoq' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1fTX1wt_daS1xWJbRIdq6GkJiWY1EEzoq" -O save_0.pth.tar && rm -rf /tmp/cookies.txt
$ wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1dyLZga_a5v2934sPKKHkTVZ0EyA0zOuq' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1dyLZga_a5v2934sPKKHkTVZ0EyA0zOuq" -O save_5.pth.tar && rm -rf /tmp/cookies.txt 
$ wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1AMa7I5TbgLaNlvs8DZUL29QaZNXlhSgy' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1AMa7I5TbgLaNlvs8DZUL29QaZNXlhSgy" -O save_9.pth.tar && rm -rf /tmp/cookies.txt
  • please extract file in datasets after download .tar files

Usage

Train

you can easily use our run_train.sh

scripts/run_train.sh

#!/bin/sh

EPOCH=20
BATCH_SIZE=10
COLOR_PATH="datasets/Capitals_colorGrad64/train"
LATENT_DIM=1024
LEARNING_RATE=1e-3
EXPNAME='selector'
SAVE_FPATH='selector/save_8_1530.pth.tar'
SCHEDULE_FACTOR=5e-1
SCHEDULE_PATIENCE=5
MIN_LR=0
LAMBDA_VAL=2000

python3 train/train.py --epoch $EPOCH --batch_size $BATCH_SIZE --color_path $COLOR_PATH --latent_dim $LATENT_DIM --learning_rate $LEARNING_RATE --expname $EXPNAME --gpu --scheduler --schedule_factor $SCHEDULE_FACTOR --schedule_patience $SCHEDULE_PATIENCE --min_lr $MIN_LR --lambda_val $LAMBDA_VAL

# if want to load, add below command
# --load --save_fpath $SAVE_FPATH
  • you can set every hyper-parameters here.
  • if you want to load pretrained model add --load and --save_fpath options!
  • your training model will save automatically in results/EXPNAME !
  • if you end with setting type below
$ bash scripts/run_train.sh

Test

you can also easily use our make_font.sh

scripts/make_font.sh

#!/bin/sh

INPUT_LOCATION="pocari.png"
SOURCE_CHARACTER='pocar'
PRETRAINED_LOCATION='results/selector/save_9.pth.tar'
OUTPUT_NAME='pocar_font.png'
BATCH_SIZE=512
LATENT_DIM=1024

python3 test/test_users.py --input_location $INPUT_LOCATION --source_character $SOURCE_CHARACTER --pretrained_location $PRETRAINED_LOCATION --output_name $OUTPUT_NAME --batch_size $BATCH_SIZE --latent_dim $LATENT_DIM --gpu
  • if you do not want gpu option just delete --gpu
  • choose 5 letter font and type location in INPUT_LOCATION
  • your input size should be (64,320,3)
  • plz type source character in SOURCE_CHARACTER
  • output will save in results/OUTPUT_NAME !
  • this phase will takes really really long time.. (because of select phase)
  • if you end with setting type below
$ scripts/bash make_font.sh

Test_non_users

if you can't wait select phase, follow below

scripts/run_test.sh

#!/bin/sh

INPUT_LOCATION="datasets/Capitals_colorGrad64/train/akaPotsley.0.2.png"
PRETRAINED_LOCATION='results/selector/save_0.pth.tar results/selector/save_1.pth.tar results/selector/save_2.pth.tar '
OUTPUT_FOLDER='outputs/'
OUTPUT_NAME='output_test_0.png output_test_1.png output_test_2.png'
LATENT_DIM=1024
COLOR_PATH="datasets/Capitals_colorGrad64/train"
NONCOLOR_PATH="datasets/Capitals64/BASE"
BATCH_SIZE=1

python3 test/test.py --input_location $INPUT_LOCATION --pretrained_location $PRETRAINED_LOCATION --output_folder $OUTPUT_FOLDER --output_name $OUTPUT_NAME --latent_dim $LATENT_DIM --color_path $COLOR_PATH --noncolor_path $NONCOLOR_PATH --batch_size $BATCH_SIZE
  • usage is almost same with test
  • but this require the images in datasets/Capitals_colorGrad64/
  • you can have multiple location and output_name.
  • if you end with setting type below
$ bash scripts/run_test.sh

Team

We are team 22 in cs470, KAIST

License

License

About

2019 Fall, CS470 term project

License:MIT License

Languages

Language:Python 96.9%Language:Shell 3.1%