Awesome

Learning by Planning: Language-Guided Global Image Editing

Instroduction

This is the Pytorch implementation for paper "Learning by Planning: Language-Guided Global Image Editing".

Dependency

• Pytorch >= 1.0.0
• opencv-python
• panopticapi
• pycocotools
• easydict
• tensorboardX
• tensorflow<2.0.0
• tabulate
• dominate
• kornia

Installation

• Clone this repo

  git clone https://github.com/jshi31/T2ONet.git --recursive
  

• Install the submodule pyutils/edgeconnect according to its README.

  The critical thing is to download pre-trained model.

Dataset

All the working directory for the following commands are project root.

MIT-Adobe FiveKReq

• Download the fiveK image Download from Google Drive and unzip the file to data/FiveK/images in order to keep the data structure show in the file tree.

• Test the dataloader by running

PYTHONPATH='.' python datasets/FiveKdataset.py

GIER

Go to GIER webpage, download the following files:

• Download the GIER images into data/GIER/images
• Download the GIER mask into data/GIER/masks
• Download the GIER feature to data/GIER/features

Finally, the data structure should be the same as the file tree.

Test the data loader by running

PYTHONPATH='.' python datasets/GIERdataset.py

Plan Action Sequences

FiveK

Generate action sequence using operation planning

PYTHONPATH='.' CUDA_VISIBLE_DEVICES=0 python preprocess/gen_greedy_seqs_FiveK.py

Or download the sequence from Google Drive and unzip it to output/actions_set_1

GIER

Generate action sequence using operation planning (currenty has error)

PYTHONPATH='.' CUDA_VISIBLE_DEVICES=0 python preprocess/gen_greedy_seqs_GIER.py

Or download the sequence from Google Drive and unzip it to output/GIER_actions_set_1

T2ONet

FiveK Train

PYTHONPATH='.' CUDA_VISIBLE_DEVICES=0 python experiments/t2onet/train_seq2seqL1.py --batch_size 64 --print_every 50 --checkpoint_every 1000 --num_iter 10000 --trial 2

FiveK Test

PYTHONPATH='.' CUDA_VISIBLE_DEVICES=0 python experiments/t2onet/test_seq2seqL1.py --print_every 100 --visualize_every 100  --visualize 0 --is_train 0 --trial 1 

select the trial number indicates which model you will use. To test our provided model, first download it from Google Drive and unzip it to output/FiveK_trial_1/seq2seqL1_model

and set the trial argument as 1 in the testing model.

To visualize the result, set visualize argument as 1, and the result will be in FiveK_trial_1/test/web

GIER Train

PYTHONPATH='.' CUDA_VISIBLE_DEVICES=0 python experiments/t2onet/train_GIER_seq2seqL1.py  --dataset GIER --session 3 --batch_size 64 --data_mode global+shapeAlign --print_every 100 --checkpoint_every 1000 --num_iter 20000 --trial 1 

GIER Test

PYTHONPATH='.' CUDA_VISIBLE_DEVICES=0 python experiments/t2onet/test_GIER_seq2seqL1.py  --dataset GIER --session 3 --data_mode global+shapeAlign --print_every 20 --visualize_every 5 --visualize 0 --trial 7

To test our provided model, first download it from Google Drive and unzip it to output/GIER_trial_7/seq2seqL1_model

and set the trial argument as 7 in the testing model.