PyTorch implementation of the paper Learning Fashion Compatibility with Bidirectional LSTMs [1].
INSTALLATION
In order to run this code, you must install:
- PyTorch (install it with CUDA support if you want to use GPUs, which is strongly recommended).
- Python with packages numpy, torchvision, tensorboardX, PIL, collections, cv2 (only if you want to generate result images for fill in the blank task, tested with version 2.4.8), h5py, sklearn, nltk.
If this is your first time using nltk, you'd probably have to download wordnet corpora by executing the following lines in python:
import nltk
nltk.download('wordnet')The code has been tested with these two configurations:
- python 2.7.6 - Pytorch 0.3.0.post4 - CUDA 7.5 - numpy 1.11.1 - sklearn 0.18.1
- python 3.6.0 - PyTorch 0.4.0a0+040336f - CUDA 9.1 - numpy 1.13.3 - sklearn 0.19.1
DATASET
Download the Polyvore dataset from the authors of the paper here and extract it in the bilstm/data folder.
TRAINING
To train the model, run:
python main.py
You can take a look at the evolution of the training by running tensorboard --logdir runs.
Note that we included options to change the default parameters (learning rate, batch size, and convolutional model to extract image features). By default, these are just like in the paper: 0.2, 10 and inception_v3.
EVALUATION
We provide a snapshot for our trained model (with inception) here.
Prior to running the evaluation of the models, you must extract the features of the test images. In order to do so, run:
python src/get_features.py -m model_path -sp path_to_save_data [-t model_type]
This will save a h5 file with the image features.
Evaluation of compatibility AUC
To reproduce the results in the 2nd column of table 1 of the paper, run:
python src/evaluation.py -m model_path -sp path_to_saved_data [-t model_type]
Evaluation of fill in the blank accuracy
To reproduce the results in the 1st column of table 1 of the paper, run:
python src/outfit_generation.py -m model_path -sp path_to_saved_data -i path_to_save_results [-t model_type]
This will generate outfit images in the folder specified in -i. First row of the image is the question with an empty space. Second row shows the possible answers, with the predicted item squared in green (if correct) or in red (if wrong, with the correct answer in green).
[1] Han, Xintong, et al. "Learning Fashion Compatibility with Bidirectional LSTMs." Proceedings of the 2017 ACM on Multimedia Conference. ACM, 2017.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent