A pytorch implementation of LMSER, which is a bidirectional architecture with some built-in natures. In this project, we implement the Dual Connection Weight (DCW) property of LMSER. We implemented both LMSER with and without the pseudo inverse constrain, and compare their performance on reconstructing MNIST and Fashion-MNIST with AutoEncoder. We also add the supervised learning to the middle-layer embedding vector (16-d vector in our implementation), which aims to test the supervised learning effect on original reconstruction task.
- Linux, Ubuntu 16.04 or 18.04
- Python 3.6+
- Pytorch 1.4.0+
- CUDA 10.0+
- Create a conda virtual environment and activate it.
conda create -n lmser python=3.6
conda activate lmser
-
Install Pytorch and torchvision following the official instruction.
-
Clone the LMSER-Pytorch repository.
git clone https://github.com/Simon-Fuhaoyuan/LMSER-pytorch
- Install build requirements
cd LMSER-pytorch
pip install -r requirements.txt
We offer both training and testing codes.
To train a model, please run
python train.py [-h] [--epochs EPOCHS] [--batch_size BATCH_SIZE] [--lr LR] \
[--mean MEAN] [--std STD] [--dataset DATASET] \
[--weight_dir WEIGHT_DIR] [--image_dir IMAGE_DIR] [--cpu_only] \
[--no_test] \
model
For example, to train AutoEncoder, run
python train.py AutoEncoder --epochs 50 --batch_size 64 --dataset mnist
To test a model, run
python test.py [-h] [--dataset DATASET] [--batch_size BATCH_SIZE] \
[--mean MEAN] [--std STD] [--image_dir IMAGE_DIR] [--cpu_only] \
model weight
For example, to test a trained model, whose weight model is saved at ./weight_dir/AutoEncoder.pth, run
python test.py AutoEncoder ./weight_dir/AutoEncoder.pth --dataset mnist
The usage of train_supervise.py and test_supervise.py is quite similar, but they will load the supervised version of model and train or test.
For more usages, please refer to experiments_command.txt.