Diffusion Models for Generative Outfit Recommendation

This is the pytorch implementation of our paper at SIGIR 2024:

Diffusion Models for Generative Outfit Recommendation

Yiyan Xu, Wenjie Wang, Fuli Feng, Yunshan Ma, Jizhi Zhang, Xiangnan He

DiFashion Overview

An overview of DiFashion: it gradually corrupts outfit images with Gaussian noise in the forward process, followed by a parallel conditional denoising process to reconstruct these images. The denoising process is guided by three conditions: category prompt, mutual condition, and history condition.

Generated Examples

Generated images (enclosed in small black squares within each outfit) for the PFITB task.
Generated outfits for the GOR task.

Environment

Anaconda 3
python 3.8.13
torch 2.0.1
torchvision 0.15.2
Pillow 9.0.1
numpy 1.24.4
transformers 4.32.1
open-clip-torch 2.20.0
accelerate 0.20.3
diffusers 0.18.2
xformers 0.0.22
pytorch-fid 0.3.0
lpips 0.1.4

Usage

Dataset

The experimental data are in './datasets' folder, including iFashion and Polyvore-U. Each dataset consists of the following files:

train.npy: dict with four keys uids, oids, outfits, and category for training.

{
    'uids':[uid1,...],  # list of user ids
    'oids':[[oid1,...],...],  # user interacted outfit ids
    'outfits':[[[iid1,iid2,iid3,iid4],...],...],  # item ids within each interacted outfit
    'category':[[[cid1,cid2,cid3,cid4],...],...]  # category ids within each interacted outfit
}

valid_grd_dict.npy & test_grd_dict.npy: same data format as train.npy.
fitb_valid.npy & fitb_test.npy: same data format as train.npy for validation and test, respectively. For the Fill-In-The-Blank task, we randomly set one item id of each outfit to be 0, indicating the missing item that needs to be filled in.

fitb_valid_dict.npy & fitb_test_dict.npy: another format of fitb_valid.npy and fitb_test.npy.

{
    uid1:{
        oid1:[iid1,iid2,iid3,iid4],
        oid2:[...],
        ...
    },
    ...
}

valid_grd.npy & test_grd.npy: ground-truth outfit information of each oid within validation and test sets.

{
    oid1:{
        'outfits':[iid1,iid2,iid3,iid4],
        'category':[cid1,cid2,cid3,cid4]
    },
    ...
}

fitb_valid_retrieval_candidates.npy & fitb_test_retrieval_candidates.npy: we randomly sample 4 items from the item set to serve as negative samples, which are combined with the ground-truth item to form the entire candidate set for each outfit in fitb_valid.npy and fitb_test.npy for retrieval.
```
{
    uid1:{
        oid1:[cand1,cand2,cand3,cand4,cand5]  # cand1 is the ground-truth
        oid2:[...],
        ...
    },
    ...
}
```

train_history.npy, valid_history.npy & test_history.npy: user interacted category id and corresponding item id.

{
    uid1:{
        cid1:[iid1,iid2,...],  # user interacted items within category cid1
        cid2:[...],
        ...
    },
    ...
}

id_cate_dict.npy: category id and corresponding semantic category.
```
{0:'skirt', 1:"women's canva shoe", ...}
```
cate_iid_dict.npy & iid_cate_dict.npy: the correspondance between category id and item id.
To acquire additional data, kindly retrieve the necessary files from iFashion-cloud and Polyvore-cloud and put them into corresponding data path, respectively. Each source comprises two files and one folder.
- /processed: data processed for training, validation, test and evaluation. If you didn't download the files in the folder,
  - You can set the argument '--data_processed' to False in /DiFashion/train.py, which will initiate the processing of above base data into a suitable format and save them in the folder '/datasets/dataset_name/processed'.
  - Run the code '/Evaluation/extract_hist_embs.py' to obtain the remaining files required for evaluation.
- all_item_image_paths.npy: list of item image paths arranged according to item ids.
- cnn_features_clip.npy: extracted clip image embeddings of all the images arranged according to item ids. You can find relevant code in '/Evaluation/extract_hist_embs.py'.

For item images of iFashion,

Download item information item_info.npy from here with the format:

{
    iid1:{
        'original iid':ori_iid,
        'category':semantic_category,
        'url':image_url
    },
    ...
}

Download each image through the url into path '/img_folder_path/semantic_category/ori_iid.png', corresponding to the image path format in all_item_image_paths.npy.

Process these item images via function process_image.

from PIL import Image

def process_image(image):
    image = convert_to_rgb(image)
    image = pad_to_square(image)
    target_size = (512, 512)
    image = resize_image(image, target_size)
    
    return image

def convert_to_rgb(image):
    if image.mode != 'RGB':
        if image.mode != "RGBA":
            image = image.convert('RGBA')
        background = Image.new("RGBA", image.size, (255,255,255))
        result_image = Image.alpha_composite(background, image)
        result_image = result_image.convert("RGB")
        
        return result_image
    else:
        return image

def pad_to_square(image):
    # Transform the image into a square format, ensuring a uniform white background.
    width, height = image.size
    padding = abs(width - height) // 2
    
    if width < height:
        padding_mode = (padding, 0)
    else:
        padding_mode = (0, padding)
    
    squared_image = Image.new('RGB', (max(width, height), max(width, height)), (255, 255, 255))
    squared_image.paste(image, padding_mode)
    
    return squared_image

def resize_image(image, target_size):
    resized_image = image.resize(target_size, Image.LANCZOS)

    return resized_image

Add an empty image into the image folder path without semantic category.

from PIL import Image

empty_img = Image.new('RGB', (512, 512), (255,255,255))  # a pure white image
folder_path = '/img_folder_path/'
empty_img.save(os.path.join(folder_path, "empty_image.png"))

For item images of Polyvore-U,

Download all the images from here.
Unzip the file 291x291.tar.gz and put it into an appropriate path '/path/to/291x291'.

Add an empty image into the image folder.

from PIL import Image

empty_img = Image.new('RGB', (512, 512), (255,255,255))  # a pure white image
folder_path = '/path/to/291x291'
empty_img.save(os.path.join(folder_path, "empty_image.png"))

Training

Please configure correct paths of all data in train.py and set appropriate gpu id in config.yaml.

cd ./DiFashion
sh run_eta0.1.sh

Inference

Download the checkpoint released by us from here.
Put the checkpoint into appropriate folder.
Configure correct paths of all data in inf4eval.py and set appropriate gpu id in config.yaml.
Run inf4eval.py

cd ./DiFashion
sh run_inf4eval.sh

Evaluation

Download finetuned inception from here and put it into '/Evaluation/finetuned_inception/'.
Run the evaluation code through the .sh files. For example, to evaluate the performance of DiFashion on iFashion dataset within the Fill-In-The-Blank task, execute the corresponding evaluation code evaluate_fitb.py using the provided run_eval_fitb.sh.

cd ./Evaluation
sh run_eval_fitb.sh ifashion DiFashion all test log 3

YiyanXu/DiFashion