Course project (optional task 1) of SJTU CS3612: Machine Learning, 2023 spring.
Attention: Discussion & reference welcomed, but NO PLAGIARISM !!!
- You should design and train a VAE on LFW dataset.
- You should use Linear interpolation to generate images with specific properties.
For example, given two output features of the encoder, i.e.,
$z_1$ and$z_2$ , the decoder takes$z_1$ as the input to generate a face image of a woman, and takes$z_2$ as the input to generate a face image of a man. You can use Linear interpolation to obtain a new feature$z = \alpha z_1 + (1 − \alpha)z_2 ,\alpha \in (0,1)$ . Then, the decoder takes 𝑧 as the input to generate a new face image. You are required to do experiments with different values of 𝛼, e.g., 𝛼=0.2, 0.4, 0.6, 0.8, so as to obtain a set of face images. • You can conduct experiments on any two categories (e.g., male and female, old and young)
left: original input / right: reconstructed output
Pairwise comparison:
Random integration:
Male and female:
Old and young:
- Note that you may need to manually download the LFW dataset via this link, and place the unzipped folder under directory
dataset. - To train with default settings, make sure to leave no less than
32 GBof memory. - Train model with tuned parameters
python train.py
- Test image reconstruction and generation
python main.py
- Generate the Old-Young fusion example
python example.py
For further information, refer to Section 2 of the project report here.