MultitaskNet implementation in PyTorch
Prerequisites
- Linux
- Python 3.7.0
- CPU or NVIDIA GPU + CUDA CuDNN
Getting Started
Installation
git clone https://github.com/imvinod/MultitaskNet
cd MultitaskNet
pip install -r requirements.txt
Dataset preparation
sunrgbd dataset
FuseNet train/test
visdom visualization
- To view training errors and loss plots, set
--display_id 1
, run python -m visdom.server
and click the URL http://localhost:8097
- Checkpoints are saved under
./checkpoints/sunrgbd/
train & test on sunrgbd
python train.py --dataroot datasets/sunrgbd --dataset sunrgbd --name sunrgbd
python test.py --dataroot datasets/sunrgbd --dataset sunrgbd --name sunrgbd --epoch 100
Results
- We use the training scheme defined in MultitaskNet
- Loss is weighted for SUNRGBD dataset
- Learning rate is set to 0.0001 for SUNRGBD dataset
- Results can be improved with a hyper-parameter search
More details coming up !